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.0004823 | Barriers and Recommended Interventions to Prevent Melioidosis in Northeast Thailand: A Focus Group Study Using the Behaviour Change Wheel | Melioidosis, an often fatal infectious disease in Northeast Thailand, is caused by skin inoculation, inhalation or ingestion of the environmental bacterium, Burkholderia pseudomallei. The major underlying risk factor for melioidosis is diabetes mellitus. Recommendations for melioidosis prevention include using protective gear such as rubber boots and gloves when in direct contact with soil and environmental water, and consuming bottled or boiled water. Only a small proportion of people follow such recommendations.
Nine focus group discussions were conducted to evaluate barriers to adopting recommended preventive behaviours. A total of 76 diabetic patients from northeast Thailand participated in focus group sessions. Barriers to adopting the recommended preventive behaviours and future intervention strategies were identified using two frameworks: the Theoretical Domains Framework and the Behaviour Change Wheel.
Barriers were identified in the following five domains: (i) knowledge, (ii) beliefs about consequences, (iii) intention and goals, (iv) environmental context and resources, and (v) social influence. Of 76 participants, 72 (95%) had never heard of melioidosis. Most participants saw no harm in not adopting recommended preventive behaviours, and perceived rubber boots and gloves to be hot and uncomfortable while working in muddy rice fields. Participants reported that they normally followed the behaviour of friends, family and their community, the majority of whom did not wear boots while working in rice fields and did not boil water before drinking. Eight intervention functions were identified as relevant for the intervention: (i) education, (ii) persuasion, (iii) incentivisation, (iv) coercion, (v) modeling, (vi) environmental restructuring, (vii) training, and (viii) enablement. Participants noted that input from role models in the form of physicians, diabetic clinics, friends and families, and from the government via mass media would be required for them to change their behaviours.
There are numerous barriers to the adoption of behaviours recommended for melioidosis prevention. We recommend that a multifaceted intervention at community and government level is required to achieve the desired behaviour changes.
| Melioidosis is a serious infectious disease caused by the Gram-negative environmental saprophyte, Burkholderia pseudomallei. Infection in humans occurs following skin inoculation, inhalation or ingestion. Recommendations for melioidosis prevention include using protective gear such as rubber boots and gloves when in direct contact with soil and environmental water, and consuming bottled or boiled water. Northeast Thailand is a hot spot for melioidosis, but only a small proportion of people follow such recommendations. Here, we evaluated barriers to the adoption of preventive behaviours in diabetics (who are at highest risk for melioidosis), and systematically identified key functions required for future interventions. Our study participants had no knowledge of the disease, believed that there was no harm in not adopting the recommended preventive behaviours, and were not inclined to use boots and gloves while working in muddy rice fields. Participants reported that input from numerous role models (physicians, diabetic clinics, friends and families), and from the government via mass media would be required for them to change their behaviours. We recommend that a multifaceted intervention at community and government level is required to bring about the desired changes.
| Melioidosis is a serious community-acquired infectious disease caused by the Gram-negative bacillus Burkholderia pseudomallei, which is present in soil and water in many tropical countries in Central and South America, sub-Saharan Africa, South Asia, Southeast Asia and northern Australia [1, 2]. The bacterium is intrinsically resistant to a wide range of antimicrobials, and treatment with ineffective antimicrobials has case fatality rates exceeding 70% [3]. An estimated 165,000 human melioidosis cases occur each year worldwide, of which 89,000 (54%) die [4]. In northeast Thailand, melioidosis is the second most common cause of community-acquired bacteremia, and the number of people dying there from melioidosis is now comparable to deaths from tuberculosis, and exceeds those from malaria, diarrheal illnesses and measles combined [5, 6]. Diabetes mellitus is the major underlying risk factor for melioidosis, and is present in more than 50% of all melioidosis cases [3]. The risk of people with diabetes acquiring melioidosis is about 12 times higher than the rest of the population [6, 7]. People with diabetes are therefore the major target population for melioidosis preventive measures [8].
Melioidosis is potentially preventable, as infection occurs by skin inoculation, inhalation or ingestion of bacteria in soil and water in endemic areas [3]. No melioidosis vaccine is currently available for human use [8]. Evidence-based guidelines for the prevention of melioidosis in Thailand recommends that residents and visitors should avoid direct contact with soil and water, wear protective gear such as boots and gloves when in direct contact with soil or water and only drink bottled or boiled water [9]. These recommendations have been repeatedly promoted by the Ministry of Public Health (MoPH) of Thailand to prevent leptospirosis, particularly since the sharp increase in the incidence of leptospirosis in Thailand between 1997 and 2000 [10]. In addition, in 2012 the MoPH launched a platform for improved preparedness, response and prevention to infectious diseases under a ‘One health concept’, which recognizes that human health is strongly connected to animal health and the environment [11]. Nonetheless, our previous study found that only a small proportion of people living in northeast Thailand followed such recommendations [9]. For example, many people still work in rice fields without protective gear, and drink untreated water [9]. In addition, there is no national campaign that is specific for melioidosis, public awareness of melioidosis is very low, and more than 90% of Thais have no knowledge of the disease [12].
Changing behaviour is typically complex, and a systematic approach is required to understand factors that influence adherence to recommendations so as to inform the design of future preventive interventions. In general, providing information alone does not change their behaviour [13, 14]. Frameworks have been developed that structure a wide range of possible influences on behaviour, including the Theoretical Domains Framework (TDF) and associated Behaviour Change Wheel (BCW [15–17]). The TDF is a useful framework for understanding the barriers and factors influencing specific behaviours [15, 16, 18, 19], while the BCW is a comprehensive framework that links this understanding to design interventions [17, 20]. Examples of using these frameworks to investigate implementation problems are the delivery of sepsis care bundles [21], antibiotic stewardship in healthcare facilities [22], and changing dietary behaviours in overweight children [23].
In this study, our aim was to evaluate barriers to and facilitators of behaviours recommended for melioidosis prevention. We then systematically identified key functions of interventions likely to be effective in increasing adherence to recommendations.
Focus group interviews were conducted to evaluate barriers to adopting (i) the use of protective gear such as rubber boots and gloves during direct contact with soil and environmental water, and (ii) the consumption of bottled or boiled water. These behaviours were selected from the list of recommendations for melioidos is prevention [9] on the assumption that these would be highly effective if adopted. The questions for the focus group interviews were grouped according to the Theoretical Domains Framework (TDF) (Table 1 and S1 Table), which comprises 14 key domains related to behaviour (S2 Table). TDF was chosen because it has been validated and proved useful to inform interventions aimed at bringing about behavioural changes [15].
In April 2012, the focus group interviews were conducted at Det Udom Royal Crown Prince Hospital, Warin Chamrap hospital and Don Mot Daeng hospital, Ubon Ratchathani province, northeast Thailand, where melioidosis is highly endemic [6]. These three district hospitals were chosen based on the comparable size of the hospital (range from 30 to 90 beds) and distance from the center of Ubon Ratchathani province (within 100 kilometers). The majority of the population in these districts live in rural settings, and most adults (around 80%) are engaged in agriculture, particularly rice farming. Diabetic patients who came to follow-up visits at diabetic clinics were invited in sequential order to participate in focus group interviews. Eligible participants were males and females aged 18–60 years old who had been diagnosed with diabetes for at least 3 months. Diabetics in northeast Thailand were selected as the study population because they are at the highest risk of melioidosis and are the target population for prevention [24]. We excluded participants with a history of melioidosis. Participants were encouraged to compare and discuss their viewpoints within the group. Nittayasee Wongsuwan acted as a moderator and probed participants to elaborate on comments as necessary. After there were no new viewpoints raised, indicating saturation point, no further interviews were conducted. Each focus group lasted about 60–90 minutes.
Data from focus group discussions were recorded in video formats, and detailed notes were taken by a note taker (Mayura Malasit). All videos were transcribed verbatim, and supplementary notes were added to ensure that all relevant participant comments and ideas were captured. To familiarize themselves with the data, two of the authors (Pornpan Suntornsut and Direk Limmathurotsakul) watched the videos and read the transcripts twice.
First, we used TDF and a deductive analytic process to classify responses [25, 26] according to the domains within the TDF [15]. For instances where coding differed between coders, differences of interpretation were discussed and an agreement was reached by consensus.
Second, we used a second framework, the Behaviour Change Wheel (BCW [17]) to identify the intervention functions most likely to be effective in changing the TDF domains identified. BCW was chosen because it was developed systematically, fitted well with the TDF [15], and has been found to be a robust starting point for designing interventions and planning policy [20]. Details of the BCW and explicit links between TDF domains and the BCW are given in the BCW guide [20]. In brief, the BCW is composed of a simple model of behaviour, COM-B, comprised of Capability, Opportunity, Motivation and Behaviour [20]. This sits at the hub of the ‘wheel’ linked to an inner ring of nine intervention functions (education, persuasion, incentivisation, coercion, training, restriction, environmental restructuring, modeling and enablement) which are in turn linked to an outer ring of seven policy categories (environment/social planning, communication/marketing, legislation, service provision, regulation, fiscal measures and guidelines [17, 20]). COM-B components represent sources of behaviour, which would need to be changed for the desired behaviour to occur. Capability is divided into physical and psychological, opportunity into physical and social, and motivation into reflective and automatic [17, 20]. We then linked general intervention functions to specific behaviour change techniques (BCTs) [27]. We used a set of criteria called APEASE to select the most appropriate intervention functions, BCTs and modes of delivery for our setting. APEASE criteria are (i) Affordability, (ii) Practicability, (iii) Effectiveness/cost-effectiveness, (iv) Acceptability, (v) Safety/side-effects and (vi) Equity in making context-based decisions [17, 20].
Approval for the study was obtained from the Ethics Committee of the Faculty of Tropical Medicine, Mahidol University, Bangkok, Thailand. Written informed consent was obtained from each participant prior to conducting each focus group interview. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
A total of 76 diabetic patients participated in nine focus group interviews (a median of 8 diabetic patients per group, ranging from 7 to 11). Overall, 20 (26%) were male and 56 (74%) were female. The median age of participants was 54 years (interquartile range, 47 to 61 years). Fifty-five participants (72%) were rice farmers and 13 (17%) were non-rice farmers. The majority of participants attained primary school education (89%; 68/72) and had a personal income less than 5,000 baht/month (equivalent to about 140 dollars/month; 86%; 64/72). Information about education and socioeconomic level was not available from two participants.
Most participants had no knowledge of the disease, believed that there was no harm in not adopting the recommended preventive behaviours, and were not inclined to use boots and gloves while working in muddy rice fields. Also, participants tended to drink water without boiling. Factors influencing the behaviours recommended for melioidosis prevention were related to five domains: (i) knowledge, (ii) beliefs about consequences, (iii) intention and goals, (iv) environmental context and resources; and (vi) social influences (Table 1). These were elaborations of four COM-B components: ‘psychological capability’, ‘reflective motivation’, ‘physical opportunity’ and ‘social opportunity’ (Table 2).
Guided by the links between COM-B and intervention functions (S3 Table), we identified nine intervention functions that could be used to change the targeted COM-B components. We found that ‘restriction’ would not be applicable in our context. Therefore, the recommended intervention functions included ‘education’, ‘persuasion’, ‘incentivisation’, ‘coercion’, ‘modelling’, ‘environmental restructuring’, ‘training’ and ‘enablement’ (Table 2).
Guided by links between Intervention functions and BCTs (S4 Table), we considered that BCTs appropriate for our context included ‘information about health consequences’, ‘feedback on behaviour’, ‘feedback on outcomes of the behaviour’, ‘prompts/cues’, ‘self-monitoring of behaviour’, ‘credible source’, ‘demonstration of the behaviour’, ‘instruction on how to perform a behaviour’, ‘commitment’, ‘behavioural practice/rehearsal’, ‘adding objects to the environment’, ‘restructuring the physical environment’, ‘social support’, and ‘goal setting’. Examples of each BCT were developed and are listed in Table 3.
Guided by links between Intervention functions and policy categories (S5 Table), we considered that the following policies would support the delivery of intervention functions for our context: ‘communication/marketing’, ‘guidelines’, ‘environmental/social planning’ and ‘service provision’. We found that ‘fiscal measures’, ‘regulation’ and ‘legislation’ would not be applicable in our context.
Using a taxonomy of modes of delivery for intervention functions (S1 Fig), we considered that the distance individual mode (phone helpline, mobile phone text and individually accessed computer programme) was unpractical and irrelevant to our setting. Therefore, the recommended modes of delivery included face-to-face mode (individual and group levels) and distance mode (population level). Distance mode at population levels could include broadcast media (television and radio), outdoor media (posters and billboards), digital media (internet and mobile phone applications) and print media (leaflets, newspaper and other written materials).
This study shows that the barriers for melioidosis prevention in Thailand are related to multiple domains, and we suggest that multiple intervention functions, BCTs and policies are required for the changes to be successful. The barriers identified are not limited to a lack of knowledge of the disease and measures to prevent it, and so providing information alone is unlikely to lead to the necessary behaviour changes. They also believe that there is no harm in not adopting the recommended preventive behaviours and give reasons for not boiling water due to lack of time and not using rubber boots and gloves in muddy rice fields due to discomfort. Understanding these barriers are also crucial as they point to the behaviours that require modification in order for the prevention to be effective. To our knowledge, this kind of systematic approach, using frameworks such as TDF and BCW to evaluate barriers and design an intervention after recommendations have been developed [10] is rarely performed in Thailand. This is highlighted by the fact that many people still work in rice fields without protective gear [9], despite the fact that wearing protective gear has been recommended in Thailand for many years [10, 11]. Our recommendations are based on the identified barriers, a systematic approach and local context. Furthermore, no single recommended BCTs or policy changes could affect all of the barriers. Therefore, all of the recommendations will need to be considered for the development of future interventions for melioidosis prevention in northeast Thailand.
Finding that most of our focus group participants (95%) have never heard of melioidosis is not surprising. Our result is comparable with the previous national survey finding that 74% of adults in Thailand have not heard of melioidosis [12]. The difference (95% vs. 74%) could be mainly because the survey respondents in the previous study were relatively younger and had higher education levels [12]. Lack of knowledge was found to be a major barrier in behaviour changes for many infectious diseases in Southeast Asia, including chronic obstructive pulmonary disease in Malaysia [28], sexual transmitted diseases in Cambodia [29] and liver fluke infection in Thailand [30]. Having accurate knowledge of the disease is fundamental for people to change their behaviour [17]. Therefore, providing knowledge about melioidosis and its prevention is still needed, and should be one of the main components in behaviour change interventions.
It was striking that most participants believed that there was no harm in not adopting the recommended preventive behaviours. Our finding is consistent with several studies in tropical countries where beliefs about consequences are one of the major barriers to the adoption of preventive behaviours. For example, Filipino farmers and laborers believed that wearing gloves during spraying pesticides would cause an illness called pasma rather than protecting themselves from pesticides [31]. Cambodian parents thought that HPV vaccine was unnecessary as they had traditional beliefs that whatever was going to happen would happen [32]. As we identified that ‘belief about consequences’ is a major barrier for melioidosis prevention, future interventions should also include ‘persuasion’ and ‘modeling’ as part of main intervention. This could be delivered through several BCTs in the interventions, as shown in Table 3.
A major concern raised by the focus group participants was that Wellington boots are hot and make walking difficult in muddy rice fields. Environmental context and resources might have been frequently overlooked in previous campaigns, in which Wellington boots have frequently been provided to farmers in Thailand as part of the previous campaign to prevent leptospirosis. The problem of boots was also raised in studies in other tropical developing countries in Sri Lanka [33] and Philippines [31]. Our pilot studies show that over-the-knee boots, hip boots and half-body waders can be used in flooded rice fields without causing difficulty in walking, but may still be uncomfortable in hot weather. Further studies are needed to focus on developing and trialing specifically designed boots that could allow farmers to walk easily in muddy paddy fields and comfortably in tropical developing countries.
Our study highlights the importance of ‘credible source’ and ‘social support’ as possible components of an intervention. This is consistent with other tropical developing countries, where these factors are very important in rural settings. For example, a study of oral poliovirus vaccines in Nigeria showed that having religious leaders, town announcers and health workers as primary sources of health information were strongly related to an individual’s probability of receiving the vaccine [34]. A study of acute respiratory infections in Bangladesh showed the importance of family and community people in decision-making [35]. Our participants even questioned whether the burden of melioidosis was real because they have never seen any information or campaign from the government via mass media. Our systematic approach also shows that multiple policy categories are required. Although guidelines (as one of the policy categories) for melioidosis prevention are now available [9], additional guidelines including all changes to service provision for policy makers is still needed. This suggests that commitment and action by the government are essential for the preventive interventions to be successful.
Our study has several strengths. First, we used a systematic approach (TDF and BCW) to evaluate barriers to behaviours recommended for melioidosis prevention, and we provide a range of recommendations for policy makers to use for behaviour change interventions in the future. If the interventions are designed without a systematic approach, it is common that a number of relevant intervention functions, BCTs and modes of delivery could be overlooked [20]. This systematic approach is useful to encourage intervention designers to be comprehensive in considering all options to intervene and then to systematically select those that are most promising for the context as shown in the previous successful examples [21–23]. Second, we selected only two behaviours that we consider would be highly effective for melioidosis prevention. This is consistent with the recommendation of behaviour change theory that the intervention should initially focus to just one or a few behaviours, and that building on small successes is more effective than intervening many behaviours simultaneously [20]. Third, the target behaviours in our study are also target behaviours recommended under the ‘One Health concept’ that could prevent other infectious diseases such as leptospirosis and acute diarrhea if adopted [11].
The major limitation of this study is that the identified barriers and recommended interventions to prevent melioidosis may not be equally relevant to all age and socioeconomic groups of the diabetic population in Thailand and beyond. It is possible that some barriers vary and that the intervention functions would need to be adjusted based on local context.
In conclusion, we recommend that health care providers together with policy makers should consider multifaceted interventions for melioidosis prevention. Health care providers should focus on delivering behaviour change interventions based on our recommended BCTs. Policy makers should focus on delivering disease education and implementing its preventive measures through healthcare providers and, particularly, through mass media.
|
10.1371/journal.pcbi.1000985 | Analysis of Stochastic Strategies in Bacterial Competence: A Master Equation Approach | Competence is a transiently differentiated state that certain bacterial cells reach when faced with a stressful environment. Entrance into competence can be attributed to the excitability of the dynamics governing the genetic circuit that regulates this cellular behavior. Like many biological behaviors, entrance into competence is a stochastic event. In this case cellular noise is responsible for driving the cell from a vegetative state into competence and back. In this work we present a novel numerical method for the analysis of stochastic biochemical events and use it to study the excitable dynamics responsible for competence in Bacillus subtilis. Starting with a Finite State Projection (FSP) solution of the chemical master equation (CME), we develop efficient numerical tools for accurately computing competence probability. Additionally, we propose a new approach for the sensitivity analysis of stochastic events and utilize it to elucidate the robustness properties of the competence regulatory genetic circuit. We also propose and implement a numerical method to calculate the expected time it takes a cell to return from competence. Although this study is focused on an example of cell-differentiation in Bacillus subtilis, our approach can be applied to a wide range of stochastic phenomena in biological systems.
| When exposed to stress, organisms react by taking actions that help them protect their DNA. ComK protein is a key regulator which activates hundreds of genes, including the genes encoding the DNA-uptake and recombination systems. In Bacillus subtilis, stress in the environment activates a sequence of chemical reactions that, driven by cellular noise, stochastically increases the level of ComK in some bacterial cells driving them from their original vegetative state into a competent state. Entrance into and exit from competence are stochastic switching events that the cell undergoes. In this work, we present a novel numerical method that allows the analysis of stochastic events in biological systems. We illustrate our method by computing the probability with which Bacillus subtilis enters in competence. We also present a method to analyze the sensitivity of stochastic events. We use this method to study the sensitivity of the probability of entrance in competence with respect to various gene expressions and degradation rates. We finally present a numerical method to calculate the expected time it takes a cell to return from competence. Although we studied the competence regulatory genetic circuit, our approach can be applied to a variety of stochastic events in biological systems.
| Competence is the ability of a cell, usually a bacterium, to bind and internalize transforming exogenous DNA. Under stressful environments, such as nutrient limitations, some cells enter competence while other cells commit irreversibly to sporulation. Entry in competence is a transient probabilistic event that facilitates copying of the exogenous DNA [1], [2]. It has been shown that among a group of cells only a randomly chosen fraction enters in competence [3], [4]. Proper modeling and correctly accounting for noise in the model of this phenomenon is crucial to understanding the underlying biological explanation. The few cells that enter competence express a high concentration of the key regulator ComK, which activates hundreds of genes, including the genes encoding the DNA-uptake and recombination systems [5]–[7]. Competence is understood as a bistability pattern [4], [8] and the nonlinear system describing the competence regulatory circuit is an excitable dynamical system.
Auto-activation of the regulator ComK is responsible for the bistable response in competence development. Auto-activation of ComK, is essential and can be sufficient to generate a bistable expression pattern [9]–[11]. Specifically, the concentration of an inducer must cross a certain threshold to start the positive feedback. Different experimental studies concluded that an auto activation of ComK is the only needed factor for bistability to occur in the expression of this protein [9], [11], [12]. In [9], Smits et. al discuss the factors that determine the required threshold for the activation of ComK and deduce that other transcription factors can raise or lower the threshold. Although many proteins are involved in the regulation of competence, there are two main proteins that play a major role. Süel et al. [13] propose a deterministic model driven by an additive noise to describe the dynamics of competence regulation. We use the reduced order Stochastic Differential Equation model (SDE) presented in [13] to develop a discrete stochastic model for competence. Calculating the probability and the expected time for entering and returning from competence, requires solving for the splitting probabilities and the first moment passage time. The problem of calculating the first passage time has been studied heavily in the literature for the stochastic difference equations, Fokker Planck equations and some special cases of the CME (separable kernels or single specie). For a detailed treatment of this topic see [14]–[17] and references therein. Researchers usually use Monte-Carlo simulations to calculate the distribution of the first passage time when working with he CME (e.g. see [18] and references therein). We propose in this work, an alternative approach that makes it possible to calculate the states in which the system will be as time evolves. The main idea here is to aggregate regions of the state space over which specie evolve into absorbing states. This technique is useful in analytically computing the distribution of the first passage time, by providing a way to deal with the infinite dimension of the state space over which the system evolves.
The contributions of this paper are threefold. First, it provides a new method to calculate exact probabilities of biological phenomena where transient behaviors such as competence, which is the topic we chose to study here, occur. Second, it shows how to calculate sensitivities of the probabilities of passing to the transient state with respect to the system's parameters. Third, it gives a methodology to calculate the expected time that it takes a cell to return from its transient state. All these methods can be used to analyze any biological system that has the characteristic of switching between two states, while staying for a while in the unstable state.
In this paper we start by describing the chemical reactions and the deterministic model. We then generate the Chemical Master Equation (CME) of our proposed discrete stochastic model. The CME characterizes the evolution of the probability density of the different discrete states. We simulate it using the Stochastic Simulation Algorithm (SSA) and show how the solution can be approximated using the Finite State Projection method (FSP). We then conduct a sensitivity analysis studying the effect that the various system parameters have on the probability with which a cell enters in competence. This analysis shows the usefulness of our proposed numerical method in analyzing the roles of the different affinity, transcription and degradation rates, etc., in driving the cellular switching (between competence and vegetative states in this case). Finally, we analyze the roles of these parameters in determining the expected time a cell stays in competence.
We introduce at the beginning the modeling techniques used to propose a set of equations that capture the behavior of interest. We then present our discrete stochastic Chemical Master Equation (CME) model, followed by the Stochastic Simulation Algorithm (SSA) used to approximate the solution of the CME. We proceed to present our Finite State Projection based method that makes it possible to analyze the CME exactly. We show how such a method can be tailored to answer many questions of biological interest.
Competence is a physiological state that enables cells to bind and internalize transforming DNA. This state is accompanied by blockage of the essential cell's functions, and since this state is driven by the transcriptional factor ComK, it is no surprise that ComK synthesis is subject to a number of finely tuned regulatory circuits [19]. The gene regulatory model for competence has been presented and described in [13]. Entrance of a cell in competence is controlled by a set of molecular interactions. Initially ComK and ComS are present in the cell at basal levels. The transcriptional factor ComK activates its own expression through positive feedback. The MecA complex is a multiprotein assembly that includes the ClpP-ClpC proteases. Bound to MecA, ComK is degraded under the action of the ClpP-ClpC proteases. In stressful environments, the level of ComS is high and that favors entrance into competence since ComS competes with ComK to bind to MecA. Inhibition of the binding of ComK to MecA by competitive binding with MecA-ComS allows a higher number of free ComK molecules to be present, which finally triggers the positive feedback that further raises the number of ComK molecules driving the cell in competence. This rise in the number of ComK is specific to competence. Once the number of ComK molecules reaches a certain level, it acts as an inhibitor for ComS through repression. The increase in the level of ComK will also favor the binding of MecA-ComK complex which degrades ComK through the ClpP-ClpC proteases, starting the return from competence. At this point ComS is below its basal level because of the aforementioned repression from ComK. The level of ComK starts to decrease by degradation. The degradation has two effects: (1) the decrease in the level of ComK will affect the transcriptional auto regulatory positive feedback loop of ComK, and (2) the absence of ComK in high levels, favors the synthesis of ComS by releasing the ComK-mediated ComS repression. This continues until the cell eventually exits the state of competence. The above mentioned molecular interactions are described [13] by the following chemical reactions:The rate equations describing the dynamics of the molecular reactions between the species model are the following:(1)(2)(3)(4)where K, S, , and are the concentrations of ComK,ComS, MecA, MecA-ComK and MecA-ComS respectively. We give in Table 1 the values and the description of each of the parameters in Eq. 1–4.
If one further assumes that the reactions of degradation of and are much faster than the other reactions, and can then be eliminated through time scale separation [13], [20] and the conservation law:giving the following reduced model for the dynamics of competence:(5)(6)whereand
In their paper Süel et al. [13] analyze the excitable dynamical system described above. They present a phase diagram where they study the nullclines and the vector field of the dynamical system. Their analysis gives insight about the vegetative and competent states analyzed in this work. As we already stated, under the same conditions some cells enter into competence while other cells do not. Entry in competence is a random event, and in order to properly model the cell's behavior, we need to include the effect of noise on the dynamics of competence. In their analysis Süel et al. [13] account for stochasticity by adding white gaussian noise terms in Eq. 6. This drives the excitable dynamical system presented in Eq. 5–6 into long excursions when the noise magnitude is large enough. These long excursions correspond to a high level of ComK indicating entry into a state of competence. The problem with this approach is that reaching a competent state is highly dependent on the magnitude of the additive noise. The dynamics of the system in Eq. 5–6 are such that if the initial number of molecules of ComK and ComS is in the neighborhood of the fixed point of the dynamical system described in Eq. 5–6, the number of molecules for both species will stay in the vicinity of that point without taking long excursions. If on the other hand, the number of molecules is driven beyond a threshold, the dynamical system in Eq. 5–6 will have a totally different behavior. The number of molecules of ComK will increase significantly because of ComK auto-activation through positive feedback; In other words, the cells will enter in competence. Here we would like to analyze the stochastic behavior of the dynamics of the competence regulatory circuit taking into account the internal noise in the environment of the cell without having a direct control on the magnitude of the noise driving the regulatory circuit. To do so, we model the stochasticity in the chemical reactions using the CME. We look at the problem at the molecular level and propose four reactions to model the system in Eq. 5–6. The four reactions are:(7)with the following reaction rates:These reactions will serve as the starting point for developing and simulating a discrete stochastic model for competence in the next section.
In order to compute the probability of entering into competence we use the CME to describe the stochastic chemical kinetics. Once we derive the CME, we simulate it using the Monte-Carlo based SSA. We then use the FSP method to obtain a finite dimensional solution to the infinite dimensional CME. In the CME, the state vectors indicate the number of molecules of each of the two species of interest: ComK and ComS. The CME describes the evolution of the probability that the number of molecules of each of the species has a given value. The dynamics of the evolution of the probability density vector are directly related to the chemical reactions. Starting from a number of molecules , the probability of being at molecules at time has the following dynamics:(8)where is the propensity vector and it represents the change that reaction will have on the number of molecules of each of the species. For example reaction increases ComK by one molecule and leaves the number of molecules of ComS unchanged so the propensity vector is , denotes the probability that the reaction will occur in the next infinitesimal time interval . Written in vector form, the CME becomes(9)where corresponds to the number of reactions that the species would go through. Let be a vector of the possible states of the system. Let be the corresponding vector of probabilities of the states in computed at time . evolves according to the equation(10)In general, may be infinite, resulting in an infinite dimensional system.
Getting the exact value for the solution to the CME is not generally an easy task. In this part we introduce the SSA that is normally used to simulate Eq. 9. The SSA is a Monte-Carlo based algorithm that generates sample paths for the underlying stochastic process. Gillespie introduced this algorithm in 1977 [21]. Reactions are modeled as a random event whose occurrence depends in a non linear manner on the number of molecules through the reaction rates. The algorithm can be summarized as follows:
What we described above is a a basic summary of the algorithm, interested readers are referred to [21] for more details.
The CME derived in Eq. 9 describes the evolution of the probability density vector of the number of molecules. Using SSA to get an estimate of the probability of entering into competence is easy to implement. However, a large number of simulations is required for a reasonably accurate estimate to be obtained. Aside from being time consuming, the algorithm has the drawback of lacking an accurate bound on the estimation error. In addition, analyzing the effect that different parameters have on the probability with which a cell enters in competence, requires the repetition of a large number of SSA simulations while changing those parameters of interest. This is numerically very costly. An alternative method in dealing with the CME is to compute an analytical expression for the probability of being in each state. The FSP method introduced in [22] provides a way to compute these probabilities. The probability density vector described in Eq. 9 allows molecules to evolve on an infinite lattice (Fig. 1) and therefore gives an infinite dimensional system. The idea behind FSP is to choose a suitable subset of the lattice in which one retains all the states and chemical reactions (transitions) found in the original system, while aggregating the remaining states in the lattice into one absorbing state. Transitions that drive the states outside the region are retained, while those that allow return to the selected finite region are deleted (see Fig. 1 for illustration). The finite state projection method gives the probability of being at any of the states inside the specified region at any point in time [22]. In this problem we are interested in finding the probability with which the pair (ComK,ComS) enters a region corresponding to the cell entering a state of competence. The sum of the probabilities of a cell being in and of the probability of being anywhere else in the state space has to equal one at all times. Moreover, if we divide the state space of the two proteins ComK and ComS in two regions, then the probability of the cell being inside the first region without ever leaving it, and the probability of leaving the first region once within a time should sum to one. These properties make FSP a very well suited numerical method to solve our problem.
In the finite model all the states outside the projection region are aggregated into one absorbing state: (see Fig. 1 for an illustration). The probability vector at time is given as in Eq. 10 by(11)where is an infinite matrix and is the initial distribution of the probabilities, that is a vector with infinite entries, where each entry corresponds to a probability with which the system starts with a given number of molecules. Using FSP we can project the infinite system in Eq. 11 into the following finite system:(12)In this case, A becomes a finite matrix, and is the finite vector of projected states. We build the finite matrix as followswhere and are the terms appearing in Eq. 8. If denotes the underlying stochastic process, gives the probability of being in any of the states listed in during the time , conditioned on the event of never leaving the inside region for any time . We can rewrite the probability as the conditional probability(13)where is the state to which the outside region is aggregated. Remember that is an absorbing state. The probability of being inside the region without ever leaving it during the interval and the probability of visiting once should sum to one. Therefore(14)Eq. 14 gives the probability of entering the region at least once within a time . The boundary of the region that is aggregated into the absorbing state , is chosen to include the states with a high number of ComK molecules. This indicates that the systems reaching the absorbing state corresponds to the cell being in a state of competence. Denoting by and by it can be seen that the probability of competence at time , , is given by(15)where is chosen so that the columns of the state transition matrix add up exactly to zero.
One advantage of having an analytical solution of the probability of competence is that we can use the solution to run a sensitivity analysis with respect to different model parameters. This makes it possible to shed light on the importance and roles that the different parts of the regulatory circuit play in reaching competence.
We start this section by introducing the equations we used to compute the sensitivity for the probability with respect to all the parameters. We then compare answers obtained by this method to estimates of sensitivities that we obtained using a finite difference method.
Recall that and suppose that we are interested in looking at the sensitivity of with respect to a parameter , which could be any of the parameters presented in Table 1. The entry in is given by , where , is an vector with in the entry and zero everywhere else. We have from Eq. 9 that . Letting take values in the set of parameters , and using the fact that , we get that where is defined to be . Similar equations hold for . The sensitivity of the probability with which a cell enters in competence evolves according to the following dynamical system:(16)Solving the above linear system, we obtain the sensitivity of the exit probability to all the parameters. We evaluate the solution at the nominal values given in Table 1. The results are reported in Table 2.
For comparison, we calculated the same terms computed above by using a finite difference method. The sensitivity of the probability of entering competence with respect to the various parameters is calculated according to the formula , where denotes the normalized sensitivity and denotes the nominal value of the parameter of interest. In order to change study the sensitivity to each parameter, we update the value with small steps using the equation below(17)
In summary, the sensitivity results presented in Table 2 are calculated using two different methods:
Method #1: We solve the double order system in Eq. 16. This results in more accurate answers but is more computationally expensive.
Method #2: We use the solutions for the original system describing the evolution of the probabilities of the states presented in Eq. 15 in addition to the numerical approximation method presented in Eq. 17 with . This method is less accurate than the first but is considerably faster to implement.
We study here the time it takes for a cell to return from a state of competence to its original vegetative state. We use once again the analytical solution of the CME to conduct this analysis. We use a similar concept to the one explained earlier, with the difference that in this case, we aggregate into an absorbing state the region of the state space that corresponds to the vegetative state, indicating that the cell returned from competence. We also assume that the cell starts from a state of competence and that it is allowed to return from that state, i.e., competent states are no longer absorbing in this case. Starting from competence corresponds to starting from a pair (ComK,ComS) that falls anywhere in the region . We assume that the cell can be at any state in equally likely. This assumption translates to setting the initial probability vector in a way that gives equal probability to all the states in . Return from competence is mapped to the region defined by . We set the initial probability vector to take the value at the entries corresponding to the states in and zero everywhere else. Here is the cardinality of the competence region in . Having defined a region to be the region in the state space corresponding to return from competence, we aggregate all the states of return from competence into one absorbing state . Hence, for the purpose of this calculation, once a trajectory ‘returns’ from competence, it cannot go back to it.
Having described the dynamics of the probability for return from competence in a similar manner to the description we had presented for the probability of entering in competence, we find the probability of returning from competence as a function of time by solving a set of differential equation just like we did earlier. We still need to deal with the infinite dimensions of the original model. For this purpose we add another absorbing state. This state is an aggregation of the region outside the finite state space that we consider, , into a single state . The finite state space is chosen so that the probability of reaching in the time interval of interest remains small. This small probability gives an upper bound on the approximation error due to the reduction of the infinite system into a finite one, as can be seen in the FSP algorithm [22]. Define to be the probability of returning from competence within . Denote by , the probability of returning from competence at time , and by , the probability of exiting to the outside region at time .
The system becomes:(18)
Now consider a partition of the interval as follows:We can approximate the expected value of return time as follows:(19)
We applied SSA to both the full model presented in Eq. 1–4, as well as to the reduced model presented in Eq. 7. We say that a cell entered in competence when the pair enter in the region . SSA simulations start from a number of molecules for and all runs simulate hours of molecular reactions. The initial number of molecules for ComK and ComS corresponds roughly to the mean steady state values of the reduced model. We are interested in studying the probability with which a cell enters in competence. For the return from competence analysis, we defined the region . A cell return trajectory is the one it takes when going from to (see Fig. 2 for illustration). We should point out that the boundaries of the region may be selected regardless of their shape.
In Fig. 3 we show seven different SSA runs, for hours each. It can be seen that two of the runs behave differently from the remaining five runs. The long excursions seen in Fig. 3 correspond to a high number of ComK molecules, i.e., the state of competence. In Fig. 4 we show one SSA run where both ComK and ComS concentrations were plotted. Competence is clear in this case, and it is detected by both the high level of ComK and the negative correlation between ComK and ComS corresponding to the negative feedback from ComK to ComS when the number of molecules of ComK is high. In SSA runs, we found that the cell entered in competence times, corresponding to an approximate probability . Using the Chernoff inequality, the accuracy in this case is described as , where and [23]. Using the bound on and Equation 19, we find an upper bound on the error in the calculation of .
Using the FSP based method as described in Eq. 15, we find that the probability of entering in competence at least once in 40 hours is , this probability is calculated with an error of no more than .
Now that we presented the SSA, and FSP method, we first use the SSA algorithm to compare the reduced model in Eq. 5–6 to the full model in Eq. 1–4 both presented in [13]. In order to do this, we simulate both models using SSA and compare the probability of entering in competence as the parameters presented in Table 1 were changed. We show the results for the parameters , and for demonstration purposes, but we note that the behavior of the full and reduced model were very close for all the parameters. We then compare the results given by the SSA and the FSP method, when applied to the reduced model. We show in Figs. 5 and 6, these results.
We show next the insights our numerical methods allowed us to have about how the molecules involved in competence, affect the time a cell spends in this state. In Fig. 7 we show how changing the parameter affects the time a cell stays in competence. This parameter corresponds to the saturation expression rate of the ComK positive feedback. The plot shows results obtained by both FSP and SSA. We can see that the plots exhibit similar behaviors, keeping in mind that such a calculation requires a lot more SSA simulations. In addition to giving more accurate results, the FSP approach allows us to combine multiple points from which we consider the cell as being in competence, while a different set of SSA simulations should be run for each different initial condition (starting number of molecules). Combining initial conditions is extremely useful in this case, since we care more about regions that the states go through than about specific points. It is not crucial to know the specific number of molecules of ComK or ComS when the cell is entering and returning from competence.
We saw earlier that increasing will increase the probability of cells entering competence. We now know that it will also keep the cell in competence for a longer time. Competence is an exhausting but occasionally necessary state for the cell. In this work we develop the CME accounting properly for the internal noise driving the competence switching dynamical system. The stochastic behavior of cell switching to competence has been studied in the literature. For example in their work, Süel et al. [13] account for the stochasticity by introducing an additive noise term to their model. The intensity of the noise and its distribution were parameters that are determined by the authors. In this work, we accounted for noise in its natural intrinsic form, eliminating therefore any controlled excitation of the excitable system.
We applied FSP to come up with an analytical solution, whereas other researchers always reverted to Monte-Carlo simulations, in their analysis. Finding an analytical solution made it possible for us to describe to a great extent the role of each of the molecules in driving cells into and out of competence. We discuss our results below.
We start by addressing the roles of the different expression and degradation rates in a cell entering competence. Fig. 5 shows that an increase in the saturating expression rate of ComK positive feedback increases the probability of entering in competence. Fig. 7 also shows that it makes returning from competence slower. Although Figs. 5 and 6 show that ComK and ComS have similar roles in driving a cell into and back from competence, Table 2 suggests that changes in ComS affected by the values of the expression and degradation rates of ComS affect the probability of entering and staying in competence more than changes in ComK affected by the values of . This leads to the expectation that the genetic circuits controlling ComS levels need to be much more sophisticated and complex than those regulating ComK in order to keep ComS concentration at specific values. Our normalized sensitivity analysis showed that increasing the basal expression rate and the saturating expression rate of ComK has an almost canceling effect to increasing the degradation rate of Comk as far as the probability of entering in competence is concerned. It also showed that the expression and degradation rates of ComS , had a similar canceling effect. This means that each of these molecules plays a dual role. As it turned out, while the expression rate of Comk drives the cell in competence, its degradation rate brings it back to its vegetative state. Similarly, a high concentration of ComS drives the cell in competence by competing over free MecA with ComK molecules, leaving more ComK molecules free. On the other hand, a decrease in ComS is necessary to return from competence as we will see next. This is true because low levels of ComS allow free MecA molecules to bind to ComK decreasing therefore the level of ComK molecules. We saw as well that high levels of ComK and ComS drive the cell into competence with probability 1. This is in agreement with experimental results reported in [24], where Leisner et al use an approximate SDE model in which they account for noise by introducing an additive gaussian noise term, in contrast to our approach which uses CME directly.
We now study the roles of the different molecules in the return from competence. Figs. 7 and 8 suggest that the degradation rate has a larger effect than when it comes to the expected time for which a cell stays in competence. We found similar results for and . This implies that once a cell is in a state of competence, the degradation rate acts fast bringing it back to its vegetative state. The degradation rate is faster than the rate at which the free molecules try to keep the cell in competence. Fig. 8 suggests that increasing the value of will decrease the time for which a cell stays in competence. We also know from Table 2 that an increase in diminishes the probability with which Bacillus subtilis enters in competence. Our calculations also show that an increase in has a similar effect to an increase in in the sense that they both decrease the probability with which a cell enters in competence and the expected time it takes for a cell to return form competence. Recall that is the degradation rate of ComK, and is the degradation rate of ComS. Also recall that whenever the number of ComS molecules is sufficiently small, more MecA molecules will be free to bind with ComK decreasing therefore the number of ComK molecules. Similarly, a higher ComK degradation rate, will lead to a decrease in the number of ComK molecules. A lower number of ComK molecules drive the cell back to its vegetative state and/or decreases its probability of entering in competence. This explains the similarity in the effect of and on the probability of entering competence and the expected return time.
In this paper we developed a discrete stochastic model for competence in Bacillus subtilis. We performed simulations of the model using Monte Carlo based SSA and verified that the reduced order model gave a valid approximation of the full model. We then applied the recently developed FSP method to the reduced model and computed the probability of competence, where competence was defined in terms of a trajectory leaving a pre-defined region of the state space. Having the analytical solution, we were able to conduct a sensitivity analysis of the probability with which a cell enters in competence as the model parameters vary. We were also able to compute interesting terms such as the expected time it takes for a cell to return from competence.
This paper presented numerical methods that are applicable to many biological systems that exhibit a transient switching behavior. These methods were shown to be very useful in studying the genetic circuit regulating competence in a bacteria, and in answering questions about exact probabilities of stochastic events in this bistable biological behavior. They were also useful in studying sensitivities of these probabilities when expression rates, degradation rates, repression rates or activation rates of proteins were changed. Finally, the methods introduced in this paper showed how to calculate the expected time for return from transient states. Many other terms characterizing different transient physiological behaviors, such as the number of molecules that are most likely to enter in the transient states, and the return trajectories that are most likely to be taken can be computed using similar approaches to the one discussed here. Our approach should be easily extendible to analyze many biological system exhibiting a bistable switching behavior.
|
10.1371/journal.pcbi.1000402 | Active Dendrites Enhance Neuronal Dynamic Range | Since the first experimental evidences of active conductances in dendrites, most neurons have been shown to exhibit dendritic excitability through the expression of a variety of voltage-gated ion channels. However, despite experimental and theoretical efforts undertaken in the past decades, the role of this excitability for some kind of dendritic computation has remained elusive. Here we show that, owing to very general properties of excitable media, the average output of a model of an active dendritic tree is a highly non-linear function of its afferent rate, attaining extremely large dynamic ranges (above 50 dB). Moreover, the model yields double-sigmoid response functions as experimentally observed in retinal ganglion cells. We claim that enhancement of dynamic range is the primary functional role of active dendritic conductances. We predict that neurons with larger dendritic trees should have larger dynamic range and that blocking of active conductances should lead to a decrease in dynamic range.
| Most neurons present cellular tree-like extensions known as dendrites, which receive input signals from synapses with other cells. Some neurons have very large and impressive dendritic arbors. What is the function of such elaborate and costly structures? The functional role of dendrites is not obvious because, if dendrites were an electrical passive medium, then signals from their periphery could not influence the neuron output activity. Dendrites, however, are not passive, but rather active media that amplify and support pulses (dendritic spikes). These voltage pulses do not simply add but can also annihilate each other when they collide. To understand the net effect of the complex interactions among dendritic spikes under massive synaptic input, here we examine a computational model of excitable dendritic trees. We show that, in contrast to passive trees, they have a very large dynamic range, which implies a greater capacity of the neuron to distinguish among the widely different intensities of input which it receives. Our results provide an explanation to the concentration invariance property observed in olfactory processing, due to the very similar response to different inputs. In addition, our modeling approach also suggests a microscopic neural basis for the century old psychophysical laws.
| One of the distinctive features of many neurons is the presence of extensive dendritic trees. Much experimental and computational work has been devoted to the description of morphologic and dynamic aspects of these neural processes [1], in special after the discovery of dendritic active conductances [2]–[4]. Several proposals have been made about possible computational functions associated to active dendrites, such as the implementation of biological logic gates and coincidence detectors [5],[6], learning signaling via dendritic spikes [7] or an increase in the learning capacity of the neuron [8]. However, it is not clear whether such mechanisms are robust in face of the noisy and spatially distributed character of incoming synaptic input, as well as the large variability in morphology and dendritic sizes.
Here we propose to view the dendritic tree not as a computational device, an exquisitely designed “neural microchip” [6] whose function could be dependent on an improbable fine tuning of biological parameters (such as delay constants, arborization size, etc), but rather as a spatially extended excitable system [9] whose robust collective properties may have been progressively exapted to perform other biological functions. Our intention is to provide a simpler hypothesis about the functional role of active dendrites, which could be experimentally tested against other proposals.
We study a model where the excitable dynamics is simple, but the dendritic topology is faithfully reproduced by means of a binary tree with a large number of excitable branchlets. Most importantly, branchlets are activated stochastically (at some rate), so that the effects of the nonlinear interactions among dendritic spikes can be assessed. We study how the geometry of such a spatially extended excitable system boosts its ability to perform non-linear signal processing on incoming stimuli. We show that excitable trees naturally exhibit large dynamic ranges — above 50 dB. In other words, the neuron could handle five orders of magnitude of stimulus intensity, even in the absence of adaptive mechanisms. This performance is one hundred times better than what was previously observed in other network topologies [10],[11].
Such a high performance seems to be characteristic of branched (tree) structures. We believe that these findings provide important clues about the possible functional roles of active dendrites, thus providing a theoretical background [4] on the cooperative behavior of interacting branchlets. We observe in the model the occurrence of dendritic spikes similar to those already observed experimentally and recently related to synaptic plasticity [7]. Here, however, such spikes are just an inevitable consequence of the excitable dynamics and we propose that even dendritic trees without important plasticity phenomena (like those of some sensory neurons) could benefit from active dendrites from the point of view of enlargement of its operational range.
Our results also suggest that, under continuous synaptic bombardment, dendritic spikes could be responsible for another unintended prediction of the model, namely, that the neuron transfer function needs not to be simply a Hill-like saturating curve; rather, a double-sigmoid behavior may appear (as observed experimentally in retinal ganglion cells [12]). The model further predicts that:
So, why do neurons have active dendrites? As a short answer, we propose that neurons are the only body cells with large dendrites because they need to work with a large stimulus range. Owing to the enormous number of afferent synapses and the large variability of input rates, highly arborized and active dendrites are crucial to enhance the dynamic range of a neuron, in a way not accounted for by passive cable theory and biophysical neuron models with few compartments (reduced models) [13]. Other phenomena, such as backpropagating spikes, could have been later exapted to more complex functional roles. One should, however, consider first a generic property of extended excitable media: that, due to the creation and annihilation of non-linear pulses, the input-output transfer function of such media is necessarily highly non-linear, with a very large dynamic range as compared with that of a passive medium.
In computational neuroscience, the behavior of an active neuronal membrane traditionally is modeled by coupled differential equations which represent the dynamics of its electric potential and gating variables related to the ionic conductances. This modeling strategy was then further extended by detailing the dendritic tuft through a compartmental approach [14]. Motivated by the abundant evidence that dendrites have active ion channels that can support non-linear summation and dendritic spikes [1],[6], this line of research currently aims at examining the possibility that these extensive tree-shaped neuronal regions may be the stage for some kind of “dendritic computation” [4],[5],[15].
Many efforts within this framework of biophysical modeling have been devoted to unveiling the conditions under which the regenerative properties of dendritic active channels may be unleashed to generate a nonlinear excitation (e.g. at the level of a single spine [16] or upon temporally synchronized and locally strong input at the level of a branchlet [17]). Nonlinear cable theory can further help predict whether and how a single dendritic spike will propagate along the branches, for instance highlighting the relative importance of a given channel type for the propagation of action potentials [18]. Detailed biophysical models also correctly predicts e.g. that two counter-propagating dendritic spikes annihilate each other upon collision [19],[20] (instead of summing), but this is true for most – if not all – extended excitable media. However, at the present state of the art of neuronal simulations, biophysical modeling may not necessarily be the approach best suited for addressing the much more difficult question of what happens when many dendritic spikes interact, specially in a more natural scenario where they would be continuously created at different points of the dendritic tree at some stochastic rate.
Understanding the net effect of the creation and annihilation of dendritic nonlinear excitations under massive spatio-temporal patterns of synaptic input requires 1) knowledge of the key properties of these excitations and their interactions (which cable theory gives us) and 2) a theoretical framework which addresses the resulting collective behavior. We therefore borrow from cable theory the facts that dendritic spikes may (or may not) be created by integrated synaptic input at some branchlets, then may (or may not) propagate to neighboring branchlets, and annihilate upon collision owing to refractoriness. Then, by employing a simplified excitable model for each branchlet, but a realistic multicompartment dendritic tree, we are able to focus on their collective behavior and to cast the dynamics of the dendritic tuft into the framework of extended excitable media, where both numerical and theoretical approaches have been successfully applied [9]–[11], [21]–[27].
Conventional wisdom in computational neuroscience is that in the limit of a very large number of compartments the model would be physically accurate. But in this same limit, conventional wisdom in statistical physics (say, renormalization group arguments) tells us that collective behaviors should be very weakly dependent on the detailed modeling of the basic (compartmental) unit [28]. Macroscopic properties of extended media would rather depend more strongly on dimensionality, network topology, symmetries, presence of parameter randomness (disorder), noise, boundary conditions etc. Therefore modeling should concentrate efforts on these more decisive aspects, the use of simple excitable dynamics for the elementary units being justified as a first approximation.
We define the apical activity as the number of excitations ( states) produced at the site, averaged over a large time window (104 time steps and five realizations, unless otherwise stated). In the following we will be interested in understanding the function , which is somehow analogous to the neuron frequency versus injected current curves studied in the neuroscience literature. We suppose that, in the absence of lateral inhibition, the neuron firing frequency produced at the axonal trigger zone will be proportional to the apical activity , which is assumed by some biophysical models [30] and supported by recent experimental evidence in the Drosophila olfactory system [31].
For readers familiar with statistical physics models we observe that is the order parameter and is an external field that drives the system to an active state with . Our model is an out-of-equilibrium system with one absorbing state [32]. This means that, in the absence of external drive () the dynamics eventually takes the system to a global resting (quiescent) state from which it cannot escape without further external stimulation. In biological terms, this simply means that our dendritic tree will not show spontaneous dendritic spikes without external synaptic input and any activity in the tree will eventually die if is turned to zero.
Dendritic trees are responsible for processing incoming stimuli which impinge continuously on the many synaptic buttons spread on the dendrites (a single olfactory mitral cell can have around 30,000 synapses, whereas cerebellar Purkinje cells have around 200,000 synapses). Of course these numbers vary also between individual cells of the same type. So, we first consider a classical question asked (and not clearly answered) in the literature: given a constant activation in cells with different arbor sizes, will they fire at very different levels [33]? The answer is not obvious since they may have a huge difference of absolute number of synapses and branchlets and we could have the prejudice that cells that have more synapses should fire more easily (or at least need to implement some homeostatic mechanism for controlling their firing rate).
The answer provided by our model is very interesting: for low excitation rate h, the output increases linearly with the number of branchlets, so that having a large branched tree is indeed important to amplify very weak signals (see Fig. 1C). In this context there is a clear reason for a neuron to maintain a costly number of branchlets. However, for moderate and high activation levels, the activity depends very weakly on (it grows sub-logarithmically with N, see Fig. 1C for , say, larger than 5,000). That is, in this regime the output reflects, in an almost size-independent way, mostly the Poisson rate h, not the absolute number of branchlets activated on the tree.
Large dendritic arbors therefore aid the detection of weak stimuli, but for higher activation levels (i.e. higher imbalance between excitatory and inhibitory signals) all the neurons code in a similar way the activation rate , irrespective of their arbor size. Note that this “size invariance” is an intrinsic property of the excitable tree, not based on any homeostatic regulatory mechanism [19],[34]. This sublogarithmic dependence of on means that neurons function as reliable transductors for the signal : the specific number of branchlets, developmental defects, or asymmetries of the dendritic tuft have only a secondary effect in the global neuron functioning.
Given that the cell output depends weakly on , now we turn our attention to how depends on the activation rate . Note that not much modeling work has been done on addressing the collective activity of the dendritic tree subjected to extensive and distributed synaptic input [35],[36], particularly as the activation rate is varied. However, this is one of the simplest questions one may ask regarding dendritic signal processing. In particular, studies with models where the whole dendritic tree is reduced to a small number of compartments (reduced compartmental models [37]) can hardly address this issue, since the complex spatio-temporal information of the tree activation is lost by definition.
As is well known, the average firing rate dependence on stimulus rate of several cells has a saturating aspect like that of Fig. 2A. Our cell presented a similar behavior (Fig. 2B), although of course it is not the simple Hill function usually employed to fit experimental data. Indeed, for some values of axial transmission pλ, we saw an unexpected double-sigmoid behavior (see below).
A possible critique to our modeling approach is that it lacks biological realism. We notice that this is only true at the level of the biophysical dynamics of each compartment, but we believe that the idealization of such compartment as a generic excitable element (a cyclic automaton) is immaterial. This has already been demonstrated in studies of the dynamic range of networks composed by cellular automata, non-linear discrete time maps, nonlinear differential equations and conductance-based models (Hodgkin-Huxley compartments) [22],[27], as well as in a biophysically detailed model of the vertebrate retina [38].
Our model has realistic biological aspects not reproduced by most works in computational neuroscience with detailed biophysics:
Notwithstanding the fact that artificial input protocols like punctual current injection are useful for comparison with experimental measurements [39], we believe that spatio-temporal Poisson activation is a step toward a more realistic modeling of the dendritic arbor dynamics under natural circumstances [19],[35],[36].
As can be viewed in Fig. 3, large active dendritic trees perform strong signal compression, which is the ability of coding many orders of magnitude of stimulus intensity through only one decade of output frequency. This question is particularly important in sensory processing, where many orders of magnitude of stimulus intensity are present. Interestingly, olfactory glomeruli, constituted primarily by large active dendrites of mitral cells in vertebrates and dendrites of principal cells in insects have large dynamic range [40]–[42]. We conjecture that a similar situation occurs in the problem of fine motor control and sensory-motor integration in the cerebellum [43], which also involves the necessity of handling sensory-motor feedback signals varying by orders of magnitude. In correspondence to our hypothesis, Purkinje cells, which are involved in these tasks, have indeed enormous active dendritic arbors [13],[44].
Previous work [10], [11], [21]–[27] has shown that the non-linear summation of spikes enhance the dynamic range of excitable media. The tree topology, however, has not been studied in these works. Surprisingly, we found that its performance is largely superior to the others. This motivates the proposal, first made here (to the best of our knowledge), that the main functional role of active dendrites is to enlarge the cell dynamic range.
As a particular application, we discuss now the case of the dynamic range of olfactory glomeruli. Recent results for second-order projection neurons of the Drosophila melanogaster antennal lobe clearly exhibit strong weak-stimulus amplification and enhanced dynamic range as compared to olfactory receptor neurons (ORNs) [42]. To account for this observation, we can interpret our model as representing a Drosophila principal cell (analogous to a mitral cell) inside the glomerulus. Also, the signal propagation from ORN axons to principal cell dendrites and the proportionality between apical activity and somatic firing measured by Root et al. in the Drosophila is compatible with our identification of with the somatic neuron response [31] in this particular case. These authors show that it is mainly the ORN activity that drives the projection neuron firing rate, the isolated effect of synapses from interneurons (excitatory and inhibitory) being not sufficient to induce spikes and having mostly a modulatory role.
Of course, in the case of other biological systems like the mammalian olfactory bulb (where strong lateral inhibition occurs) or pyramidal cells, the identification of with the somatic firing rate is problematic, but we claim that the model is still useful for understanding of the large dynamic range (as measured by Calcium fluorescence) observed in the neuronal tuft [40],[41].
It is important to notice that large dynamic ranges as observed here means that the output varies slowly with the input. Therefore, if experiments are done over only one or two orders of magnitude of stimulus intensity (10–20 dB), the observed effect could be confounded with an almost constant response. This may be an alternative explanation for the concentration invariance property observed in olfactory processing [45].
Another important prediction of our model is that dendritic size (and the respective number of branchlets and synapses) has a weak effect on the apical activation, being important mostly in the small excitation regime. It is mainly the branchlet activation rate h, not the total number of branchlets, that controls the apical rate F. This is a desirable robustness property since there is a high variability of dendritic size and spine density within a neuron population and along time in the same neuron.
Whichever function one wishes to assign to active dendrites, it must be fault tolerant in relation to gross dendrite morphology, branchlet excitability and synaptic density, which vary with age and time: for example, 30% of spine surface retracts in hippocampal neurons over the rat estrous cycle [46]. Due to the sublogarithmic dependence of on (see also the Model Robustness section), our model demonstrates that such gross independence from branchlet number, detailed branchlet dynamics, dendritic axial conductance and tree morphology is possible, and that enhancement of dynamic range is one of the most visible properties of these excitable trees.
Double-sigmoid response functions have been reported recently for retinal ganglion cells of the mouse [12]. This unusual shape contrasts with the standard Hill fitting function. One wonders whether the habit of fitting Hill functions to data could have prevented further double-sigmoid curves from having been reported in the literature.
It is very interesting that such double sigmoid behavior is a distinctive feature of our model in a certain range of parameter space. Can we interpret the findings on retinal ganglionar cells in terms of our simplified model of dendritic response? Ganglionar cells have dendritic arbors but their size is small compared to, say, mitral cells or our typical model with branching order around .
However, in a structural analogy between the visual and olfactory systems, Shepherd proposed that some ganglionar cells are the retinal equivalent of mitral cells [44]. Here we pursue this analogy and suggest that the ganglionar dendritic arbor plus the retinal cells connected to it by gap junctions (electrical synapses) can be viewed as an extended active tree similar to the one studied here, with a large effective .
We show in Fig. 2D that an appropriate choice of the model parameters can lead to a response function which fits the experimental data. Of course, the quantitative fit, although good, is not the important message, but the qualitative one: that double-sigmoid response functions can appear solely due to the tree topology, without invoking any secondary activation processes or complicated mechanisms to produce the unusual shape.
What is the physical origin of the double sigmoids in our model? We believe that it is related to the two different modes of activation of the apical site. The first one is the direct excitation due to its local rate. This direct excitation, if large, drives the system to its maximum firing rate, which scales with the inverse of its refractory period. This mechanism would be responsible for the saturation in the right side of (region of large ), see Fig. 2B.
But the apical site also receives signals from its extended dendritic tree, which is very sensitive to small activity (extending the curve to the small regime). However, it is plausible that the tree excitability saturates for a smaller frequency, due to the complicated interations between the spikes in the tree. So, we conjecture that the first sigmoid represents a bottleneck effect related to saturation in the flux of the activity along the subtrees connected to the apical site. Indeed, this is compatible with the observation that if we disconnect the apical site from the dendritic tree (, Fig. 2B), the double sigmoidal behavior disappears and only the second (large ) sigmoid is maintained. Of course, a more detailed analysis of the origin of the first sigmoid is needed.
We also observed curves with three sigmoids (see Model Robustness section), but postpone the discussion of these results to future works. We only note here that the plateau in these curves could also be related to the concentration invariance reported for olfactory systems [45].
As can be seen in Fig. 2B, some response curves in our model can present an unusual shape, with curves for higher probability of axial transmission falling below curves for lower . How can more efficient trees present a response below less efficient ones for the same level?
This question can be answered by looking at Fig. 4A, where we plot a family of curves for fixed . For some (intermediate) values of , this curve is non-monotonic, suggesting a kind of resonance through which activity in the primary dendrite is maximized for an optimal coupling among sites all over the tree. Why is this so?
Note that, on the one hand, for low enough , excitations created in distal sites may not arrive at the primary site due to propagation failure. For too strong coupling, on the other hand, the topology of the tree leads to a dynamic screening of the primary dendrite: backward propagation of activity (backspikes) effectively can block forward propagation of incoming signals, as shown in Fig. 4B. Activity is therefore maximized at some intermediate value of coupling. We called this phenomenon “screening resonance”. That such screening resonance indeed depends on backspikes is confirmed by an asymmetrical propagation variant of the model (see Model Robustness section). As backpropagation goes to zero, the crossing between curves disappears (Figs. 4C and 5C).
The transmission probability accounts for the joint effects of membrane axial conductance and density of regenerative ionic channels (Na+, Ca2+, NMDA etc). A possible experiment to test whether this screening resonance indeed exits could involve the manipulation of the density (or efficiency) of those channels in the dendritic tuft: the model predicts that more excitable trees may present lower activity than less excitable ones due to resonant annihilation of dendritic spikes.
As discussed above, annihilation due to collision of dendritic spikes is the central mechanism in our model behind both the dynamic range enhancement (by preventing the tree response to be proportional to the rate ) and the screening resonance phenomenon (by blocking forward-propagating dendritic spikes with backward-propagating ones).
With rare exceptions [19],[20], the fact that nonlinear summation often implies spike annihilation has been somewhat underrated in the literature. Recent simulations with biophysical compartments show the propagation and collision of dendritic spikes [19],[20]. To fully evaluate our ideas, one should examine better this phenomenon in in vitro dendrites. The computational results suggest the following simple experimental tests:
One consequence of spike annihilation is that under moderate stimulation backspikes will fail to reach more distal branches, owing to collisions with forward-propagating dendritic spikes and/or refractory branches [19],[47]. Indeed, we have observed this phenomenon in our model.
This is compatible with recent observations that backspikes are strongly attenuated in the presence of synaptic input in medial superior olive principal neurons [48]. So, the use of somatic backspikes as a backpropagating signal under massive synaptic input seems to be problematic. Somatic backspikes show up naturally in excitable trees but plays no functional role here.
We conjecture that somatic backspikes may be epiphenomena or perhaps, if they have a functional role in learning processes, it is a recent evolutionary exaptation from previous robust functions like signal amplification by dendritic spikes. This can be tested: our model predicts that active dendrites will be found even in neurons without any plasticity or learning phenomena.
Our modeling approach also suggests a microscopic (neural) basis for Stevens law of psychophysics [49],[50], which states that the perception of stimulus intensity grows as a power law . In a previous work with disordered networks [10], by assuming a linear relationship between psychophysical perception and the network activity, we have found a Stevens-like exponent for the input-output function of excitable media with value . For planar networks we found [26]. Here we found for the dendritic tree architecture that the Stevens exponent is very small ( or even 0.1 for large trees with ), which means that the response function could be confounded with a logarithmic (Weber-Fechner) law [49]. Of course, the macroscopic psychophysical law would be a convolution of all these non-linear transfer functions between the sensory periphery and the final processing (psychological) stage.
What our model shows is that any excitable medium naturally presents a nonlinear input-output response with exponent , that is, large dynamic range, and that perceptual “psychophysical laws” could be a very early phenomenon in evolution. A simple precondition is that the sensory network should have an excitable spatially extended dynamics, like the one already found in bacterial chemotaxis channel networks, for example [51],[52].
In our model, variable branchlet diameter and size is described by a spatial dependence and disorder in pλ. We do not expect the results concerning the dynamic range to change qualitatively with this type of generalization. The same model robustness appears for changes in the refractory time and the use of continuous dynamical variables (maps or differential equations). This latter property has already been demonstrated in multilevel modeling studies which used cellular automata and nonlinear differential equations to describe the neuronal excitable elements [22],[27],[28]. We now explicitly show the results for three variants of the model in order to address the robustness of the dynamic range enhancement.
Several detailed biophysical models of dendritic trees have already been presented in the literature, but we are not aware of studies confirming the enlargement of the dynamic range by active dendrites in such arbors. To see this effect, it is necessary that such models incorporate inputs distributed along the full dendritic tree, and that the branchlet activation rate be varied by orders of magnitude.
We believe that biophysical multi-compartmental models (with a large number of branchlets) seeking to probe the robustness of our results would be most welcome. In particular, they would be able to address the effect of post-synaptic potentials (PSPs, both excitatory and inhibitory) which manage to generate somatic – but not dendritic – spikes despite the presence of active channels in the dendrites (a phenomenon which might be artificially adapted to our model, but for which biophysical models are better equipped). Also, the modulatory effect of such subthreshold PSPs and other passive phenomena are better studied in biophysical simulations.
Other future tasks will be the study of the dendritic response due to non-Poisson input distributions (say, noise), correlated input on the arbor, time-dependent inputs, asymmetric dendritic trees etc. We believe that new signal processing features may appear, but the dynamic range enlargement and sensitivity enhancement by active dynamics will continue to be present.
Why do neurons have active channels in extensive dendritic trees? Our proposal is that active large dendrites are able to detect and amplify very weak signals and, at the same time, saturate slowly for stronger tree activity. This universal “dynamic range” problem, related to the trade-off between sensitivity and saturation of signal processing, is important both for individual neurons, large neural networks, whole sensory organs and organisms. We conjecture that the large dynamic ranges found in neurons with active dendritic arbors could even help to explain macroscopic psychophysical laws, providing a neural account for the century old findings of Fechner, Weber and Stevens [10],[49],[53].
|
10.1371/journal.pcbi.1004426 | Discovery of Drug Synergies in Gastric Cancer Cells Predicted by Logical Modeling | Discovery of efficient anti-cancer drug combinations is a major challenge, since experimental testing of all possible combinations is clearly impossible. Recent efforts to computationally predict drug combination responses retain this experimental search space, as model definitions typically rely on extensive drug perturbation data. We developed a dynamical model representing a cell fate decision network in the AGS gastric cancer cell line, relying on background knowledge extracted from literature and databases. We defined a set of logical equations recapitulating AGS data observed in cells in their baseline proliferative state. Using the modeling software GINsim, model reduction and simulation compression techniques were applied to cope with the vast state space of large logical models and enable simulations of pairwise applications of specific signaling inhibitory chemical substances. Our simulations predicted synergistic growth inhibitory action of five combinations from a total of 21 possible pairs. Four of the predicted synergies were confirmed in AGS cell growth real-time assays, including known effects of combined MEK-AKT or MEK-PI3K inhibitions, along with novel synergistic effects of combined TAK1-AKT or TAK1-PI3K inhibitions. Our strategy reduces the dependence on a priori drug perturbation experimentation for well-characterized signaling networks, by demonstrating that a model predictive of combinatorial drug effects can be inferred from background knowledge on unperturbed and proliferating cancer cells. Our modeling approach can thus contribute to preclinical discovery of efficient anticancer drug combinations, and thereby to development of strategies to tailor treatment to individual cancer patients.
| Fighting cancer with combinations of drugs increases success of treatment. However, due to the large number of drugs and tumor variants, it remains a tremendous challenge to identify efficient combinations. To illustrate this, a set of 150 drugs corresponds to more than 10.000 possible pairwise drug combinations. Experimental testing of all possibilities is clearly impossible. We have developed a computational model that allows us to identify presumably effective combinations, and that simultaneously suggests combinations likely to be without effect. The model is based on specific cancer cell biomarkers obtained from unperturbed cancerous cells, and is then used to perform extensive automated logical reasoning. Laboratory testing of drug response predictions confirmed results for 20 of 21 drug combinations, including four of five drug pairs predicted to synergistically inhibit growth. Our approach is relevant to preclinical discovery of efficient anticancer drug combinations, and thus for the development of strategies to tailor treatment to individual cancer patients.
| It has long been envisaged that future anticancer treatment will adopt combinatorial approaches, in which several specific anti-cancer drugs together target multiple robustness features or weaknesses of a specific tumor [1–3]. The effectiveness of combinatorial anti-cancer treatments can be further maximized by exploiting synergistic drug actions, meaning that different drugs administered together exhibit a potentiated effect compared to the individual drugs. Drug synergy is attractive because it allows for a significant reduction in the dosage of the individual drugs, while retaining the desired effect. Synergies therefore hold the potential to increase treatment efficacy without pushing single drug doses to levels where they lead to adverse reactions. Hence, synergies identified in preclinical studies represent interesting candidates for further characterization in cancer models and clinical trials.
Current efforts to identify beneficial combinatorial anti-cancer therapies typically rely on large-scale experimental perturbation data, either for deciding on specific patient treatment [4], or for pre-clinical pipelines to suggest new drug combinations [5–8]. This work, however, faces challenges posed by the large search space that needs to be supported by experimental data, making systematic searches for efficient combinations challenging. Moreover, the number of conditions for testing dramatically increases when considering higher-order combinations, multiple drug dosages, temporal optimization of drug administration, and diversity of cancer cell types and patients. Thus, workarounds must be sought to reduce the experimental search space of drug combinations and their application modes in order to obtain a qualified repertoire of combination therapies for clinical trials, and ultimately to support delivery of personalized treatment.
Computational models are increasingly used to predict drug effects [6,9], with the aim to rationalize and economize the experimental bottleneck. In order to enable substantial reduction of the number of relevant conditions that need to be tested, such models would ideally be constructed without the need for massive experimental drug perturbation data. Approaches where the formulation of predictive models can be based on molecular data from unperturbed cancer cells are therefore attractive.
We decided to focus on Boolean and multilevel logical models, as they enable a relatively straightforward formalization of the causalities embedded in molecular networks, such as signal transduction and gene regulatory networks. Moreover, logical model simulations can be used to automate reasoning on network dynamics, even with scarce knowledge of kinetic parameters [10–15], and have been used to describe and predict the behavior of molecular networks affected in human disease [13,14]. Such modeling efforts have contributed to the understanding of mechanisms underlying growth factor induced signaling in cancer cells and the selection of candidate target proteins for novel anti-cancer treatment [16–23]. While previous studies have demonstrated the power of logical models to predict single drug actions, we extend the use of logical modeling to predict effects of combinatorial inhibition of two or more signal transduction components.
We report the construction of a logical model encompassing molecular mechanisms central to controlling cellular growth of the gastric adenocarcinoma cell line AGS. After an initial assembly of a comprehensive signaling and regulatory network from general signal transduction knowledge, the logical rules associated with each of the 75 model components were refined using baseline data obtained from actively growing AGS cells. The resulting logical model was used to assess drug synergy potential among 21 pairwise combinations of seven chemical inhibitors, each targeting a specific signaling component. Model simulations suggested five combinations of inhibitors to be synergistic, four of which could subsequently be confirmed in cell growth experiments. Importantly, none of the combinations predicted by the model to be non-synergistic displayed synergistic growth inhibitory effects in our cellular assays, i.e. no false negatives were observed. Our results demonstrate that our logical model, constructed without the use of initial large-scale inhibitor perturbation data, recapitulates key molecular regulatory mechanisms underlying growth of AGS cells in a manner that allows successful prediction of the synergistic effect of inhibitor combinations in experimental cell cultures. Guided by the model, we identified two established synergistic drug interactions and discovered two synergies not previously reported.
In order to discover combinatorial drug treatments synergistically exerting inhibition of cancer cell growth, we developed a workflow combining computational and experimental analyses to predict and validate drug synergies (Fig 1).
Our modeling procedure integrates a priori biological knowledge on intracellular signaling pathways with baseline data from AGS gastric adenocarcinoma cells. The design principles of our analysis are guided by the premise that growth of cancerous cells is largely driven by mechanisms which enable these cells to exploit a wide range of growth promoting signals from the environment. This aspect of intrinsic, sustained multifactor-driven cancer proliferation [1] is accommodated by constructing the regulatory network as a self-contained model: we include only nodes that are regulated by other nodes in the model. The chosen design avoids the need to model effects of specific growth factor receptors, considering instead the integrated responses from a multitude of growth promoting stimuli, as observed when assessing the activity of signaling entities (proteins and genes) included in the model. It follows from this that the de facto growth promoting configuration of such a self-contained model can be established by observing baseline biomarkers measured in the cancer cells.
After a model reduction step, where nodes and logics pertaining to drug targets and phenotypic outputs are retained, the model is used for exhaustive simulations of the effect of pairwise node inhibitions using seven known chemical inhibitors. Finally, the growth inhibitory effects of these drug combinations on AGS cells are tested experimentally.
In order to assess combinations of inhibitions for synergy, we focused on the systematic inhibition of seven model nodes and their 21 pairwise combinations. These seven nodes (labelled with thick borders in Fig 2) were chosen because potent and specific chemical inhibitors were available for targeting the corresponding protein kinases in biological experiments (Table 1).
Using an asynchronous updating policy (see Materials and Methods), we simulated the effect of chemical inhibitions by forcing the state of specifically targeted model nodes to be 0 (inactive), and then computing the resulting attractor. Each inhibition of single nodes or pairs of nodes led to a unique attractor. In a few cases the system reached a complex attractor, in which a subset of states is traversed repeatedly (see Materials and Methods and S1 Text). The computation of potential complex attractors is challenging because of the combinatorial explosion of states for large logical models. To cope with this problem, we used a model reduction method to obtain a compressed model preserving the selected drug targets, and compacted the state transition graphs in a hierarchical manner (see Materials and Methods and [14]). The reduced logical model (see Fig 3 and S3 Dataset) was obtained by iteratively removing components not targeted by drugs, and was sufficiently small to allow exhaustive asynchronous simulations and thorough characterization of both stable states and complex attractors, thereby enabling the analysis of all single and pairs of inhibitions.
To ease interpretation, we defined the overall response growth, by subtracting the value of Antisurvival from the value of Prosurvival readout nodes (each multi-valued with state ranging from 0 to 3), with a value range from -3 to +3. If the attractor contained a unique stable state, the computation of growth was straightforward. In the case of complex attractors we used the mean values of the difference Prosurvival–Antisurvival over all states belonging to the attractor. We inferred synergy whenever the combination of two inhibitors produced a value for growth lower than the smallest value of the inhibitors individually:
growth (perturbation1 & perturbation2) < Min (growth (perturbation1), growth (perturbation2)),
where perturbationN is the perturbation of component N.
For example, growth (perturbationMEK & perturbationAKT) = 0.5; which is a value lower than observed with perturbations of either MEK or AKT: growth(perturbationMEK) = 1.5; growth (perturbationAKT) = 2.
The simulations predicted five synergistic combinations (<25% of the 21 possible pairs). Three of these combinations involve MEK, together with PI3K, AKT or p38. The two remaining synergies involve TAK1 with either PI3K or AKT (Fig 4).
To assess the validity of our model predictions, a real-time cell assay was used to test chemical inhibitors of the seven proteins (Table 1) for their ability to limit AGS cell growth in single and combinatorial formulations.
The effect of chemical inhibitors was analyzed using a strategy based on Loewe’s definition of synergy [28], which states that a synergistic interaction performs better than the expected additive effect observed when an inhibitor is combined with itself in a ‘zero-interaction’ experiment. To quantify synergistic interactions, a combinatorial index (CI) was calculated [29], based on growth measured 48 hours after adding inhibitors. CI values range from zero to infinity, and values below 1 indicate synergistic interactions.
Four of the five synergies predicted by our logical model were confirmed experimentally, with CI values well below 0.5, which indicates strong synergy. Indeed, a profound effect on AGS cell growth was found when MEK or TAK1 inhibitors were combined with PI3K or AKT inhibitors. The corresponding growth curves (Fig 5) indicate that cell growth in the presence of two inhibitors combined, each at half their GI50 concentrations (purple curves) is markedly lower than growth in the presence of either inhibitor alone at its full GI50 concentration (green and blue curves). In contrast, the combination of MEK and p38 inhibition could not be confirmed in the cell growth experiments. Importantly, we observed no false negative predictions, meaning that the remaining inhibitor combinations, predicted to lack synergetic effects, indeed failed to display synergy in our cellular assays. Taken together, our model simulations proved to be highly accurate, correctly predicting the effects of 20 of the 21 combinations.
Synergies of PI3K-MEK or AKT-MEK inhibitions have already been observed in a variety of tumor cells [6,30–34], thus providing further confidence to the synergies of TAK1-PI3K and TAK1-AKT inhibitions. Hence, these novel combinatorial inhibitions are promising candidates readily amenable to experimental testing in a range of cancer cell types.
Understanding the signaling mechanisms underlying synergistic inhibitions is of high interest because it can contribute to the identification of biomarkers informative of treatment response, which may serve as guides to select from an arsenal of established drug synergies the right treatment for the individual patient.
Examination of simulated perturbation effects with our AGS logical model revealed that FOXO, representing pro-apoptotic transcription factors inactivated by phosphorylation [35], was synergistically activated by combined MEK and PI3K or MEK and AKT inhibition (see S1 Text). Interestingly, single inhibitory perturbation of MEK, PI3K or AKT did not change FOXO activity (see S1 Text for more details). These observations match experimental findings in human umbilical vein endothelial cells (HUVEC), where inhibitors targeting MEK and AKT are reported to synergistically activate FOXO [36], which suggests that our model simulations can provide a basis for biologically relevant hypotheses on molecular effects downstream of specific inhibitors.
To further investigate the mechanisms involving FOXO we simulated the inhibitor perturbations in a FOXO knock-out model, and found that combined inhibition of MEK and PI3K displayed no enhanced growth inhibitory effect compared to their corresponding single inhibition. This indicates that the MEK-PI3K synergy indeed depends on FOXO. For the combined MEK and AKT inhibition, the FOXO knock-out model simulations showed only a minor reduction of the synergistic effect of combined MEK and AKT inhibition. The synergy between MEK and AKT inhibitors thus appears to be less dependent on FOXO. Taken together, these simulation results suggest potentially interesting differences between pro-apoptotic signaling events when comparing MEK-AKT inhibition versus MEK-PI3K inhibition.
The mechanistic basis of the synergies observed when inhibiting TAK1-AKT, or TAK1-PI3K, is unknown. Interestingly, AGS model simulations show that FOXO is activated when TAK1 is inhibited in combination with either PI3K or AKT, but not by single inhibitions. Activation of FOXO is thus a potential mediator also for the synergies involving TAK1. In support of this, simulations showed that both TAK1-PI3K and TAK1-AKT synergies were abolished when FOXO is knocked-out, similarly to the finding of MEK-PI3K inhibition (See S1 Text). Potentially, ERK could be involved in signaling downstream of TAK1. In that case the MAP kinase cascade could represent a common mechanism implicated in the synergies involving MEK-PI3K and MEK-AKT, and those involving TAK1-PI3K and TAK1-AKT. However, our AGS model simulations predict that ERK is still active after combined inhibition of TAK1 and PI3K or of TAK1 and AKT. This may indicate that MEK/ERK is not involved in the downstream effects of the inhibitory perturbations involving TAK1. Another kinase could potentially function as a point of crosstalk for TAK1 and PI3K/AKT signaling. In this respect, NLK (Nemo-like kinase) is an interesting candidate as it is known to act downstream of TAK1 [37], mediating inhibitory phosphorylation of FOXO [38].
The model-based suggestion that FOXO activation may be important for synergistic growth inhibition does find experimental support in numerous accounts of FOXO proteins acting as mediators of cytotoxic chemotherapeutic drugs [39]. This suggests that the dynamical behavior of our logical model recapitulates generic properties that may be relevant for a range of different tumor types.
The development of novel anti-cancer medication predominantly focuses on drugs directed against specific molecular targets. However, clinical applications have often been disappointing, resulting only in transient responses followed by drug resistance which hinders therapy benefits. This has led to the consideration of therapies based on combinations of drugs targeting different signaling pathways or cellular processes, with the aim to restrain the evolution of drug resistance and at the same time allow for a reduction in drug dosage, to lower drug-induced toxic effects [2,3,40,41].
These strong incentives for combinatorial drug treatment are challenged by the numerous combinations to consider and by the fact that the efficacy of a given drug combination is dependent on the nature of the specific tumor. Thus, to discover apt drug combinations at a pace compatible with the vast search space posed by the many drugs and diverse cancer cell spectrum, it is mandatory to develop efficient strategies to predict beneficial combinatorial treatment for individual cancers.
Current efforts to come to a rational choice of drug combination therapy by using primary tumor cell cultures and xenograft studies are confronted by high costs and a variable rate of success in tumor cell growth inhibition, and struggle to obtain highly accurate predictions within the timeframe limited by disease progression [4,42–44]. While cancer cell line cultures rarely allow for discoveries that can be directly transferred to a clinical setting, they do allow for experimental investigation of mechanisms underlying biological diversity and robustness and can thus be used to explore strategies to identify potentially effective drug combination therapies. They can therefore contribute to establish a large arsenal of advantageous drug combinations accompanied by prognostic tools enabling the choice of the right combination for the individual tumor. However, even in these cellular models, it is not feasible to test all potential drug combinations and application modes for a sufficient spectrum of cancer cell types. In this context, computational modeling can be of great help to reduce the experimental search space.
We have demonstrated how a logical model built from known signal transduction network information can be tailored to a specific cancer cell system using baseline data, so that it can be used to predict synergistic and non-synergistic combinatorial growth-impeding treatments. Four of the five predicted synergistic combinations were confirmed experimentally with no false negative predictions. With such a success rate, it would have been sufficient to test only a quarter of the 21 possible drug combinations investigated and still not miss any synergistic pair. Our results are encouraging in light of the success rate reported from the recent DREAM challenge [7], where the best-performing method of synergy prediction would have allowed halving the size of screening experiments. However, there are important differences between our study design and that of the DREAM challenge: DREAM analyzed transcriptome changes following broad-acting chemotherapeutic drug treatments, while we investigated the action of inhibitors with specific targets, relying only on information from the unperturbed system.
Contrary to network-based strategies, which commonly use correlation analysis of large-scale datasets from different disease phenotypes [45,46], or large-scale cell culture drug perturbation data to train models for drug response predictions [6,7,9,47,48], our modeling-based strategy exploits mechanistic molecular pathway knowledge, available in databases, along with baseline data from the unperturbed cancer cells of the chosen experimental system. This means that our approach allows for the selection of interesting candidates for efficient drug combinations before performing actual drug perturbation experiments. To our knowledge this has not been successfully demonstrated before.
The majority of regulatory network modeling approaches focus on signaling events driven by specific hormone receptors. This applies to studies investigating logical modeling to understand consequences of interfering with specific growth factor signal transduction responses [16,18,49–51], as well as to quantitative and semi-quantitative modeling approaches used to predict the effect of synergistic signal transduction perturbations [6,31,52]. In contrast, our approach demonstrates that it is possible to effectively use a model representing a cell fate decision network in actively growing cells without considering explicitly any external growth-promoting stimulus (e.g. growth hormone). Indeed, we argue that using the attractor of a self-contained model of a proliferating cell as the reference point for drug synergy analysis provides a good proxy for the state of actively growing cancer cells. Cancer cell growth is considered to be driven by a multitude of growth promoting stimuli. Not only is the potential repertoire of these signals substantial, relatively little detail about their signaling mechanisms is known. We therefore assume that we can summarize their effect by considering this multitude of signals to provide a context promoting robust growth and that we can therefore dismiss any further detail. On this basis we accommodate a sustained multifactor-driven proliferation [1] by employing a self-contained model, where all components included are regulated by other nodes in the model. The configuration of component activities can then be inferred from baseline biomarkers measured in cancerous cells. Together, these model design principles enable us to generate a dynamic model tailored to specific cancer cells, yet not dependent on explicit extracellular input from specific growth promoting agents (e.g. growth hormones) and without the need for initial large-scale inhibitor perturbation data that would be difficult and costly to obtain.
Our definition of drug synergy in experimental validation is based on Loewe additivity [28], and synergies are quantified with the combinatorial index [29]. Several mathematical frameworks have been proposed to determine drug synergy [53]. We chose the Loewe method as it is extensively used, and it correctly handles the sham zero-interaction experiment where a drug is ‘combined with itself’. Regarding the predictions from the logical model, we identify potential synergies by selecting drug pairs that have a more profound effect on the global output ‘growth’ than either of the single drugs. While the experimental synergy analysis allows quantifying the degree of synergy, the logical model is discrete and cannot provide synergy quantifications. Even though our definitions of synergy in experiments and simulations are not identical, our model-based classification proved to be highly accurate with regard to synergies assessed experimentally. Tentatively, a translation of logical variables into continuous ones could be considered (see for example [54]), to estimate synergies in a manner more analogous to the computation of experimental combinatorial indexes.
The AGS gastric adenocarcinoma cell line was chosen as a model system because its gene expression profile is highly similar to profiles of gastric adenocarcinoma [55,56] (intestinal subtype, Lauren’s histopathological classification). Despite increased understanding of the molecular underpinning of gastric cancer, it remains the second leading cause of cancer death globally and as such an austere reminder of the need for improved treatments [57]. In Western countries, two thirds of gastric cancer cases are discovered at a stage where radical treatment is not feasible, and more than half of the patients who can be radically treated will experience relapse. For patients with advanced gastric cancer, the 5 year survival is less than 10% [58].
The synergy between PI3K and MEK inhibitors in AGS cells is in line with previously published observations [6,30,31,33], and is currently being pursued in clinical trials for advanced solid cancer (including pancreatic, breast, non-small cell lung cancer and colorectal cancer) [59]. Similarly, the synergistic effect of MEK and AKT inhibitors has been previously observed [32,34], and is currently investigated in clinical trials (including multiple myeloma, breast, endometrial, colorectal, non-small cell lung cancer, pancreatic cancer, ovarian cancer) [60]. The novel growth inhibitory synergies between TAK1-PI3K and TAK1-AKT discovered here are interesting candidates for further investigations.
Our knowledge concerning molecular regulatory mechanisms underlying cell fate decision networks is rapidly increasing. This poses both opportunities and challenges to integrate many details into an extensive understanding, enabling global mechanistic reasoning on regulatory networks and construction of comprehensive models that can provide in silico predictions of drug effects. Translation of these predictive capabilities to clinical settings is facilitated by the increasing availability of patient omics data, which can provide biomarkers informative of cellular signaling status associated with disease and treatment response. The integration of biomarker information with models of combinatorial drug responses may provide important clues to improve health care for patients who currently lack effective treatment.
We used pathway information from databases as a source of signaling components, and included interactions for MAPK pathway (JNK, p38, ERK), the PI3K/AKT/mTOR pathway, the Wnt/β-catenin pathway and the NF-κB pathway. Each protein was annotated with its official gene symbol, Uniprot protein identifier, and mutational status in our AGS gastric adenocarcinoma experimental system. All interactions were substantiated by bibliographical references documenting experimental evidence. The model was composed of 75 signaling components and 149 directed interactions.
AGS (human gastric adenocarcinoma, ATCC, Rockville, MD) were grown in Ham’s F12 medium (Invitrogen, Carlsbad, CA) supplemented with 10% fetal calf serum (FCS; Euroclone, Devon, UK), and 10 U/ml penicillin-streptomycin (Invitrogen). Growth experiments were performed with medium with 5% FCS.
Chemical inhibitors PI-103 (Merck), AKT-i-1,2/AKT inhibitor VIII (Merck), PD0325901 (Sigma-Aldrich), PKF118-310 (Sigma-Aldrich), BIRB 0796 (Axon), and CT99021 (Axon) dissolved in DMSO at stock concentrations of 20 mM, except PI103 which was dissolved in DMSO at a stock concentration of 10 mM.
Cell growth was measured without labelling, in real-time, with the xCELLigence RTCA SP (96-well) or xCELLigence RTCA DP (16-well) growth assay (Roche Applied Science). This system utilizes culture plates with gold electrode arrays at the bottom of each well in multi-well E-plates (Roche Applied Science). Real-time measurements of the impedance across the gold arrays were reported in the dimensionless unit of cell index which is taken to correspond to the number of cells. In agreement with manufacturer’s instructions, cells were split 1:1 the day before experiments to ensure that cells were in an exponential growth phase at the time of seeding cells for xCELLigence analysis. First, complete medium was added to wells in 50 μl aliquots to measure background signal, next 100 μl of cell suspension was added, at a seeding density of 5x103 cells per well. The well plate was then put back in the RTCA SP/DP instrument, where cells are allowed to adhere overnight (20 hours). The well plate was then removed from the instrument, and 50 μl aliquots of complete medium with chemical inhibitor of interest were added to each well, to a total volume of 200 μl. Real-time monitoring of cell proliferation was performed for 72 hours, at which time the effect of growth arrest was stable (see S1 Text).
|
10.1371/journal.pcbi.1000979 | Evaluating Gene Expression Dynamics Using Pairwise RNA FISH Data | Recently, a novel approach has been developed to study gene expression in single cells with high time resolution using RNA Fluorescent In Situ Hybridization (FISH). The technique allows individual mRNAs to be counted with high accuracy in wild-type cells, but requires cells to be fixed; thus, each cell provides only a “snapshot” of gene expression. Here we show how and when RNA FISH data on pairs of genes can be used to reconstruct real-time dynamics from a collection of such snapshots. Using maximum-likelihood parameter estimation on synthetically generated, noisy FISH data, we show that dynamical programs of gene expression, such as cycles (e.g., the cell cycle) or switches between discrete states, can be accurately reconstructed. In the limit that mRNAs are produced in short-lived bursts, binary thresholding of the FISH data provides a robust way of reconstructing dynamics. In this regime, prior knowledge of the type of dynamics – cycle versus switch – is generally required and additional constraints, e.g., from triplet FISH measurements, may also be needed to fully constrain all parameters. As a demonstration, we apply the thresholding method to RNA FISH data obtained from single, unsynchronized cells of Saccharomyces cerevisiae. Our results support the existence of metabolic cycles and provide an estimate of global gene-expression noise. The approach to FISH data presented here can be applied in general to reconstruct dynamics from snapshots of pairs of correlated quantities including, for example, protein concentrations obtained from immunofluorescence assays.
| Programs of gene expression lie at the heart of how cells regulate their internal processes. Some dynamical gene-expression programs, such as the cell cycle, are well known and studied, others, such as metabolic cycles, have only recently been recognized, and many other dynamical programs including switches are likely to be discovered. Traditional bulk studies typically fail to resolve such cycles or switches, because individual cells are out-of-phase with each other. On the other hand, standard techniques for studying single cells are limited in time resolution and scope. RNA Fluorescent In Situ Hybridization (FISH) is a single-cell technique that offers both high time-resolution and precise quantification of mRNA molecules, but requires fixed cells. We have explored how, when, and with what prior information FISH snapshots of pairs of genes can be used to accurately reconstruct gene-expression dynamics. The technique can be readily implemented, and is broadly applicable from bacteria to mammals. We lay out a principled and practical approach to extracting biological information from RNA FISH data to reveal new information about the dynamics of living organisms.
| Cells are well known to respond to external conditions by altering their gene expression. In recent years, many examples of altered gene expression programs have been revealed by population level studies, including microarray studies of yeast, mammalian, and bacterial cells. But many cells are also known to alter gene expression is ways that are heterogeneous across a cell population. Examples include the acquisition of competence for DNA uptake [1], [2] and spore formation [3] in Bacillus subtilis, induction of the lac operon in Escherichia coli depending on “memory” of previous exposure to lactose and the presence of lactose permease [4], [5], and the response of Saccharomyces cerevisiae (budding yeast) temperature-sensitive mutants to a shift to non-permissive temperature depending on the position of cells in their division cycle [6], [7]. Heterogeneous changes in gene expression in response to homogeneous external cues may be purely stochastic as in the switch to competence in B. subtilis [1], [2], [8], or may depend on pre-existing non-genetic differences such as the phase of the cell cycle in budding yeast [6], [7].
Since population level studies are not well suited to reveal heterogenous behavior, how can heterogeneous changes in gene expression be studied and quantified? Fluorescent reporter proteins have been used successfully to report on expression of a small number of genes either via FACS analysis or fluorescence microscopy. However, the use of fluorescent reporters is generally limited to highly expressed genes, with time resolution severely limited by fluorescent protein maturation and the low turnover rates of the fluorescent marker. Moreover, construction of fluorescent reporters can be laborious and impractical for studies of large-scale transcriptional responses.
A promising approach that has recently been used to study gene expression on a cell-by-cell basis is Fluorescence In Situ Hybridization (FISH) [9]–[11]. In FISH, fixed cells are exposed to fluorescently labeled probes of specific mRNA transcripts, so that the number of these mRNAs can be counted in each cell by the number of bright spots. Advantages of FISH include: (1) absolute quantification since the actual number of mRNAs can be counted, (2) time resolution since there is no delay for reporter maturation, (3) ability to directly study wild-type cells, and (4) the ability to probe simultaneously for multiple mRNAs, e.g. by employing probes with different fluorescent spectra [10], [12]. A significant disadvantage of FISH is the requirement to fix cells. This disadvantage presents a particular challenge when it is the dynamics of gene expression that is of central interest. For example, each individual drawn from an asynchronous yeast population represents a particular moment in the cell division cycle. In essence, the problem we wish to address is how to reconstruct the dynamics of gene expression from what amount to “snapshots”, where each individual cell represents a different point in time.
Here, we present an approach to extracting information about the dynamics of gene expression from FISH data by considering correlations of expression between pairs of genes (cf. Fig. 1). The approach applies even if the dynamics of interest occurs heterogeneously in a population. One class of dynamics we consider are cyclic oscillations of gene expression. Common examples are the cell cycle, circadian oscillations, and metabolic oscillations [13]. Cyclic oscillations of gene expression, such as the cell cycle, have been studied at the population level by synchronizing cells, but for many organisms synchronization is difficult without strongly perturbing the cells. A non-perturbative approach to studying oscillatory gene expression is likely to be of value in these cases. To study metabolic oscillations, cells of the yeast Saccharomyces cerevisiae have been synchronized in chemostats [13], but those cells demonstrably continue to influence each other via levels of dissolved oxygen and other chemical species. To ascertain if Saccharomyces undergoes metabolic oscillations outside the chemostat, Silverman et al. [14] recently obtained an extensive FISH data set, and argued for the existence of metabolic oscillations based on correlations in gene expression. Using the same data set, we apply our approach to reconstructing oscillatory dynamics, and confirm the existence of metabolic cycles in unsynchronized yeast populations [14]. Our approach can also be applied to transient oscillations in response to external stimulation, such as in the bacterial SOS response to DNA damage [15] or in the analogous eukaryotic p53-Mdm2 system [16]. Another class of dynamics we consider are stochastic switches among different states of gene expression. Examples include persister cells in Escherichia coli [17], competence [1], [2], [8] and swimming/chaining in Bacillus subtilis [8], the stringent response in mycobacteria [18], and galactose utilization in Saccharomyces cerevisiae [19].
Specifically, we show how Maximum Likelihood Estimation (MLE) [20] can be applied to FISH data obtained for multiple pairs of genes to reconstruct the underlying dynamics of gene expression. MLE consists of finding the set of parameters within a particular family of models for which the observed data is most “likely”. MLE has been applied successfully to biological data analysis in many contexts, from reconstruction of evolutionary trees [21], [22] to estimation of genetic parameters [23] to understanding the evolution of gene structure [24]. We show using synthetic FISH data that MLE can accurately reconstruct dynamics, even in the presence of substantial noise, provided the number of genes and the number of FISH observations per gene pair are sufficient. Reconstructing gene-expression dynamics is most challenging in the “bursty” regime where mRNAs are often present at very low levels or not at all in the cell, except when transcriptional bursts occur. For this regime, we present a robust approach based on thresholding the FISH data into binary form, followed by MLE analysis. In this case, we show that Principal Component Analysis (PCA) of the covariance matrix performs nearly as well as MLE. We suggest that the two-step approach of thresholding followed by MLE or PCA is likely to prove the best practical approach to reconstructing gene-expression dynamics for most real FISH data sets, and we demonstrate this approach using the data set of Silverman et al. [14].
Importantly, the method we present here for inferring intracellular dynamics from data in the form of “snapshots” is quite general, relying only on measurements of pairs of quantities in single cells, with no requirement for exact counts. The method can therefore be applied with little modification in other contexts such as quantitative immunofluorescence or single-cell sequencing studies.
We presume that production of mRNA transcripts is a stochastic process. Transcription factors bind to DNA at random times, with a probability that depends on other signals, and which can therefore also vary with time. Binding of one or more transcriptional activators, or unbinding of repressors, typically leads to production of a “burst” of mRNA transcripts. One can distinguish three regimes, two of which are illustrated in Fig. 1. In the first regime, many bursts typically contribute to the total concentration of a particular mRNA species at any moment. The distribution of mRNA is therefore approximately Gaussian with a mean and variance that can vary with time, e.g. over the cell cycle. We refer to this case as the continuous regime. The second regime is the opposite limit where mRNA production is highly intermittent [10] – typically there are very few mRNAs of a particular species, and when there are more than a few, they all stem from the same burst. We refer to this case as the bursty regime. The third regime is the intermediate case, where a few bursts typically contribute to the number of mRNA present at any moment. In what follows we focus on the two first regimes. Optimal treatment of the intermediate regime requires a more detailed and/or empirical noise model, but the thresholding method we develop for the bursty regime can also be usefully applied in the intermediate case, as demonstrated by our analysis of FISH data for metabolic cycles in yeast [14].
For each regime of mRNA expression, our approach consists of defining a class of possible dynamics, and choosing the one for which the observed data is most likely. Specifically, for a given set of model parameters, we calculate the probability of the observed data, and then ask for the particular set of parameters that maximizes this probability. Since the probabilities don't sum to one over all models (i.e. sets of parameters), they are called “likelihoods” and hence this approach to parameter inference is called Maximum Likelihood Estimation (MLE). Below, we demonstrate the practicality of the MLE approach using synthetically generated FISH data in both the continuous and bursty mRNA regimes.
In practice the parameter optimization in MLE can be a challenge, and algorithms used to search parameter space for the maximum likelihood can get stuck in local maxima. However, the general formulation of the maximum likelihood approach is conceptually distinct from the detailed choice of algorithms used to optimize parameters, and so we have chosen to present only fully optimized results in the main text. In Methods, we present a practical method for searching parameter space that typically quickly finds the model parameters that maximize the likelihood of the data.
It is important to recognize one absolute limitation of using FISH data to reconstruct the dynamics of gene expression. Because cells must be fixed before mRNAs are measured, only “snapshots” of individual dynamical trajectories are available. As a consequence, it is impossible from FISH data alone to determine the overall time scale of the dynamics of gene expression. Thus, while it is possible to infer from correlated FISH data that cells undergo cycles of gene expression, and even practical, as we will show, to accurately reconstruct such cycles, it is not possible, even in principle, to determine the period of these oscillations. Similarly, it is not possible, even in principle, to determine which direction around the cycle of gene expression corresponds to the forward arrow of time. In many cases, we anticipate that other methods, e.g. fluorescent reporters or population-level assays, can be used to provide this missing information. In some cases, the insensitivity of FISH data to cycle period may actually prove advantageous. In bulk studies of synchronized cell populations, different cycle periods of individual cells lead to loss of synchrony and therefore loss of signal. In contrast, for single-cell FISH studies, differences in cycle period among cells will not affect mRNA correlations. Hence, variations of period will not affect the ability to reconstruct cycle dynamics from mRNA snapshots. However, cell-to-cell variations of the shape of the cycle constitute noise even for FISH, which can at best allow reconstruction of the mean cycle waveform.
At a qualitative level, the regime of continuous mRNA production allows for relatively straightforward reconstruction of cyclic gene-expression dynamics. In the absence of noise, FISH data for even a single pair of genes is sufficient. One can simply plot the ordered pairs of FISH data as in Fig. 2A, and infer the dynamics from the smooth trajectory that joins the data points. (The fundamental limitations of FISH are already clear in this case – from the FISH data points alone one cannot in principle infer the period of the trajectory nor its direction.) Noise complicates the reconstruction somewhat, and requires a computational means of inferring the trajectory that best fits the data. Our solution presented below is to find the trajectory most likely to account for the data, within a family of harmonic functions, , , etc. The ability to accurately reconstruct trajectories in the presence of noise is greatly improved by FISH data for multiple pairs of cycling genes. Geometrically, the true trajectory is a path in the space of all the cycling genes. Each set of pairwise FISH data represents a projection of this trajectory onto a plane as in Fig. 2A. The more such projections are available, i.e. the more sets of pairwise FISH data, the more accurate the reconstruction of the true trajectory will be. This approach can be readily extended to the case of a stochastic switch between distinct gene-expression states, with the same improvements expected from multiple FISH pairs.
Reconstruction of gene-expression dynamics in the regime of bursty mRNA production is more challenging. In this case, the data consists of the presence or absence of bursts of mRNAs, with rare coincidences of bursts for two genes (cf. Fig. 1D). All of the information in the data is therefore captured by a single number for each pair of genes, namely the covariance of their mRNA bursts. However, as we show below, the matrix of these covariances for multiple gene pairs in principle contains enough information to reconstruct the underlying parameters of cyclic trajectories or stochastic switches (albeit in some case with degeneracies that require additional constraints to resolve). Since coincident bursts of mRNAs are likely to be rare, one expects the covariance matrix derived from the data to be noisy. Nevertheless, with a sufficient number of sets of pairwise FISH data, we find that accurate reconstruction of the underlying gene-expression dynamics is feasible.
We first consider the continuous regime where many bursts typically contribute to the instantaneous mRNA number. To demonstrate the MLE algorithm, we reconstruct the dynamics of gene expression using synthetic FISH data for which the underlying dynamics is known. We focus on analyzing cyclic dynamics, e.g. the cell cycle or a metabolic cycle; the results can be readily extended to stochastic switches, which are introduced in a later section. We denote the mean expression level of mRNA for gene by , which is taken to be periodic with the same period for genes . For concreteness, we denote the period as , although cannot be inferred from FISH data alone. FISH observations are generated for pairs of genes at randomly chosen times: , where the s reflect fluctuations in mRNA number around the mean, as well as noise in the measurement. is assumed to be a Gaussian random variable of mean zero and standard deviation . We assume that is not a function of the mean expression , but it is straightforward to extend the method to the more general case. (A natural extension of the model is to consider where characterizes the measurement noise and is the characteristic size of the independent events of mRNA production leading to the total mRNA number.) We aim at maximizing the likelihood of the observations within a family of harmonic functions of period . Bayes Theorem for the probability of a particular model (i.e. set of parameters) given the data states:(1)We neglect the term, , corresponding to prior knowledge of the parameters, as there is no obvious choice for what such prior knowledge should be; moreover, with sufficient data, including such priors generally has little effect on the results of optimization. The probability of the data is a constant normalization factor, and so does not affect the relative likelihood of models. Therefore the probability of the model given the observations is proportional to the probability of the observations given the model .
For each FISH observation, one therefore has:(2)and for the combined likelihood over all observations,(3)where the product runs over all FISH observations. In what follows we maximize assuming harmonic oscillations of mRNA levels,(4)The method can be systematically extended to periodic trajectories that are not simple sine waves by including higher harmonics. It is also straightforward to extend the method to more detailed noise models. For example, non-Gaussian noise can be incorporated by appropriately modifying the Gaussian integrand in Eq. (2). Similarly, global transcriptional noise [25] can be modeled in Eq. (2 via a single additional random variable multiplying both and . Later, we consider both higher harmonics and global noise in detail for the more physiologically relevant case of bursty mRNA production.
We generate synthetic FISH data by first choosing the parameters in Eq. (4) for the oscillating mRNA levels , and then generating FISH observations based on these parameters. Specifically, we choose random variables uniformly on , for genes . We then define the model parameters in Eq. (4) as , and . This construction ensures the positivity of the mRNA levels , and also ensures that the genes considered oscillate in time with significant amplitudes. The noise amplitudes are random variables, distributed continuously, . The synthetic FISH data are generated by choosing for each gene pair , random times and random noise values , and constructing . In this way, the synthetic data correspond to a set of independent, pairwise FISH observations. An example is shown in Fig. 2 for . The red ellipse indicates the true mean-mRNA-level trajectory , and the crosses are the randomly generated FISH data points. The blue ellipses correspond to reconstructions of the mean trajectory via maximization of the likelihood in Eq. (3).
To test the accuracy of reconstruction of mRNA dynamics using our MLE approach, we generated a large number of sets of parameter, and for each parameter set generated synthetic FISH data and then applied MLE to reconstruct the true dynamics. Specifically, for each parameter set defining a trajectory of mean mRNA levels , we maximized the likelihood with respect to . To ensure that we always found the global maximum of the likelihood, the initial guess for the parameters was taken to be the true parameters describing the mean dynamics. (In Methods, we present a simple algorithm that almost always finds the global likelihood maximum without prior knowledge of the true parameters. However, in Fig. 2, we chose to present the true MLE optimum as the fundamental limit of reconstruction accuracy, not limited by a particular algorithm.)
As shown in Fig. 2, with synthetic FISH data for only two genes (dashed blue ellipse) the reconstruction is rather poor, in this case mistakenly assigning too large a noise to each gene and missing the phase shift. However, the addition of pairwise information from two more genes to make (for a total of gene pairs) is enough to correct these errors and provide a very accurate reconstruction. Since each pairwise data set is independent, the total amount of data grows as .
To quantify the accuracy of the MLE algorithm, we computed the reconstruction error , which characterizes how much the reconstructed dynamics varies from the true mRNA dynamics,(5)where is the MLE reconstructed trajectory for mRNA , and is the (true) average number of mRNA over the period . Results for the reconstruction error are shown in Fig. 2B for , , and . Each point is averaged over 20 randomly generated parameter sets. As expected, the results improve with the number of genes and the number of FISH observations per gene pair, but at this noise level the results are already good for and .
We now consider the bursty regime where a cell will typically either have few (or no) mRNAs of a particular type, or the mRNAs present will come from a single recent burst of transcription. In this limit, the information provided by FISH is essentially binary - either mRNAs for a particular gene are present at significant levels, indicating a recent burst, or they are not. Formally, if a significant number of mRNAs for gene are present, then , otherwise . The optimal threshold to set for the “presence” of mRNAs will depend on burst size and duration, measurement noise, and the total number of FISH observations – see Discussion. FISH data yields an estimate for the mean probability that mRNAs are present above the threshold value for each gene (in the expression for , the variable reports the absence or presence of mRNA for observation of the pair , and the sum is made on the observations that probe for mRNA of gene .) The FISH data also yields an estimate for the covariance for each pair of genes. We aim to accurately reconstruct the mRNA dynamics from these quantities and , which capture all the information provided by the binarized FISH data in the bursty regime. However, even with perfect knowledge of mean expression and covariance, the reconstruction of mRNA dynamics has fundamental limitations in this regime. We illustrate by considering both cyclic dynamics and stochastic switches.
We denote by the probability that the number of mRNAs of type present at time is larger than some threshold , and we call such an event a burst in what follows. Assuming is any periodic function with period , it can be expanded in harmonics:(6)with more harmonics generally required to capture more complex oscillation patterns. Note that if is twice the number of harmonics considered, the number of parameters is , the coming from the invariance with respect to the overall phase. For the following discussion it is sufficient to keep only the first two harmonics, shown explicitly in Eq. (6). In this case, denoting by the average over a cycle, one finds:(7)(8)where and denote the true cycle-averaged mean and covariance, respectively. One immediately sees that the transformation for all leaves both and unchanged. Thus, in this bursty regime, pairwise FISH data alone cannot disentangle different harmonics without prior knowledge. However, any additional constraint, including even a single triplet FISH data set, can readily resolve the ambiguity between harmonics. (A triplet FISH observation, i.e. simultaneous measurement of three different mRNA types, leads to terms , which do not have the problematic symmetry.)
For simplicity, let us now consider only the lowest harmonic, as in the previous section. We introduce the -dimensional vectors and , defined as and . Each component can vary independently of the others, as there are parameters, and coordinates for the two vectors. Then by inspection the covariance matrix from Eq. (7) can be written:(9)which shows that is in general of rank 2. If the second harmonics are included, is of rank 4, etc. Note that all symmetric matrices of rank can be written in the form of Eq. (9), with eigenvectors, implying that for a covariance matrix of even rank there is always an interpretation of the dynamics in terms of cyclic trajectories. Unfortunately, this interpretation is not unique except for the case of a single harmonic (), which can be seen as follows. The observed mean probabilities for mRNA bursts of each type leads to constraints. The observed covariances lead to an additional constraints. Being a symmetric matrix of rank , the covariance matrix can be defined by this many coefficients, i.e. the number necessary to describe the eigenvectors, enforcing orthogonality among them. The expression for the number of covariance constraints is true for sufficient large , but in general is .) The total number of constraints provided by FISH is therefore . Thus the number of unconstrained parameters is , which is zero for , but is already 5 for . Hence, for two harmonics at least 5 triplet FISH data sets or other constraints are required to be able to infer all the parameters.
An important consideration in analyzing FISH data is that overall transcription rates may vary from cell to cell. Indeed, measurements of gene-expression noise in single yeast cells at the protein level reveal global fluctuations [25]. How can the dynamics of bursty gene expression be reconstructed against the background of these global correlations? We consider the case of a simple harmonic cycle. The probability that the number of mRNAs of type present at time is larger than some threshold now reads:(10)where , representing the fluctuating global level of transcription, is a random variable of mean unity and standard deviation . One then obtains for the true cycle-averaged mean and covariance :(11)(12)Introducing the definitions , , and , the covariance matrix from Eq. (12) can now be written:(13)which shows that is now of rank 3. If the second harmonics are included, is of rank 5, etc. The maximum likelihood reconstruction for the model of Eq. (10) provides an estimate of the level of global transcriptional noise, as we show below for our reconstruction of metabolic cycles in yeast.
We now consider a model where the expression pattern can switch stochastically among distinct states, as illustrated in Fig. 3. We assume all genes of interest switch their expression synchronously, consistent with control by a single transcription factor, and without delays, consistent with state lifetimes long compared to mRNA lifetimes. On average, each state occurs with probability . In state , the true probability that a burst of mRNAs of type is present is denoted . For such models, the number of parameters is therefore , taking into account that . One finds, following simple arithmetic:(14)(15)where is the state-averaged burst probability and is the covariance matrix. is a -dimensional vector of components .
It is straightforward to see that , which implies that the different vectors are not independent. Together with Eq. (15), this dependence implies that the covariance matrix is in general of rank . Thus, pairwise FISH data alone cannot distinguish a 3-state model from simple harmonic dynamics (or a 5-state model from a cycle including second harmonics, and so on). Moreover, even if one assumes that the dynamics is a switch, the parameters cannot be resolved uniquely: the number of constraints set by the measured means and covariances is , so that the number of unconstrained parameters is , which is 1 for a 2-state switch, and 3 for a 3-state switch. The corresponding number of triplet FISH data sets or other constraints are therefore required for parameter inference; however, if this additional data is available, switching parameters can be inferred even in the presence of global noise, as discussed above for the case of a simple cycle.
For either cyclic or switching dynamics, maximum likelihood parameter estimation in the regime of bursty mRNA production requires the following steps, (1) estimating the mean burst probability and covariance from the FISH data, (2) determining the uncertainty of these estimates, and (3) obtaining the parameters for which the observed data is most likely. Taking an average over FISH data provides an estimate of the cycle- or state-averaged probability for a burst of mRNAs of type to be present. Specifically, , where reports the absence or presence of mRNA for observation of the pair , and where the sum is made on the observations that probe for mRNA of gene . Similarly, FISH data provide an estimate of the covariance, namely . For a finite number of data points, these estimates will be noisy, i.e. and , where the right hand sides are the exact values. Since coincident bursts of mRNAs of type and will be rare, the covariance estimate from finite FISH data may deviate significantly from the true covariance , and one must allow for this uncertainty in the maximum likelihood calculation. In contrast, one may safely neglect the uncertainty in the FISH estimate of the mean burst probability, both because single mRNA bursts are much more frequent than coincident bursts, and because each mRNA type is probed times more frequently than each pair. In practice, we therefore demand that the MLE parameters yield exactly.
To estimate the uncertainty in the covariances, we first note that the true variance in is(16)where the overbar indicates the cycle or state average. We can then estimate the relevant quantity from the data, , to obtain an estimate for the variance . Using this estimate for , the probability of obtaining a covariance estimate if the true covariance is is given by:(17)Since the observations for the different mRNA pairs are independent, the likelihood of the observed covariance estimates for a given set of parameters is readily obtained from Eq. (17).
As discussed above, for bursty mRNA production the means and covariances alone cannot distinguish cyclic from switching dynamics. However, if one has prior evidence that gene expression is cyclic, maximum likelihood estimation can be usefully employed to reconstruct the dynamics. If and for a sufficient data, the algorithm works well, as illustrated in Fig. 4 for harmonic dynamics. A larger data set is needed than in the continuous mRNA regime because observations of coincident bursts are rare. Note that for , the constraints from the observed means and covariances are fewer than the parameters, and reconstruction requires additional constraints, e.g. from triplet FISH data.
A stochastic switch between 2 states implies a covariance matrix of rank 1, and therefore can be distinguished from cyclic dynamics, which leads to a minimum rank of 2 (unless all the genes are exactly in phase). Still, one piece of additional information is required to reconstruct the dynamics. For example, it is sufficient to know the expression level of a single gene in one state. Here, we instead assume that the probability of being in one state is known, and given that constraint we infer all the levels of gene expression from synthetic FISH data. (Note that FISH data can only reveal the probabilities to be in each state, not the kinetics of switching, e.g. interval durations or branching ratios.) The MLE algorithm works well, as shown in Fig. 3A, as long as and for sufficient data. To quantify the accuracy of the MLE parameter estimation for switching dynamics, we have plotted in Fig. 3B the reconstruction error which measures the deviation of the reconstructed rates from the true rates (normalized by the state-to-state variation and weighted by the state probabilities) and averaged over all measured genes:(18)where the are the reconstructed rates.
In principle, with enough FISH data it should be possible to reconstruct more than just the probability of observing a burst. For example, the entire distribution of mRNAs of each type in each switching state could be obtained using MLE, e.g. via Expectation Maximization (EM) [26], by treating the full distributions rather than just the mean burst probabilities as unknowns. However, the approach proposed above of thresholding and binarizing the data has the advantage of reducing noise, and thereby reducing the required number of FISH observations, while still allowing for inference of the basic gene-expression dynamics.
In the regime of bursty mRNA production, all of the information from FISH is contained in the mean burst probabilities and the covariance matrix, suggesting that Principal Component Analysis (PCA) could be usefully applied. For example, for a 2-state switch the covariance matrix has rank 1. Thus, according to Eq. (15), performing PCA by diagonalizing directly yields , the vector of differences of burst probabilities between the two states, as the only eigenvector with a non-vanishing eigenvalue. Together with the mean burst probabilities, , this yields full information on the switching dynamics. One caveat is that all the diagonal terms are missing from the estimated covariance matrix , as one cannot obtain an estimate of directly from FISH data. To solve this problem, we initially diagonalize the matrix with a zero diagonal, and obtain the principal eigenvalue and eigenvector. We then approximate the diagonal terms of with the diagonal terms of the rank-1 matrix built using this single eigenvector . We repeat this procedure iteratively to convergence, and take the converged principal eigenvector as an estimate of . This PCA approach performs similarly well to MLE for the case of a 2-state switch, as shown in Fig. 3. The PCA approach can be easily extended to cases in which has a higher rank, where it also performs well, see Fig. 4. Of course, like MLE, PCA has the same fundamental limitations discussed above that are inherent to coincidence detection.
In practice, elements of the PCA and MLE approaches can be usefully combined. The main utility of PCA lies in diagonalizing to infer its rank. (The iterative approach to filling in the diagonals of can help refine this procedure.) From the rank of , one has a direct estimate of the “complexity” of the dynamics. Complexity here means the number of states in a switch model, or the number of harmonics to be considered for cyclic dynamics. This suggests the following heuristic approach to FISH data analysis: First diagonalize . Then isolate a group of eigenvalues that are significantly larger than the rest. Use prior information to select between the different models (cyclic or switching) leading to such a rank, and finally compute the model parameters using maximum likelihood estimation.
In recent years, McKnight and coworkers demonstrated that the yeast Saccharomyces cerevisiae grown in chemostats can undergo synchronized metabolic oscillations [13], [27]. As shown in Fig. 5, the mRNA levels of three clusters of genes – Oxidative, Reductive Building, and Reductive Charging – were found to cycle together, with the expression of each cluster peaking at a different phase of the cycle. These population-level chemostat studies raise the question - is there an intrinsic metabolic cycle in individual cells in unsynchronized cultures? To address this question, in [14] FISH data were obtained from single, unsynchronized yeast cells. Specifically, correlations of mRNA levels were determined for pairs of genes, each of which cycled in the chemostat. The correlations observed in single cells closely matched those found in the chemostat studies, leading to the conclusion that metabolic oscillations do occur in individual cells in unsynchronized populations as well as in synchronized chemostats. However, in [14] no attempt was made to go beyond correlations to reconstruct the dynamics. Here we use MLE to infer metabolic gene dynamics in unsynchronized populations. Our results support the conclusion of [14] that the gene clusters observed in the chemostat persist in individual cells in unsynchronized cultures. In particular, we find that the genes of the Oxidative cluster oscillate together and so do the genes of the Reductive Building cluster . The situation is still unclear for the Reductive Charging genes, but is likely to be clarified by additional FISH data.
To analyze the dynamics, we first binarized the FISH data of [14] as appropriate for bursty gene expression. The data consists of 79 pairwise FISH experiments involving a total of 25 genes. To set an appropriate binary threshold of expression for each gene, we found the median of the mRNA distribution for each gene . Only 7 genes have a median larger than zero (and in all cases ), indicating that most genes are indeed bursty – despite the fact that those 25 genes were selected, among other criteria, to have a high expression level [14]. denotes the probability that the number of observed mRNA of gene is strictly larger than and is directly measurable from the data. We found , with the lower range coming from genes for which the median .
We assumed that the dynamics is cyclic and considered the expansion of Eq. 6 up to the first harmonic. Such a model has 74 independent parameters for 25 genes. Moreover, the number of data points per pair of genes varies from 175 to 16032, with only 29 pairs having more than 2000 data points. Thus some of the correlations are well-characterized, but others are not. If only the 29 gene pairs with more than 2000 data points are considered, even a single-harmonic model is under-constrained. To circumvent this problem, we are guided by the observation apparent from Fig. 5 that the gene expression in all clusters becomes much smaller than its mean at some point in time. This suggests a simplified model where the probability of expressing more mRNAs than the median mRNA number for gene cycles as:(19)Therefore, once is extracted from the data there is a single free parameter per gene, namely its phase .
Next, the likelihood of all the observed FISH correlations was maximized with respect to the phases . The global maximum was found by considering various random initial phases, relaxing to a maximum, repeating, and choosing the maximum with the largest likelihood. We consistently found the same maximum after the order of 10 optimization runs. Results for the reconstructed dynamics are shown in Fig. 6 for the 14 most tested genes (per gene number of observations). Genes belonging to each metabolic cluster identified by the chemostat studies are represented by distinct colors as indicated in the legend. The location of the maximum probability for each gene is indicated by an arrow. From the positions of the arrows it is apparent that genes belonging to the Oxidative cluster also cluster in an unsynchronized population, and so do the genes of the Reductive Building cluster. From the existing data we cannot yet conclude whether the Reductive Charging genes also cluster.
To quantify our results statistically, we define for each cluster, , the quantity , where is the number of genes in cluster . The , which characterize the average cluster activity, are plotted in Fig. 6. If the genes belonging to a cluster are perfectly synchronized, i.e. are identical for all , then will reach zero along the cycle. More generally, the lower the minimum of , the more synchronized the cluster is for fixed . We find that the Oxidative and Reductive Building genes are indeed clustered: the probability of finding such low minima for the two corresponding curves would be only and respectively ( when considered together) if the phases were random. On the other hand, the minimum of the Reductive Charging cluster is comparable to the typical value for random phases.
From the chemostat studies [13], we expect the amplitudes of oscillation of metabolically cycling genes to be large (10 fold), and so global transcriptional noise (2 fold [25]) should not significantly affect our results. However, to test that our reconstruction of the metabolic cycle is robust with respect to global transcriptional noise, we reconstructed the dynamics allowing for a global correlation among mRNA levels as in Eq. (10). Specifically, we extended the model of Eq. (19) by adding the possibility of a varying global level of transcription :(20)where is a random variable of mean unity and standard deviation . The results of the reconstruction are essentially identical to those shown in Fig. 6, where the global level of transcription was assumed to be fixed (for a comparison see Fig. S1). Moreover, from the reconstruction we infer the amplitude of the global noise of transcription to be (i.e. 55%), which is significant, but considerably smaller than the typical variation during a cycle [13].
In Fig. 6, the lack of evidence for coherent oscillations of the Reductive Charging genes may reflect a real feature of unsynchronized populations. Alternatively, it may reflect the limited data and/or the simplicity of the model of Eq. (19). To investigate the limitations of this model we considered how the pairwise gene covariances it predicts compare with the observed FISH covariances, as shown in Fig. 7. Even the underlying “true” model should not capture the FISH correlations perfectly, especially since some observations are very noisy due to the limited data. However, some general trends appear. In particular our model in Eq. (19) systematically underestimates the largest covariances. This may be due to the fact that the single cosine wave that we use to fit the dynamics is less peaked than the typical expression profile observed in the chemostat [13]. Accordingly higher harmonics should be included to obtain a more accurate description of the gene-expression dynamics, an approach that will be achievable once the data set is enlarged to include additional gene pairs.
In general, Maximum Likelihood Estimation (MLE) requires finding the set of model parameters for which the observed data are most likely. Finding the global maximum in the space of model parameters can be a challenging task, particularly as there may be many local maxima in which a search algorithm can get stuck. For synthetic FISH data in the regime of continuous mRNA production, we found that such local maxima occurred frequently. (In contrast, for synthetic FISH data in the bursty regime a simple steepest-descent algorithm invariably found the same maximum, independent of initial conditions.) To find the global maximum in the continuous regime, we developed a heuristic algorithm that worked very well in practice to reconstruct simple cycles.
One approach is to consider various initial parameter values, and to use a steepest-descent algorithm to find the local maximum of the likelihood. Then the global maximum (with the highest likelihood) could be chosen among the different solutions. However, in practice this procedure can be very time-consuming if initial conditions are chosen randomly. Here we propose two approaches to first compute estimates of the parameters, and then use these estimates to initiate the optimization protocol. In these two approaches we estimated the parameters as follows: (1) For the mean expression level we took . (2) For both the amplitude of oscillations and the noise amplitude we took half the standard deviation of the observations of the corresponding gene . (3) Empirically we found that the initial choice of phase is critical in determining if the global or only a local maximum is found. Therefore, to accurately estimate the relative phases we introduced the Pearson correlation matrix (a normalized variant of our covariance matrix) . This definition implies . yields a rough approximation of , which leads to the following two approximations, the first being extremely crude:
Results of optimization using approaches (a) and (b) to set initial parameter values are shown in Fig. 8A. The results in Fig. 8A are nearly indistinguishable from those obtained using the true parameters as initial conditions, shown in Fig. 8B, demonstrating that the above protocols performs well in identifying the global maximum of the likelihood.
The ability to count mRNA molecules in single cells by Fluorescence In Situ Hybridization (FISH) [9]–[11] allows for highly quantitative studies of cell-to-cell variation in gene expression. However, the requirement that cells be fixed before RNA FISH analysis precludes the use of RNA FISH to directly study transcriptional dynamics in single cells. Nevertheless, we have shown here how and when correlations between levels of different mRNAs can be exploited to reconstruct transcriptional dynamics, even if cells are asynchronous. All that is necessary is for FISH data to be obtained simultaneously for pairs of genes (or in some cases triplets of genes) a technique that is already well established [10], [12]. As a practical demonstration, we applied our approach to a large, pairwise FISH data set obtained from a recent study of the yeast Saccharomyces cerevisiae [14]. Our results help confirm the existence of cell-autonomous metabolic cycles in unsynchronized yeast populations [13].
To reconstruct the dynamics of gene expression from FISH data, our approach employs Maximum Likelihood Estimation (MLE) [20] to obtain the set of transcriptional parameters most likely to account for the observed data. In the regime of continuous mRNA production, apart from rescaling and inversion of time for cyclic dynamics, there is no intrinsic limit on the accuracy with which transcriptional dynamics can be reconstructed given enough data. In practice, we have shown that MLE applied to simple parameterizations for transcription (such as the leading harmonics for cyclic dynamics) allows faithful reconstruction from a moderate number of FISH observations, including noise. On the other hand, the regime where mRNA is produced in shortlived bursts [10] presents additional challenges. In this bursty regime, FISH can at most report coincidences of bursts of different mRNAs, and there are consequently fundamental limits to reconstructing the underlying dynamics. For this bursty regime, successful reconstruction will generally rely on prior knowledge regarding the class of dynamics, e.g. cycle vs. switch, and, even so, will in some cases require additional inputs, such as triplet FISH data. (In Table S1, we explicitly quantify the amount of such additional information required for complete dynamical reconstruction.)
In applying our approach, how should one choose among models to reconstruct gene dynamics? For example, when is it better to use multiple harmonics instead of a single harmonic to model a cycle? The answer depends on the type of data. We discuss first the regime of continuous mRNA production. For this case, a standard and reliable way to choose among models when fitting data is “leave-one-out” validation, which both rewards a good fit while punishing overfitting. In the leave-one-out approach, a model is selected and its parameters are optimized on the entire data set, but with one data point left out. The resulting parameterized model is then used to fit the neglected data point. The average fitting error, taken over all possible left-out data points, is a robust measure of the quality of the model. Among competing models, the one that minimizes this error can be selected as the better choice. In the regime of continuous mRNA production, leave-one-out validation can be applied within the MLE framework by using the log(likelihood) of the left-out data point in place of the fitting error. Among competing models, the one with the largest average log(likelihood) is the best choice.
In contrast, finding the “best” model for data in the bursty mRNA regime is generally an under-constrained problem. We showed explicitly that for many cases it is impossible in principle to distinguish among different types of models, or even to find a unique best set of parameters for a given model. Intuitively, reduction of bursty FISH data to pairwise covariances means that even as the number of FISH data points approaches infinity, the number of model constraints stays finite. So, for bursty FISH data inference alone cannot guide one in choosing the model, and one must also use common sense. Clearly, prior knowledge of the system under study should be used in selecting a model. In addition, a simple rule is that one should use models that are sufficiently parsimonious in parameters not to have degenerate solutions. For example, in analyzing FISH data on metabolic cycles, we chose the one-harmonic model because there were not enough low-noise covariances to constrain a two-harmonic model. More generally, it is advisable to choose a model with significantly more well-constrained data than parameters. If the model is barely constrained, the peak of likelihood will generally be close to flat in some directions in parameter space and the reconstruction will be poor. Figure 4 illustrates this point: is the minimal number of genes to avoid degeneracy, but it requires 3 times more data per gene (or twice as many total data points) to reconstruct as well as for . In practice, one test for the quality of the reconstruction in the bursty regime is to compare the observed covariances to the reconstructed covariances, as shown in Fig. 7 for the case of the yeast metabolic cycle genes.
Reconstruction of gene-expression dynamics from FISH data presents multiple practical challenges. One important issue is noise in the measurement of mRNA levels. For the regime of continuous mRNA production, we have shown that sufficient data can compensate for both the noise inherent in gene expression and the noise arising from uncertainty in measurement. For the regime of bursty mRNA production, “binarizing” the data into the presence or absence of a significant number of mRNA molecules substantially reduces the impact of measurement noise. A practical question here is the best threshold to use for binarizing the data. In many cases, the dynamics will be best reconstructed by setting the threshold well above 1 mRNA transcript; for example, in treating the data for metabolic cycles we chose the median expression level for each gene as its threshold. A higher threshold is less sensitive to measurement noise (fewer false positives), and to occasional transcripts produced by promoter leakage (better identification of true bursts), and a higher threshold also allows finer time resolution, as a given burst will remain above threshold for a shorter time (e.g. preventing blurring of boundaries between switching states). However, a higher threshold reduces the number of coinciding bursts in the data, requiring more overall FISH observations. An important related issue is the possibility of correlated noise in the transcription of different genes. An example of such noise is the observed global correlation among transcription rates in yeast [25]. Fortunately, global noise can be readily incorporated within the MLE framework by introducing a single additional variable in the model for gene expression, as in Eq. (10). Indeed, our treatment of global noise among genes involved in the yeast metabolic cycle yields an independent, and reasonable, estimate for this noise at 55% of mean expression. (More complex noise correlations among different genes would require case-by-case analysis.)
False-positive rates and false-negative rates are also both important considerations in analyzing FISH data. These are essentially technical issues beyond the scope of our study, but a few remarks are in order. In Ref. [14], both false positives and false negatives were reduced by the use of multiple fluorescent probes (5) for each mRNA. Only high-contrast spots above a fluorescence threshold indicative of multiple bound probes were counted. This threshold was set empirically from the fluorescence distribution of spots outside of cell boundaries, corresponding to single probes. Nevertheless, with any such thresholding method, there will be cases where the “presence” or “absence” of an mRNA is ambiguous, and in the bursty regime such ambiguities can strongly impact the binarization of the data. Fortunately, because MLE is an intrinsically probabilistic approach, ambiguities can be dealt with by treating the two possibilities, present or absent, probabilistically. As in Ref. [14], by looking at spots outside of cell boundaries, one can obtain the distribution of intensities for spots that are actually noise (typically single probes that have not been washed away), and by looking inside cell boundaries a similar distribution can be obtained for spots that correspond to real mRNAs (multiple probes). Spots inside cells that fall into the region of overlap of these two distributions can then be assigned the corresponding probabilities of being present (real) or absent (noise). MLE can then incorporate both possible interpretations of the data, with their appropriate weights, in the data set.
A related issue, highlighted by Zenklusen et al. [11], is the existence of nascent mRNA transcripts at the locus of the gene. In the regime of continuous mRNA production, an estimate of nascent transcript number, possibly non-integer, could simply be added to mRNA counts. In the bursty regime, the existence of such transcripts might well be taken as prima facie evidence for active transcription, and therefore treated as equivalent to the presence of an above-threshold burst.
Another practical issue in reconstructing gene-expression dynamics from FISH measurements is that data may come in mixed forms, e.g. pairwise FISH data+triplet FISH data+additional constraints or prior information. Again, MLE is naturally suited to incorporating mixed data types since all sources of information can be combined to produce the overall likelihood of the data given a set of model parameters, including prior information on the model parameters themselves (cf. Eq. (2)).
While these and other practical issues are important to consider, our successful reconstruction of yeast metabolic cycles using the FISH data of Silverman et al. [14] demonstrates that our approach can provide a useful tool for analyzing gene-expression dynamics. In fact, our analysis of this data raises several new questions. First, since our reconstruction was statistically significant for the Oxidative and Reductive Building clusters but not for the Reductive Charging cluster, it is possible that cycles of the latter may be weaker in unsynchronized cultures than in synchronized chemostats. Second, our reconstruction indicates a spread among the oscillatory phases of genes within each cluster – is this spread a consequence of the limited data, or are the oscillation patterns of genes within clusters distinct? We expect that additional FISH data coupled with MLE analysis will soon provide answers to these questions.
The many advantages of FISH – absolute quantification, high time resolution, use of wild-type cells, ability to simultaneously measure multiple mRNA types, and broad application across species from bacteria [28] to yeast [11], [14] to metazoans [9], [10], suggest that FISH will find many uses in future studies of gene expression, including applications beyond those currently demonstrated. For example, FISH can be applied to cells in structured environments such as tissues or biofilms, or even cells in mixed-species consortia. In all of these cases, population level studies of gene expression cannot reveal the important cell-to-cell variations. Of course, FISH is not the only technique that yields quantitative snapshots at the single-cell level. Immunofluorescence and single-cell sequencing also meet the requirements of simultaneous measurements of two or more intracellular factors. We hope that the analysis presented here can facilitate the application of FISH and other single-cell snapshot assays to cases where both cell-to-cell variation and the dynamics of gene expression are of central interest.
|
10.1371/journal.pcbi.1002137 | Structured Pathway across the Transition State for Peptide Folding Revealed by Molecular Dynamics Simulations | Small globular proteins and peptides commonly exhibit two-state folding kinetics in which the rate limiting step of folding is the surmounting of a single free energy barrier at the transition state (TS) separating the folded and the unfolded states. An intriguing question is whether the polypeptide chain reaches, and leaves, the TS by completely random fluctuations, or whether there is a directed, stepwise process. Here, the folding TS of a 15-residue β-hairpin peptide, Peptide 1, is characterized using independent 2.5 μs-long unbiased atomistic molecular dynamics (MD) simulations (a total of 15 μs). The trajectories were started from fully unfolded structures. Multiple (spontaneous) folding events to the NMR-derived conformation are observed, allowing both structural and dynamical characterization of the folding TS. A common loop-like topology is observed in all the TS structures with native end-to-end and turn contacts, while the central segments of the strands are not in contact. Non-native sidechain contacts are present in the TS between the only tryptophan (W11) and the turn region (P7-G9). Prior to the TS the turn is found to be already locked by the W11 sidechain, while the ends are apart. Once the ends have also come into contact, the TS is reached. Finally, along the reactive folding paths the cooperative loss of the W11 non-native contacts and the formation of the central inter-strand native contacts lead to the peptide rapidly proceeding from the TS to the native state. The present results indicate a directed stepwise process to folding the peptide.
| The folding dynamics of many small protein/peptides investigated recently are in terms of simple two-state model in which only two populations exist (folded and unfolded), separated by a single free energy barrier with only one kinetically important transition state (TS). However, dynamical characterization of the folding TS is challenging. We have used independent unbiased atomistic molecular dynamics simulations with clear folding-unfolding transitions to characterize structural and dynamical features of transition state ensemble of Peptide 1. A common loop-like topology is observed in all TS structures extracted from multiple simulations. The trajectories were used to examine the mechanism by which the TS is reached and subsequent events in folding pathways. The folding TS is reached and crossed in a directed stagewise process rather than through random fluctuations. Specific structures are formed before, during, and after the transition state, indicating a clear structured folding pathway.
| In recent years, extensive investigation has been undertaken of the folding of small proteins and peptides that can be approximated as two-state folders. In these systems, only two stable populations are detected (folded and unfolded), separated by a single effective free energy barrier with only one kinetically important transition state (TS), the reaching of which can be considered as the rate limiting step [1]. The determination of the transition state ensemble (TSE) of these two-state systems and how it is traversed are, therefore, fundamental to our understanding of the physicochemical basis of protein folding [2].
Recent research on the TSE has been based on results from simulations [3]–[7], experimental mapping of site-specific contacts in the TSE by protein engineering [8], [9] and mixed experimental/computational approaches [10]–[20]. To obtain insight at residue-level detail, -values are commonly used to investigate the formation of side-chain interactions in the TSE by mutating residues and assessing the effect on folding kinetics. While -values do not, by themselves, directly provide structural information on the TS, they have been broadly used as structural restraints on a range of computational models (e.g., Go models [10], [20], Monte Carlo simulations [11], high-temperature, biased molecular dynamics (MD) [14], with both implicit solvent [15], [16] and all-atom MD representations [6]). However, it has been shown that not all conformations obtained in MD simulations by using the -value as a restraint belong to the TS [21].
An alternative approach is to obtain the transition state ensemble from the maximum of the free energy surface projected onto selected folding coordinates, as has been done in many previous studies e.g., [6], [22]–[27]. The choice of proper reaction coordinates for protein folding is non-trivial, and whether a transition state identified using any given pair of progress variables corresponds to the transition state using another pair is not known a priori. In many of the above studies it was concluded that global coordinates based on the native topology of two-state proteins or peptides fully satisfy the criteria needed to accurately identify and describe the TSE. However, the need for further analysis of the folding/unfolding probabilities to validate the transition state ensemble has been emphasized [6], [14], [24], [28]. To address this, the identification of a TSE from free energy surfaces combined with validation of the TS structures found provides a secure way to characterize the transition state for folding.
Various hypotheses have been made concerning TS structures, varying from fully native [29] and partly native [17], [30] to denatured topologies [31]. Apart from the topology, there is debate as to whether the TS represents an ensemble with a single, unique nucleus [31]–[33] or a heterogeneous population of conformations, e.g., some with structure formed and others with structure absent [16], [20], [34]. On the one hand, it has been suggested that CI2 [32] and other proteins [35] contain specific contacts in the TS which are crucial to folding. On the other, heterogeneous TS theory involves the existence of multiple transition state ensembles through which parallel folding paths pass [20].
Arguably even more challenging than the structural characterization of the folding TS is the characterization of the TS from a dynamical point of view. Information on the mechanisms by which the TS can be reached and left along reactive folding trajectories, commonly named transition-path trajectories, is scarce. For this purpose, transition path sampling [4], [5] is often used, in which, given an initial reactive path, a shooting algorithm is employed to collect transition paths by perturbing the initial path. Although this method can generate an ensemble of transition paths, an initial reactive trajectory is needed, which is commonly generated through high-temperature unfolding simulations [5]. However, as the potential energy surface sampled at higher temperature is formally different from that visited at room temperature, unfolding through high temperature may well occur through pathways that are very different from the folding routes at lower temperatures. For example, a comparison between unfolding simulations performed at elevated temperature and folding simulations at room temperature has revealed that unfolding pathways lack important intermediates and often resemble œfast-track of folding [36]. Thus, an increase in temperature may actually change the folding process rather than simply accelerating it [37]. Hence, there remain significant advantages to characterizing folding processes using long-timescale simulations at room temperature.
In the present study, we examine the transition state ensemble and folding dynamics of a model system, Peptide 1, a -hairpin peptide of 15 residues (Figure 1a) [38]. Although Peptide 1 is a designed peptide, the turn sequence, NPDG, is statistically the most abundant type I turn in proteins, enhancing the relevance of the study of its folding mechanism for natural proteins [39]. This peptide has been found to fold via a two-state mechanism in 0.8 s, as determined by the T-jump technique in combination with IR [38], [40]. In our previous work, the folding kinetics of this peptide was examined using multiple independent s-timescale all-atom MD trajectories in explicit solvent, yielding a folding time in accord with the above experimental datum [41].
Here, we derive the configurations of the TSE from the free energy folding landscape of Peptide 1 generated by multiple atomistic MD simulations over a total simulation time of 15 s. The trajectories were started from fully unfolded structures and several spontaneous folding events to the NMR-derived [38] -hairpin were observed, thus enabling a dynamical characterization of the evolution of the peptide through the folding TS. To our knowledge, this is the first time that multiple s-long explicit solvent, unbiased simulations with clear unfolding-folding transitions have been used to characterize structurally and dynamically the TS in folding studies.
The MD trajectories allow determination of the mechanism by which the TS is reached and subsequent events in folding pathways. The role of non-native interactions is characterized. It is found that, rather than being reached and crossed by highly-random fluctuations (i.e., through very different and heterogeneous pathways), folding is characterized by a directed, stagewise process involving the formation of specific structures before, during, and after the transition state for folding, corresponding to a structured folding pathway.
Six s-timescale atomistic MD simulations of Peptide 1 in explicit solvent (total of 15 s) were performed, starting from unfolded structures, and folding to the native NMR-derived conformation [38] was observed in all of them [41]. The six trajectories were used to evaluate the free energy landscape of the system using two progress variables based on the native topology of the peptide: the R parameter, containing information on the backbone, and the fraction of native contacts () containing information on the sidechain packing (Figure 1b). These two order parameters were chosen in order to capture most of the information relevant to folding and are based on the native topology. Indeed, it has been shown that for two-state peptides, such as the present, global coordinates based on the native topology fully satisfy the criteria needed to accurately identify and describe TSEs [25]. The corresponding free energy map can be divided into three distinct regions: the folded state F, the unfolded state U and a single barrier, the transition state TS (see caption of the figure for the state definitions). Concerning the sensitivity of TS structures derived from the selected reaction coordinates, an additional free-energy landscape was calculated as a function of the RMSD- pair of variables. TS structures derived from the RMSD- landscape were compared with the original TS structures of the R- plane. The overlap between the two TSE is approximately 70, which is acceptable and further supports the choice of the original progress variables. The TSE was slightly refined. We have noted that in one of the folding transitions (along trajectory 6) that was initially used to define the TSE, the peptide actually folds to a conformation that falls into the F state definition (R = 4.8, = 0.65), but is actually non-native, having a clearly non-native turn. This conformation does not populate the free energy minimum, but rather a region with a A value of 10 kJ/mol (see Figure S1). We, thus, excluded the TS structures sampled along this transition from the final TSE.
To estimate the reliability of the free energy surface the convergence of the free energies associated with individual grid cells of the plane was examined (note that free energy values are defined with respect to the grid cell corresponding to the global minimum). Figure 2a shows a typical convergence plot for two given grid cells, one in the TS (Figure 2a shows a typical convergence plot for two given grid cells, one in the TS ( 14 kJ/mol) and the other in the local minimum of the unfolded state ( 2 kJ/mol). After about 5–10 of sampling for all grid cells, the values are rather stable (within 1–2 kJ/mol). Figure 2b shows the probability distribution of the free energy standard deviations, (see Methods section for the estimate of the errors), for all the grid cells. Again, these show relatively small statistical errors in the free energy values (again within 2 kJ/mol).
Overall, the free energy landscape of Peptide 1 represents a typical two-state folder, consistent with the mono-exponential folding kinetics observed in both laser-induced temperature-jump experiments [40] and in our s MD simulations as reported previously [41]. Thus, the following minimal mechanism can be assumed:
Four possible dynamical scenarios emerge (see Figure 3a): the peptide climbs from the unfolded basin to the TS and either descends forward to the folded state (forward reactive path UTSF) or falls back to the unfolded basin (non-reactive path UTSU); or the peptide climbs from the folded basin to the TS and either descends to the unfolded state (backward reactive path FTSU) or falls back into the folded basin (non-reactive path FTSF).
The occurrence of conformations belonging to the TS, F and U states can be followed in time along the six unbiased MD trajectories. Examples of the dynamical evolution of the peptide through the three states are shown in Figure 3b. All four cases described above occur. In all four scenarios, once the TS is reached, fast recrossings of the TS surface are observed before the final descent to the end-state. Particularly remarkable is the fact that many of the TS structures, which were extracted based on a thermodynamic criterion, indeed occur right at the folding and unfolding transitions along the (unbiased) trajectories (top and middle panels of Figure 3b), thus confirming the validity of the thermodynamic TS selection. In addition, TS structures occurring in non-reactive paths (bottom panels of Figure 3b) are also observed, corresponding to very short excursions into the TS compared to reactive paths.
In what follows, the kinetics and structural features of the TSE are presented. Furthermore, to determine what triggers the folding events, the TS structures visited along forward reactive paths are examined in detail.
To characterize the TS kinetics, the TS lifetime (), here defined as the mean residence time in the TS, was evaluated. The distribution of the residence times (Figure 4) was fitted by a monoexponential function (dotted line) yielding a of 10.20.5 ps (for the estimate of the error see Methods), with 52.7% and 47.3% of the TS population ending up in the folded and unfolded states, respectively. The correlation coefficient is higher than 0.9999, showing the goodness of the fit. These data are consistent with a two-state kinetic model [42], thus further indicating that the selected TSE is reliable. The relatively short (ps-timescale) TS lifetime obtained, showing that the TS is a short-lived state, is consistent with the fast recrossings of the TS observed prior to the final descent (see Figure 3b).
In order to examine the degree of native topology in the TSE, the formation of turn, middle and end-to-end inter-strand contacts was evaluated and compared with the corresponding contacts in the folded conformations. Figure 5a shows the distributions of the distances between the C atoms of residues S1-E15 (end), I4-T12 (middle) and N6-G9 (turn) in the TSE. The distributions are quite narrow, indicative of a rather homogeneous topology of the TSE population. However, a minor peak is present in the middle distribution at 0.55 nm that will be discussed below. The “end” and “turn” distributions in the TSE, peaked at 0.42 and 0.55, respectively, are native-like, the only slight difference being a shift of the “turn” peak maximum to longer values (0.55 nm versus 0.60 nm in F and TSE, respectively). In contrast, the middle-contacts distribution differs significantly in the two states, the peak maximum being at a much longer distance in the TSE (0.72 nm) than in F (0.54 nm). Thus, the distinct feature of most of the TSE is the formation of native-like end-to-end and turn contacts, while the central parts of the strands are disordered. The presence of the end-to-end contact is consistent with recent experimental results on several proteins suggesting that the TS exhibits an overall native-like topology in which the N-terminal and C-terminal regions are in close proximity [17], [32], [43]. As was found for the backbone topology, the native sidechain contacts at the end of the hairpin are present in the TSE (data not shown). Contrarily, the sidechains of the central parts exhibit some persistent non-native interactions. In particular, W11, the bulkiest of the sidechains, forms non-native contacts with two residues of the turn, namely P7 and G9, in a kind of key-lock configuration (see Figure 5b, in which the distributions are plotted of the minimum distances between the W11 sidechain and the P7 and G9 sidechains, together with representative configurations).
To analyze the origin of the above-mentioned minor peak in the “middle” distribution, possible differences in the topological features of TS structures sampled along reactive (UTSF and FTSU) and non-reactive (FTSF and UTSU) pathways were investigated. Separating the TSE into UTSF, FTSU, FTSF and UTSU structures is also useful for identifying the presence, if any, of a specific folding nucleus, i.e., a nucleus of contacts resulting in rapid assembly of the native state [24], [44], thus triggering the descent of the peptide from the TS to the folded basin. The folding nucleus is more likely to be retained in the FTSF conformations than in the UTSU conformations [24], [44].
Therefore, we calculated the distribution of “turn”, “middle”, “end-to-end” and W11-P7/W11-G9 contacts in the TS structures belonging to UTSF, FTSU, FTSF and UTSU paths separately (see Figure 6). It is found that, TS structures apart from the FTSF population are very homogeneous, with the topological features described above (i.e., the turn and end-to-end contacts are formed, the central part of the backbone is disordered and the W11 sidechain is positioned in the middle of the turn at around 0.55 nm from P7 and G9). In contrast, however, in about 65 of the FTSF structures (brown color in Figure 6), which correspond to the minor peak mentioned above, the middle backbone contacts are formed, the end-to-end contacts is slightly looser and the W11 sidechain is closer to P7 (at a distance of 0.3 nm) than to G9. Hence, according to the definition given above, the folding nucleus is likely to be characterized by an almost-native backbone topology and a persistent, non-native sidechain contact between W11 and the P7 turn residue.
The mechanism by which the peptide crosses the TS is now examined. To determine whether the end and turn contacts are formed only at the TS or prior to it, the time evolution of the corresponding C-C distances was calculated. Figure 7 shows these time series for four representative forward reactive paths. Clearly, prior to TS the turn is already formed, while the end-to-end contact forms only at the TS. These results indicate that end-to-end contact formation is a discriminating feature for the TSE of Peptide 1. This trend is observed in all forward reactive paths.
The final stage of the folding process, i.e, the descent of the peptide from the TS to the F state, involves the correct arrangement of the middle part of the hairpin. Figure 8 plots the C-C distances of the end (S1-E15), middle (Y4-T12) and turn (N6-G9) residues again as a function of time for four representative reactive folding paths. The end and turn regions of the hairpin remain in contact during the final stage, while the middle part comes into contact.
Finally, the time evolution of the non-native contacts formed by the W11 sidechain with the turn is analyzed. These contacts are present already prior to the reaching the TS, locking the P7-G9 residues into the “reactive” turn conformation (see Figure 9). In the final stage of the process, the W11 sidechain loses the non-native contacts with the turn, allowing a concomitant rearrangement of the central segments of the strands into the native -hairpin conformation (see Figure 9).
The above results indicate a clear TS folding mechanism that is summarized in Figure 10. In this mechanism, the TS is reached with the formation of end-to-end and turn contacts, with the turn appearing prior to the TS and the end-to-end interaction appearing only at the TS. Reaching the folded state from the TS requires W11 sidechain repositioning and concomitant native rearrangement of the central segment of -hairpin. However, despite a common TS topology and a single, directed folding-transition mechanism was observed from extensive MD simulation, the existence of minor folding pathways cannot be excluded.
A variety of computational models and methods has been applied to extract information on the TSE for protein folding, including Go models [6], [10], [45], high temperature [14] and implicit solvent [15], [16] MD simulations, and transition path sampling [4], [5]. However, the uncertainty involved in these methods (e.g., unnaturally high temperature or absence of explicit water) leads to uncertainty in the interpretation of the data [46]. To our knowledge, no simulation studies exist that have characterized the TSE using unbiased atomistic simulation in explicit solvent, in which folding events from fully unfolded conformations occur. In the present work, we have used this information to directly evaluate TS structure and kinetics from multiple s-timescale all-atom explicit MD simulations of the -hairpin Peptide 1.
The main findings of the study can be summarized as follows. Structurally, the TS consists of a rather homogeneous loop-like topology characterized by native end-to-end contacts and specific non-native interactions in the middle region of the loop characterized by the W11 sidechain centrally locked in between the turn residues P7 and G9. A specific folding nucleus, i.e., the nucleus of contacts resulting in rapid assembly of the native state, was also identified. This is characterized by loose end-to-end contacts and the presence of native contacts in the middle part of the hairpin, with a concomitant shift of the W11 sidechain towards the sidechain of P7.
Concerning the folding mechanism, while the turn and its non-native contacts with the W11 sidechain are already formed prior to the TS, the end-to-end interaction appears only at the TS, making it a unique feature of the TS. The TS is short-lived, with a mean lifetime on the ps-timescale. The final stage in reaching the native state, occurring after crossing the TS, involves the cooperative loss of the non-native contacts and formation of the native inter-strand contacts in the central part of the hairpin. These last events are those committing the reactive trajectories to very rapidly proceeding from TS to F.
The present results support previous experimental findings on other systems suggesting that the formation of end-to-end contacts in the TS may be a fairly general phenomenon in the folding of small proteins [17], [32], [43]. In the present case, there exists an additional structural feature specific to -hairpins that further restricts the conformational variety of the TS structures, namely the formation of the turn.
From a dynamical point of view, W11 acts as a chaperone for reaching the TS through the formation of non-native interactions with the turn. The W sidechain locks the turn in place prior to TS, allowing the ends to subsequently come into contact, i.e., the TS to be reached. Hence, also from a dynamical point of view our results are in favour of a “directed”, stagewise process [11], rather than a large number of heterogeneous pathways [47], [48] characterizing the reaching and crossing of the TS. Further studies on other peptides and proteins will clarify the generality of the present observation of the structuring of the pathway across the transition state for folding.
A series of six 2.5 s-long atomistic MD simulations of Peptide 1 (SESYINPDGTWTVTE) in explicit solvent was performed [41]. Six starting structures representing the unfolded state were extracted randomly from a simulation of 50 ns that was started from a fully extended conformation of the peptide. The MD simulations were performed with the program GROMACS [49] with the OPLS-AA all-atom force field [50] for the peptide. The water was modelled using the TIP4P representation [51]. Each of the six starting conformations was placed in a dodecahedral water box large enough to contain the peptide and at least 1.0 nm of solvent on all sides (volume48 nm). Each simulation box contained 6647 atoms. Periodic boundary conditions were applied and the long range electrostatic interactions were treated with the Particle Mesh Ewald method [52] using a grid spacing of 0.12 nm combined with a fourth-order B-spline interpolation to compute the potential and forces in-between grid points. The real space cut-off distance was set to 0.9 nm and the van der Waals cut-off to 1.4 nm. The bond lengths were fixed [53] and a time step of 2 fs for numerical integration of the equations of motion was used. Simulations were performed in the NVT ensemble with isokinetic temperature coupling [54] keeping the temperature constant at 300 K. Three Na counterions were added, replacing three water molecules, so as to produce a neutral simulation box.
All the starting structures were subjected to a two-stage energy minimization protocol using the steepest descent method. The first minimization was performed with the coordinates of the peptide held fixed, allowing only the water and the ions to move, and the second was performed on the atoms of both the peptide and the solvent molecules. The temperature of the system was then increased from 50 K to 300 K in 500 ps of MD before the 2.5 production simulations were started. For all the analyses conformations saved every 10 ps were used.
The trajectories were analyzed using order parameters that capture principal aspects of the folding process of the peptide. A robust parameter for identifying conformational transitions is ‘R’ [55], calculated as follows:where R is the i inter-strand C-C distance in the native NMR structure and R is the same distance in the MD. The five inter-strand C-C pairs in the Peptide 1 hairpin are the following: N6-T10, I5-W11, Y4-T12, S3-V13 and E2-T14. A value of R5 indicates formation of the native -hairpin. Sidechain packing was quantified via the fraction of native sidechain contacts, (a contact between two sidechains is considered to be formed when the minimum distance between the atoms belonging to the sidechains is 0.55 nm).
Given a system in thermodynamic equilibrium, the change in free energy on going from a reference state, ref, of the system to a generic state, i, (e.g., from unfolded to folded) at constant temperature and constant volume was evaluated as(1)where R is the ideal gas constant, T is the temperature and p and p are the probabilities of finding the system in state i and state ref, respectively. We describe the free energy surface as a function of two order parameters, namely the fraction of native contacts, , and the R parameter. A grid 40×40 was used to divide this plane in 1600 cells and for every cell the number of points counted and the relative probability calculated, allowing A to be determined. The reference state was chosen to be the grid cell with the highest probability, which corresponds to folded -hairpin structures.
The statistical error in different properties evaluated from the simulations, such as the TS lifetime or the free energy values, was estimated through the standard error of their mean, , calculated over subsets of the trajectories:(2)(3)(4)where is the mean of the given parameter evaluated in the i subset, and and are the mean value over the samples and the sample standard deviation, respectively. Here we used 3 independent subsets of the trajectories (, i.e. 3 groups each consisting of two trajectories), which was found to be a good compromise between the statistics within each subset and the sample size, . Assuming, as usual, a normal distribution of the mean value, , the expected value of lies with 95% confidence inside the interval.
|
10.1371/journal.pgen.1003114 | Histone Deacetylase HDA6 Is Functionally Associated with AS1 in Repression of KNOX Genes in Arabidopsis | ASYMMETRIC LEAVES 1 (AS1) is a MYB-type transcription repressor that controls leaf development by regulating KNOX gene expression, but the underlying molecular mechanism is still unclear. In this study, we demonstrated that AS1 can interact with the histone deacetylase HDA6 in vitro and in vivo. The KNOX genes were up-regulated and hyperacetylated in the hda6 mutant, axe1-5, indicating that HDA6 may regulate KNOX expression through histone deacetylation. Compared with the single mutants, the as1-1/axe1-5 and as2-1/axe1-5 double mutants displayed more severe serrated leaf and short petiole phenotypes. In addition, the frequencies of leaf lobes and leaflet-like structures were also increased in as1-1/axe1-5 and as2-1/axe1-5 double mutants, suggesting that HDA6 acts together with AS1 and AS2 in regulating leaf development. Chromatin immunoprecipitation assays revealed that HDA6 and AS1 bound directly to KNAT1, KNAT2, and KNATM chromatin. Taken together, these data indicate that HDA6 is a part of the AS1 repressor complex to regulate the KNOX expression in leaf development.
| AS1 is a MYB-type transcription repressor that controls leaf patterning by repressing class-1 KNOX gene expression. The molecular mechanism by which AS1 represses KNOX gene expression is still unclear. In this study, we found that AS1 interacted with the histone deacetylase HDA6. Furthermore, HDA6 repressed KNOX gene expression by histone deacetylation. hda6 mutants displayed serrated leaf and short petiole phenotypes. Additionally, hda6/as1-1 double-mutant plants showed a more severe phenotype compared to the single mutants, indicating that HDA6 may act together with AS1 in controlling leaf development. Taken together, our data indicated that HDA6 is an important component of the AS1 repressor complex in regulating the KNOX gene expression.
| The initiation of leaf primordia is established by recruitment of cells from the flanks of the shoot apical meristem (SAM). Meristem activity in the shoot apex is specified in part by the class I KNOTTED-LIKE HOMOBOX (KNOX) genes [1]–[3]. Lateral organs, such as leaves, are initiated on the flank of SAM, and down-regulation of KNOX genes is essential to facilitate this process [1], [4]. Moreover, the silencing of KNOX genes is important in developing organs since the ectopic KNOX expression during organogenesis resulted in patterning defects and over-proliferation of cells [5]–[7]. Thus, the balance between stem cell differentiation and proliferation that is decisive for plant development is attained, in part through the proper regulation of the KNOX expression.
In Arabidopsis, the KNOX family can be further divided into three classes. Class I KNOX genes are similar to KNOTTED1 (KN1) in maize, including BREVIPEDICELLUS (BP)/KNAT1, KNAT2, KNTA6 and SHOOTMERISTEMLESS (STM). These genes are expressed in the SAM and down-regulated in leaf primordia [8]. Class II KNOX genes comprise KNAT3, KNAT4, KNAT5 and KNAT7, which are broadly expressed. Class III only contains KNATM, which is a novel KNOX gene lacking the homeodomain. It was demonstrated that KNATM functions together with KNAT1 and BELL proteins by forming heterodimer [9]. Moreover, ectopic expression of KNATM resulted in the curled down and serrated rosette leaves in wild type plants [9].
KNOX repression is mediated by the orthologous MYB domain proteins ROUGH SHEATH2 (RS2) in maize (Zea mays) and ASYMMETRIC LEAVES1 (AS1) in Arabidopsis thaliana [10]–[13]. In addition, AS1 interacts with the LATERAL ORGAN BOUNDARIES (LOB) domain protein AS2 and directly represses the expression of BP/KNAT1 and KNAT2 [14]–[16]. Previous studies revealed that AS1 and AS2 may recruit a chromatin-remodeling protein Histone Regulatory Homolog 1 (HIRA) to regulate the expression of target genes [17]. Moreover, HIRA has also been shown to interact with a histone deacetylase (HDAC) in animal cells [18].
In this study, we investigated the interaction of AS1 with the histone deacetylase HDA6 and their involvement in leaf development. We demonstrated that HDA6 can interact with AS1 in vivo and in vitro. The hda6 mutant, axe1-5, displayed curling and serrated leaves as well as shorter petioles, suggesting that HDA6 is involved in leaf development. Additionally, HDA6 and AS1 associate directly with the promoters of KNAT1, KNAT2 and KNATM. Taken together, our data suggest that HDA6 is a part of the AS1 repression complex to regulate the expression of KNOX genes.
AS1 is a MYB-type transcription repressor that controls leaf patterning by repressing class-1 KNOX gene expression [16]. However, the molecular mechanism how AS1 represses KNOX gene expression is still unclear. In yeast and mammalian cells, many transcription repressors were found to recruit HDACs to regulate their target genes [19]. To further understand the molecular mechanism of AS1-dependent KNOX repression, we analyzed the interaction of AS1 with HDA6, a RPD3-type HDAC in Arabidopsis [20], [21] by using BiFC assays. The coding sequences of HDA6 and AS1 were fused to the N-terminal 174-amino acid portion of yellow fluorescent protein (YFP) in the pEarley-Gate201 vector (pEarleyGate201-YN) or the C-terminal 66-amino acid portion of YFP in the pEarleyGate202 vector (pEarleyGate202-YC) [22]. The Agrobacterium cells containing these constructs were co-transfected into Nicotiana benthamiana leaves. The yellow fluorescence was observed at the nuclear when HDA6-YN and AS1-YC were transient expressed in N. benthamiana leaves, indicating that HDA6 interacted with AS1 in vivo (Figure 1A). In contrast, the yellow fluorescence was not observed in the negative controls (Figure S1).
The interaction between HDA6 and AS1 was further confirmed by in vitro pull down assays. When purified MBP-AS1 recombinant protein was incubated with glutathione S-transferase (GST)-HDA6 protein, HDA6-GST was pulled down by MBP-AS1 (Figure 1B), indicating that HDA6 was directly associated with AS1. Co-immunoprecipitation (CoIP) assays were also used to analyze the interaction between HDA6 and AS1. A stable transgenic plant expressing 35S:GFP-HDA6 in the hda6 mutant (axe1-5) was generated [23]. Overexpressing 35S:GFP-HDA6 in axe1-5 complemented the mutant phenotype, suggesting that the GFP-HDA6 fusion protein is functional. Crude extracts (input) of axe1-5, as1-1 and axe1-5/35S:GFP-HDA6 were immunoprecipitated by the AS1 antibody, then analyzed by western blotting. As shown in Figure 1C, GFP-HDA6 was clearly co-immunoprecipitated by endogenous AS1. Furthermore, AS1 protein was also co-immunoprecipitated by GFP-HDA6 when immunoprecipitated by the GFP antibody (Figure 1C). Taken together, our data strongly indicate that HDA6 interacts with AS1 in vitro and in vivo.
Previous studies indicated that AS1 and AS2 can associate together both in yeast cells by yeast two-hybrid assays and in vitro by ELISA experiments using purified His-AS1 and GST-AS2 recombinant proteins [14]. By using BiFC assays, we also found that AS1 and AS2 can interact with each other in N. benthamiana leaves (Figure S2). Furthermore, both AS1 and AS2 can also interact with itself (Figure 2A). These observations indicated that AS1 and AS2 can form both homo- and hetero-dimers. The yellow fluorescence was observed at the nucleus when AS1-YN and AS1-YC, AS2-YN and AS2-YC, or AS1-YN and AS2-YC were transient expressed in N. benthamiana leaves (Figure 2A and Figure S2). Moreover, the in vivo interaction between HDA6 and AS2 was also found by using BiFC (Figure 2B). Collectively, these results together with the finding that HDA6 interacts with AS1 suggested that HDA6, AS1 and AS2 function together in the same protein complex.
We further tested the protein-protein interactions among HDA6, AS1 and AS2 in the protoplasts isolated from the mutants. By using BiFC assays, we found that HDA6 interacted with AS1 in the nucleus of as2-1 mutants (Figure S3). Likewise, the interaction of HDA6 and AS2 was also found in the nucleus of as1-1 mutants. In addition, we also showed that AS1 interacted with AS2 in the nucleus of axe1-5 mutants. Our data indicate that loss of one component of HDA6, AS1 and AS2 does not affect the interaction of two others in Arabidopsis.
Previously, we reported that the Arabidopsis HDA6 is required for flowering time control and the hda6 mutant, axe1-5, displayed a delayed flowering phenotype [23]. In addition, axe1-5 mutants also displayed the curling leaves under both long-day (LD) and short-day (SD) conditions (Figure 3A). Similar curling and serrated leaves were also found in another hda6 mutant, sil1 [25] (Figure 3A), and the HDA6-RNAi plants (Figure S4). hda6 mutants displayed the down curling phenotype on both the distal and lateral axis (Figure 3A). These results demonstrated that HDA6 functions not only in controlling adaxial-abaxial axis, but also in proximal-distal axis and in medial-lateral axis.
The as1 and as2 mutants of Arabidopsis thaliana exhibit pleiotropic phenotypes in leaf development, including the curling and serrated leaves [26]. To examine the genetic interaction between HDA6 and AS1 or AS2, we generated as1-1/axe1-5 and as2-1/axe1-5 double mutants and compared the leaf phenotype of single and double mutants. Under LD conditions, as1-1/axe1-5 and as2-1/axe1-5 double mutant plants showed more severe leaf phenotypes compared with as1-1 and as2-1 single mutant plants (Figure 3B and 3C).
We also measured the lengths of petioles and lamina in wild type and mutant plants. Compared with wild type, the lengths of the petioles were decreased in axe1-5 mutants (Figure 3D). as1-1/axe1-5 and as2-1/axe1-5 double mutants displayed shorter petioles compared with as1-1 and as2-1 single mutant plants (Figure 3B, 3C and Figure 3D). However, the lamina lengths of as1-1/axe1-5 and as2-1/axe1-5 did not show significant changes compared with the single mutants (Figure 3E).
We further measured the frequencies of leaf lobe formation in axe1-5, sil1, as1-1/axe1-5 and as2-1/axe1-5 mutants. The frequencies of leaf lobes were significantly increased in as1-1/axe1-5 and as2-1/axe1-5 double mutants (Table 1). as2 mutants produced leaflet-like structures on the petioles [26]. In as1-1/axe1-5 and as2-1/axe1-5 double mutants, the frequencies of leaflet-like structures were increased (Table 2), and some of the leaf lobes were similar to leaflet-like structures (Figure 3C). These results suggested that HDA6 acts with AS1 and AS2 in regulating leaf development.
We further analyzed the gene expression by quantitative reverse transcription (qRT)-PCR in mutant plants. Compared with Col wild type, no significant changes were found in the expression of AS1 and AS2 in the axe1-5 (Figure S5). As shown in Figure 4, the expression of KNAT1, KNAT2 and KNATM was increased in axe1-5 compared to Col wild type. Consistent with the previous study [13], the transcript levels of KNAT1 and KNAT2 were elevated in as1-1 and as2-1 mutant plants. In addition, the expression of KNATM was also up-regulated in as1-1 and as2-1 mutant plants. Moreover, the expression of KNAT1, KNAT2 and KNATM was highly increased in as1-1/axe1-5 and as2-1/axe1-5 double mutants compared with their corresponding single mutants. These data indicate that HDA6 may function synergistically with AS1 and AS2 in regulating the expression of KNOX genes.
We also analyzed the expression of PHB, PHV, CUC1 and CUC2, which were involved in leaf development through the miRNA regulated pathway [27]–[30]. However, no significant different was found in the expression of PHB, PHV, CUC1 and CUC2 (Figure S5).
To determine whether the high expression of KNOX genes in the mutants is related to histone hyperacetylation in chromatin, ChIP assays were used to analyze the histone H3 acetylation levels of KNAT1, KNAT2 and KNATM. The relative enrichment of histone H3 acetylation was determined by real-time PCR using primers specific for the proximal promoter (within 500 bp upstream of the transcription starting sites) and transcription start regions of individual genes. As shown in Figure 5, levels of histone H3 acetylation were slight elevated in the proximal promoter and transcription start regions of KNAT1, KNAT2 and KNATM in axe1-5, suggesting that HDA6 may regulate these genes expression by chromatin deacetylation. We further analyzed histone acetylation levels of KNAT1, KNAT2 and KNATM in as1-1, as2-1 and the double mutants. As shown in Figure 5B, hyperacetylation of histone H3 was found in the promoter and first exon of KNAT1, KNAT2 and KNATM in as1-1/axe1-5 and as2-1/axe1-5 double mutants. In contrast, hyperacetylation of histone H3 was not found in as1-1 and as2-1 single mutants. These results suggested that hyperacetylation of histone H3 in KNAT1, KNAT2 and KNATM found in as1-1/axe1-5 and as2-1/axe1-5 double mutants was caused by the hda6 mutation.
Histone H3K4Me3 is another chromatin mark associated with active genes. We also investigated the histone H3K4Me3 level in axe1-5 mutants. However, no significant changes in the H3K4Me3 of KNAT1, KNAT2 and KNATM were found (Figure S6A). H3K9Me2 was reported as a chromatin marker associated with gene repression. No significant changes in the level of histone H3K9Me2 was found in axe1-5 mutants (Figure S6B).
The direct association between AS1 and HDA6 suggested that AS1 may recruit HDA6 to repress the downstream target genes. Previous studies demonstrated that the AS1 repressor complex binds directly to the regulatory motif I (CWGTTD) and motif II (KMKTTGAHW) on the promoters of the KNAT1 and KNAT2 [16]. We also found the conserved motif I and motif II in two promoter regions (KNAMT-X and KNAMT-Y) of KNATM (Figure 6A and Figure S7). To investigate whether AS1 binds directly to KNAT1, KNAT2 and KNATM, ChIP analyses using the AS1 antibody were performed in Col wild type and as1-1 mutants. Consistent with the previous report [16], AS1 can bind to the promoters of KNAT1 and KNAT2 (Figure 6B). In addition, AS1 can also bind directly to KNATM (Figure 6B). In comparison, AS1 cannot bind to the control genes, ACTIN2 and TUB2. To analyze whether the binding of AS1 to KNAT1, KNAT2 and KNATM requires the presence of AS2, we also performed ChIP assays using the as2-1 mutants. We found the loss of binding of AS1 to the KNOX chromatin in the as2-1 mutant (Figure 6C), suggesting that AS2 is required for the binding of AS1 to the KNOX genes.
To examine whether HDA6 can binds directly to KNAT1, KNAT2 and KNATM., transgenic plants expressing HDA6-Myc were subjected to ChIP analysis using an anti-Myc antibody. As shown in Figure 6D, ChIP analyses revealed that HDA6 can bind to the promoters of KNAT1, KNAT2 and KNATM. We also analyze whether HDA6 recruitment is dependent on AS1. ChiP assays were performed using an anti-Myc antibody in transgenic plants expressing the HDA6-Myc in as1 mutants. As shown in Figure 6D, HDA6 cannot bind to KNAT1, KNAT2 and KNATM in as1 mutants, suggesting that AS1 is required to recruit HDA6.
To analyze whether the HDA6 binding is dependent on its catalytic activity, we performed ChIP assays using an anti-FLAG antibody in transgenic plants (HDA6 all 5 mut in axe1-5) expressing the FLAG-tagged HDA6 bearing the five amino acid mutation of the active site in axe1-5 mutants [31]. As show in Figure 6E, the active site mutant HDA6 can still bind to KNAT1, KNAT2 and KNATM, suggesting that HDA6 recruitment is independent of its catalytic activity. Taken together, our findings suggested that HDA6, AS1 and AS2 act together and directly repress the expression of KNOX genes in Arabidopsis.
The Arabidopsis genome sequence contains 9 KNOX genes, which can be further classified into 3 classes [32]. In leaves, AS1 and AS2 down-regulate class I KNOX genes, but not STM; conversely, STM represses AS1 expression in the SAM [12], [33]. Downregulation of KNOX genes expression is a vital step in leaf initiation, and silencing of these genes needs to be maintained for normal organogenesis [13], [15]. In this study, we demonstrated that hda6 mutants displayed the curling and serrated leaves and shorter petioles. Compared with the single mutants, as1-1/axe1-5 and as2-1/axe1-5 double mutants show more severer phenotypes on curling leaves, petiole lengths, and leaflet-like structures, supporting that HDA6 acts synergistically with AS1 and AS2 in the regulation of leaf development.
KNAT1 and KNAT2 were previously found to be repressed by AS1 and AS2 [14]–[16]. Our results indicated that the transcript levels of KNAT1, KNAT2 and KNATM were altered in axe1-5, as1-1 and as2-1 mutants. Furthermore, the expression of KNAT1, KNAT2 and KNATM was highly increased in as1-1/axe1-5 and as2-1/axe1-5 double mutants compared to their corresponding single mutants. In addition, levels of histone H3 acetylation was elevated in KNAT1, KNAT2 and KNATM loci in axe1-5, as1-1/axe1-5 and as2-1/axe1-5 mutants, suggesting that HDA6 is required for the repression of KNOX genes by chromatin deacetylation. ChIP analyses revealed that HDA6 and AS1 bound directly to the promoters of KNAT1, KNAT2 and KNATM. These data indicate that HDA6 and AS1 function together in controlling KNOX gene expression through histone dacetylation. In addition, AS1 is required to recruit HDA6 in KNOX repression HDA6 cannot bind to KNAT1, KNAT2 and KNATM in as1 mutants, suggesting that AS1 is required to recruit HDA6 in KNOX repression.
Microarray gene expression analyses revealed that a large number of loci are differently expressed in hda6 mutants [23], [34], indicating that HDA6 may play multiple roles in different development processes. Recent studies suggested that the expression of KNOX genes is only one important factor for leaf development [28], [30]. Further analysis is required to determine whether HDA6 is involved in other leaf development pathways.
AS1 is a Myb domain transcription factor related to RS2 in maize and PHANTASTICA in Antirrhinum [12]. Mutations in AS1 result in abnormal leaves, with marginal outgrowths or lobes [12], [13], [33], [35] AS2 encodes a LOB domain protein containing a leucine-zipper motif [36]–[38]. Mutations in the as2 gene cause a phenotype similar to as1 mutants [13], [33]. Previous studies indicated that AS1 and AS2 can associate together both in vitro and in yeast cells [14]. By using the BiFC assay, we found that AS1 and AS2 can interact and form homo and hetero-dimer in plant cells. These data suggested that AS1 and AS2 function in the same protein complex.
A recent study indicated that AS1 functions as a transcriptional repressor and binds directly to its KNOX targets when in a complex with AS2 [16]. It was found that the AS1–AS2 repressor complex binds directly to the regulatory motif I (CWGTTD) and motif II (KMKTTGAHW) in the promoters of the KNAT1 and KNAT2 [16]. Similar to KNAT1 and KNAT2, we also found the conserved motif I and motif II in the promoter of KNATM. KNATM is a novel Arabidopsis Class III KNOX gene that has a MEINOX domain but lacks the homeodomain [9]. ChIP assayes revealed that AS1 can bind directly to the promoter regions of KNAT1, KNAT2 and KNATM. These data suggested that in addition to KNAT1 and KNAT2, the AS1–AS2 complex is also targeted to KNATM by binding to the conserved motifs I and II. To our knowledge, this is the first study demonstrating that KNATM is regulated by AS1 and AS2.
Recent studies suggested that the AS1–AS2 complex binds to the KNAT1 and KNAT2 promoters and recruit the chromatin-remodeling protein HIRA to maintain the chromatin in a stable repressive state [15], [16], [39]. In mammalian cells, HIRA was shown to interact with a histone deacetylase [18]. Moreover, it was observed that Arabidopsis seedlings treated with TSA, an inhibitor of HDACs, produced abaxialized filamentous leaves, indicating the involvement of HDACs in leaf morphogenesis [24]. In this study, we provided direct evidence indicating that HDA6 is involved in leaf morphogenesis by interacting with AS1 and AS2 to regulate the KNOX expression. Compared with the single mutants, as1-1/axe1-5 and as2-1/axe1-5 double mutants show more severe phenotypes on curling leaves, petiole lengths, and leaflet-like structures, supporting that HDA6 acts with AS1 and AS2 to regulate leaf development. Taken together, our results demonstrated that histone deacetylation is one of the epigenetic components involved in AS1–AS2 complex-mediated KNOX repression. HDA6 may therefore be part of the AS1–AS2 repression complex to repress the target gene expression. Our data indicate that loss of one component of HDA6, AS1 and AS2 does not affect the interaction of two others in Arabidopsis. Previous studies indicated that the interaction between AS1 and AS2 is required for their binding to the promoters of KNOX genes, because neither AS1 nor AS2 alone was able to bind to the target DNA sequences in vitro [40]. We observed the loss of binding of AS1 to the KNOX chromatin in the as2-1 mutant, suggesting that AS2 is required for the AS1 binding. Furthermore, HDA6 cannot bind to KNOX chromatin in as1-1 mutants, indicating that AS1 is required to recruit HDA6. Taken together, both AS1 and AS2 are required for the recruitment of HDA6 to chromatin in repression of KNOX genes.
A recent work has also shown that the Polycomb Repressive Complexes (PRCs) repress KNOX transcription [40]. It was found that CLF-containing PRC2 regulates KNOX genes by trimethylation of histone H3K27 [41]. Thus, AS1 and AS2 may also recruit other chromatin factors such as PRCs to regulate class I KNOX genes. Taken together, our results suggested that HDA6 is one of the epigenetic components involved in the AS1–AS2 complex-mediated KNOX repression during leaf development in Arabidopsis.
Arabidopsis thaliana was grown in 23°C under LD (16 h light/8 h dark) or SD (8 h light/16 h dark) conditions. axe1-5, sil1, as1-1 and as2-1 are in the Col background, whereas the HDA6 RNAi lines CS24038 and CS24039 are in Ws background.
Arabidopsis leaves (0.2 g) were ground with liquid nitrogen in a mortar and pestle and mixed with 1 ml Trizol Reagent (Invitrogen) to isolate total RNA. After treated with DNase (Promega), two microgram of total RNA was used for the first-strand cDNA synthesis. cDNA was synthesized in a volume of 20 µl that contained the Moloney Murine Leukemia Virus Reverse Transcriptase buffer (Promega), 1.5 µM poly(dT) primer, 0.5 mM deoxyribonucleotide triphosphates, 25 units RNasin ribonuclease inhibitor, and 200 units Moloney Murine Leukemia Virus Reverse Transcriptase at 37°C for 1 h.
cDNAs obtained from reverse transcription were used as a template to run real-time PCR. The following components were added to a reaction tube: 9 µL of iQ SYBR Green Supermix solution (Bio-Rad), 1 µL of 5 µM specific primers, and 8 µL of the diluted cDNA template. Thermocycling conditions were 95°C for 3 minutes followed by 40 cycles of 95°C for 30 s, 60°C for 30 s, and 72°C for 20 s, with a melting curve detected at 95°C for 1 minute, 55°C for 1 minute, and detected the denature time from 55°C to 95°C. Each sample was quantified at least triplicate and normalized using Ubiquitin 10 as an internal control. The gene-specific primer pairs for quantitative RT-PCR are listed in Table S1.
ChIP assay was carried out as described [42]. Chromatin extracts were prepared from 10 day old seedlings treated with formaldehyde. The chromatin was sheared to an average length of 500 bp by sonication and immunoprecipitated with specific antibodies including anti-acetylated histone H3K9K14 (Catalogue no. 06-599, Millipore), anti-trimethylated histone H3K4 (Catalogue no. 04-745, Millipore), anti-c-Myc (Catalogue no. M4439, Sigma) and anti-FLAG (Catalogue no. F1804, Sigma). The DNA cross-linked to immunoprecipitated proteins was analyzed by real-time PCR. Relative enrichments of various regions of KNAT1, KNAT2 and KNATM in axe1-5, as1-1 and as1-1/axe1-5 over Col were calculated after normalization to ACTIN2. Each of the immunoprecipitations was replicated three times, and each sample was quantified at least in triplicate. The primers used for real-time PCR analysis in ChIP assays are listed in Table S2.
To generate the constructs for BiFC, full-length coding sequences of HDA6, AS1 and AS2 were PCR-amplified using Pfu polymerase (Finnzymes). The PCR products were subcloned into the pENTR/SD/D-TOPO or pCR8/GW/TOPO vector and then recombinated into the pEarleyGate201-YN and pEarleyGate202-YC vectors [22]. The resulting constructs were transformed into the Agrobacterium GV3101 and the Agrobacteria containing these constructs were cotransfected into five week old Nicotiana benthamiana leaves. For the protoplast transient expression, HDA6, AS1 and AS2 fused with pEarleyGate201-YN or pEarleyGate201-YC were co-transfected into protoplasts by PEG transfection [43]. Transfected leaves and protoplasts were imaged using TCS SP5 (Leica) Confocal Spectral Microscope Imaging System.
Pull-down assays were performed as previously described [44] with some modifications. 2 µg Myelin basic protein (MBP) and MBP-AS1 recombinant proteins were incubated with 30 µl of MBP resin in a total volume of 500 µl of MBP binding buffer (20 mM Tris-HCl, pH 7.5, 200 mM NaCl, 1 mM EDTA) for 2 h at 4°C, and the binding reaction was washed 3 times by the binding buffer, then 2 µg GST-HDA6 recombinant protein was added and incubated for additional 2 h at 4°C. After extensive washing (at least 8 times), the pulled down proteins were eluted by boiling, separated by 10% SDS-PAGE, and detected by western blotting using an anti-GST antibody.
Coimmunocipitation assays were performed as previous described [23]. The 20-day-old axe1-5/35S:GFP-HDA6, axe1-5 and as1-1 plants were harvested and ground in liquid nitrogen. Total proteins were extracted in an extraction buffer (50 mM Tris-HCl, pH 7.4, 150 mM NaCl, 2 mM MgCl2, 1 mM DTT, 20% glycerol, and 1% CA-630) containing protease inhibitor cocktail (Roche). Cell debris was pelleted by centrifugation at 14,000 g for 30 min. The supernatant was incubated with anti-AS1 or anti-GFP specific antibody overnight at 4°C by gently rotation, then 50 µl of protein G agarose beads (Millipore) was added. After 3 h of incubation at 4°C by gently rotation, the beads were centrifuged and washed five times with a washing buffer (50 mM Tris-HCl, pH 7.4, 150 mM NaCl, 2 mM MgCl2, 1 mM DTT, 10% glycerol, and 1% CA-630). Proteins were eluted with 40 µl of 2.5× sample buffer and analyzed by western blotting using anti-AS1 and anti-GFP (Santa Cruz Biotechnologies) antibodies.
|
10.1371/journal.pcbi.1005647 | MrTADFinder: A network modularity based approach to identify topologically associating domains in multiple resolutions | Genome-wide proximity ligation based assays such as Hi-C have revealed that eukaryotic genomes are organized into structural units called topologically associating domains (TADs). From a visual examination of the chromosomal contact map, however, it is clear that the organization of the domains is not simple or obvious. Instead, TADs exhibit various length scales and, in many cases, a nested arrangement. Here, by exploiting the resemblance between TADs in a chromosomal contact map and densely connected modules in a network, we formulate TAD identification as a network optimization problem and propose an algorithm, MrTADFinder, to identify TADs from intra-chromosomal contact maps. MrTADFinder is based on the network-science concept of modularity. A key component of it is deriving an appropriate background model for contacts in a random chain, by numerically solving a set of matrix equations. The background model preserves the observed coverage of each genomic bin as well as the distance dependence of the contact frequency for any pair of bins exhibited by the empirical map. Also, by introducing a tunable resolution parameter, MrTADFinder provides a self-consistent approach for identifying TADs at different length scales, hence the acronym "Mr" standing for Multiple Resolutions. We then apply MrTADFinder to various Hi-C datasets. The identified domain boundaries are marked by characteristic signatures in chromatin marks and transcription factors (TF) that are consistent with earlier work. Moreover, by calling TADs at different length scales, we observe that boundary signatures change with resolution, with different chromatin features having different characteristic length scales. Furthermore, we report an enrichment of HOT (high-occupancy target) regions near TAD boundaries and investigate the role of different TFs in determining boundaries at various resolutions. To further explore the interplay between TADs and epigenetic marks, as tumor mutational burden is known to be coupled to chromatin structure, we examine how somatic mutations are distributed across boundaries and find a clear stepwise pattern. Overall, MrTADFinder provides a novel computational framework to explore the multi-scale structures in Hi-C contact maps.
| The accommodation of the roughly 2m of DNA in the nuclei of mammalian cells results in an intricate structure, in which the topologically associating domains (TADs) formed by densely interacting genomic regions emerge as a fundamental structural unit. Identification of TADs is essential for understanding the role of 3D genome organization in gene regulation. By viewing the chromosomal contact map as a network, TADs correspond to the densely connected regions in the network. Motivated by this mapping, we propose a novel method, MrTADFinder, to identify TADs based on the concept of modularity in network science. Using MrTADFinder, we identify domains at various resolutions, and further explore the interplay between domains and other chromatin features like transcription factors binding and histone modifications at different resolutions. Overall, MrTADFinder provides a new computational framework to investigate the multiple length scales that are built inside the organization of the genome.
| The packing of a linear eukaryotic genome within a cell nucleus is dense and highly organized. Understanding the role of 3D genome in gene regulation is a major area of research [1][2][3][4]. Recently, genome-wide proximity ligation based assays such as Hi-C have provided insights into the complex structure by revealing various structural features regarding how a genome is organized [5][6][7]. Perhaps, one of the most important discoveries is the domain of self-interacting chromatin called topologically associating domain (TAD) [8][9]. Inside a TAD, genomic loci interact often; but between TADs, interactions are less frequent. Thus the TAD emerges as a fundamental structural unit of chromatin organization; it plays a significant role in mediating enhancer-promoter contacts and thus gene expression, and breaking or disruption of TADs can lead to diseases like cancers [10][11][12]. Therefore, a deeper understanding of TADs from Hi-C data presents an important computational problem.
Results of a typical Hi-C experiment are usually summarized by a so-called chromosomal contact map [5]. By binning the genome into equally sized bins, the contact map is essentially a matrix whose element (i,j) reflects the population-averaged co-location frequencies of genomic loci originated from bins i and j. In this representation, TADs are displayed as blocks along the diagonal of a contact map [8][9]. Despite the fact that TADs are rather eye-catching in a contact map, computational identification is still challenging because of experimental factors such as noise and inadequate coverage. Moreover, it is apparent from a visual examination of the contact map that TADs exhibit various length scales: there are TADs that appear to be overlapping, and within many TADs, there are rich sub-structures.
Mathematically speaking, it is very natural to transform a contact matrix to a weighted network in which nodes are the genomic loci (or bins) whereas the interaction between two loci is quantified by a weighted edge. In network science, a widely studied problem is the identification of network modules, also known as community detection problem [13]. A module refers to a set of nodes that are densely connected. In its simplest form, the community detection problem concerns with whether nodes of a given network can be divided into groups such that connections within groups are relatively dense while those between groups are sparse. Therefore, by viewing the chromatin interactions as a network, the highly spatially localized TADs immediately resemble densely connected modules. Motivated by the resemblance, we formulate the identification of TADs as a global optimization problem based on the observational contact map and a background model. As a network-based approach, our method goes beyond a direct adaptation of standard community detection algorithms. We introduce a novel background model that takes into account the effect of genomic distance, which is specific to the context of genome organization. The objective function is optimized using a heuristic algorithm that is efficient even if the size of the input contact map is large. Furthermore, by introducing a tuning parameter, our network approach can identify TADs at different resolutions. At a low resolution, larger TADs are found whereas, at a high resolution, smaller TADs are identified as the nucleome is viewed on a finer scale. In other words, the method can identify TADs at different length scales. We name our method MrTADFinder where the acronym Mr stands for multiple resolutions.
The identification of modules in a network is formulated as a global optimization problem on the so-called modularity function over possible divisions of the network. Consider an unweighted network represented by an adjacency matrix A. For a particular division (i.e. a mapping from the set of all nodes to a set of modules), the modularity is defined as the fraction of edges within modules minus the expected fraction of such edges in a randomized null model of the network. Mathematically, the modularity is equal to
12m∑i,j(Aij−kikj2m)δσiσj.
(1)
Here, the summation goes over all possible pairs of nodes, the value of the Kronecker data δσiσj equals one if nodes i and j have the same label σ and zero otherwise, meaning only pairs of nodes within the same module are summed. In particular, m is the number of edges in the network whereas the expression kikj/2m represents the expected number of edges between i and j in a so-called configuration model. The configuration model is a randomized null model in which the degrees of nodes ki are fixed to match those of the observed network, but edges are in other respects placed at random. High values of the modularity correspond to good partitions of a network into modules and similarly low values to bad partitions. Optimizing the modularity function leads us to the best partition over all possible partitions. More recently, a so-called resolution parameter γ has been incorporated in Eq (1) to adjust the size of the resultant modules [14].
Following the network formalism, given a Hi-C contact map represented by a weighted matrix W, we define a similar objective function Q as
Q=12N∑i,j(Wij−γEij)δσiσj.
(2)
Here, i,j index the equally binned genomic loci. N is the total number of pair-end reads. Eij is the expected number of contacts between locus i and locus j. γ is the resolution parameter that could be used to tune the size of resultant TADs. Very much similar to the network setting, the identification of TADs aims to partition the loci into domains such that Q is optimized. Nevertheless, it is important to emphasize two points. First, unlike the case in a network, the bins in a chromosome form a continuous chain and therefore genomic loci belonging to a TAD have to form a continuous segment. Second, simply because of the physical nature of chromosome, the expected number of contacts between locus i and locus j depends on their genomic distance. Two loci that are close together in a 1-dimensional sense are expected to have a higher contact frequency as compared to two loci that are far apart. This point suggests that the null model Eij in Eq (2) has to be modified.
Thus, given an intra-chromosomal contact map W, the expected null model E is defined as
Eij=κi*κj*f(|i−j|).
(3)
Here, f is the average number of contacts as a function of distance d = |i − j|. By considering all possible pairs of bins in W in terms of their distance apart and the contact frequency, we estimate f by local smoothing (see Methods). For intermediate values of d, f follows pretty well with a power-law function d−1 (see S1 Fig), which is a well-known observation first reported in [5].
As a null model, the resultant E matrix satisfies a set of constraints, namely
∑jEij=∑jWij=ci∀i,
∑ijEij=∑ijWij=2N.
(4)
The first equation means that the coverage ci, i.e. the total number of reads (one end of pair-end reads) mapped to bin i, defined in the observed map is the same as the coverage defined in the null model. The second equation is a direct consequence of the first equation, where N is the total number of pair-end reads mapped to the chromosome. As f has been estimated from the observed W, we can numerically solve all the unknowns κi* in the system of matrix equations (see Methods). Mathematically, κi* can be regarded as an effective coverage because of the correlation between κi* and the coverage ci is extremely high (r = 0.95, S2 Fig). In comparison with Eq (1), κi* is conceptually analogous to the degree ki. As shown in Fig 1, given a particular matrix W, the contact frequency of the resultant null model E are the highest in the diagonal and decrease gradually away from the diagonal. With W and E, for any given resolution parameter γ, we employ a modified Louvain algorithm to optimize Q (see Methods and Fig 1 for details). To ensure robustness, multiple runs of the modified Louvain algorithm are performed, and a boundary score is defined as the fraction of times a bin is called as a boundary. The final set of TADs is defined based on the set of consensus boundaries (Fig 1 and Methods). It is important to emphasize that the conventional Louvain algorithm used in network analysis [15] cannot be directly used because chromatin domains are continuous segments.
As a demonstration, we applied MrTADFinder to analyze Hi-C data of hES cell from [8]. Fig 2A shows a particular snapshot of the contact map (for chromosome 10) and its alignment with the identified TADs. In general, the TADs displayed agree well with the apparent block structures in the contact map. Of particular interest is the choice of γ that capture various length scales in domain organization. As shown in Fig 2A, when γ increases, a large TAD breaks into a few small TADs. On the other hand, a few large TADs merge together to form an even larger TAD as the value of γ is lowered. Statistically speaking, γ quantifies to what extent do we accept the enrichment of empirical contact frequency over the expectation. As γ increases, only matrix elements close to the diagonal contribute positively to the objective function. Therefore, in general, the size of TADs decreases (see Fig 2B) and the number of TADs increases (see Fig 2C). For example, when γ = 1.0, there are about 1000 TADs in hES cells with a median size of 3Mb. When γ = 2.25, the number of TADs increases to 2600 and the median size is roughly 1Mb.
We then further compared the TADs identified at different resolutions by MrTADFinder with TADs identified by a previous method. As quantified by the normalized mutual information (see Methods for details), TADs identified by MrTADFinder best match with TADs identified in [8] when the resolution parameter is 2.9. In general, unless the resolution is sufficiently small (γ < 1.5), the two methods are quite consistent (see Fig 2C). Nevertheless, the introduction of the resolution parameter γ opens an extra dimension in domain identification in a sense the algorithm used in [8] focuses on a particular resolution instead.
The interplay between 3D genome organization and various chromatin features has widely been investigated since some of the first Hi-C experiments were reported [5][8][9]. Nevertheless, there is no clear-cut pattern emerges by aligning a variety of chromatin features with TADs (S3 Fig), even though the occurrence of sharp peaks at the boundaries is quite apparent. By identifying TADs and their boundaries using MrTADFinder, we found the boundary signatures that are consistent with the observations previously reported [8], for instance, the enrichment of active promoter mark H3K4me3 or active enhancer mark H3K27ac, as well as the depletion of transcriptional repression mark like H3K9me3 (Fig 3A and S4 Fig). To better understand the relationship between domains organization and different chromatin features, we further examined the chromatin features near different sets of boundaries that were identified in different resolutions. We found that in general, the enrichment of peak density at boundary decreases as resolution increases. This is because the number of TADs increases as the resolution increases, various chromatin features appear in the boundaries of low-resolution TADs do not appear in high-resolution TADs (Fig 3A). More specifically, the enrichment of histone marks like H3K36me3 and H3K4me3 exhibits a monotonic drop whereas certain marks exhibit characteristic resolutions. For instance, the enrichment of mark H3K27me3 remains high up to a resolution of γ = 2.5 (Fig 3B). The observation suggests that the mark H3K27me3 in general marks the boundary of TADs up to a particular resolution (length scale).
Beside epigenetic signatures, we examined the distribution of protein-coding genes along chromosomes in relation to TAD boundaries formation. Though the starting positions of genes tend to be enriched near TAD boundaries, the enrichment is much stronger for housekeeping genes as compared to tissue-specific genes (Fig 4A). As housekeeping genes are essentially active, the pattern resembles the active promoter mark H3K4me3 shown in Fig 3B. The discrepancy between housekeeping genes and tissue-specific genes was firstly reported in Ref. [8]. Nevertheless, by extending the idea to multiple resolutions, we found that the distribution of housekeeping genes follows a different length scale compared to tissue-specific genes. As shown in Fig 4B, housekeeping genes in general marks the boundary of TADs up to the resolution γ = 1.5.
Apart from histone modifications, it is well known that certain transcription factor binding sites are enriched near the boundary regions of TADs [8]. Instead of looking at individual factors, we further explored the location of the so-called HOT regions and XOT regions on TADs. High-occupancy target (HOT) regions and extreme-occupancy target (XOT) regions are genomic regions that are bound by an extensive amount of transcription factors [16]. As expected, we found a strong enrichment of HOT regions and an even stronger enrichment of XOT regions near TAD boundaries in hES cells (Fig 5A). The observation is, in general, true for all tested resolutions. The observation agrees with the idea that HOT regions are very accessible regions in open chromatin. Nevertheless, it is still widely unknown if transcription factors bind to HOT regions simply because of thermodynamics, or the binding will result in important biological consequences.
Motivated by the observation that many factors tend to bind to the boundary regions, we further examine which factors are responsible for establishing the domain border, and more interestingly for borders in different resolutions. There are a few proteins which are widely known to be important in border establishment [17]; nevertheless, it is worthwhile to perform a systematic analysis. To do so, we formulated a classification problem which aims to distinguish, for each resolution, a set of boundaries identified by MrTADFinder (positive set) from a set of random boundaries obtained by swapping the TADs along the chromosomes (negative set). Using a logistic regression model recently proposed by [18], we integrated the binding signals of 60 transcription factors at a genomic locus to predict if it is TAD boundary (see Fig 5B and Methods for details). Generally speaking, with 10-fold cross validation, the model is quite successful in low resolutions (AUC = 0.81, S5 Fig). The result is consistent with an early work based on histone modifications [19]. Being consistent with the trend that chromatin features are less enriched at the boundaries of high resolution TADs, the predicting power of the model decreases as the resolution increases. The regression model further quantifies explicitly the influence of each of the transcription factors. In general, factors that are responsible for border formation are quite consistent across different resolutions (Fig 5B). For instance, we found that the well-known insulator CTCF, and Rad21 that is a part of cohesin, are direct key components of border establishment. In addition, the chromatin remodeler Chd7, which is often found at enhancers [20], is predicted to be a key component. On the other hand, factors like MYC have a consistently negative effect. Nevertheless, the relative importance of factors does change with resolutions. For instance, Rad21 has a higher predictive power in classifying high-resolution domains in compared with classifying low-resolution domains.
The contact maps of more deeply sequenced Hi-C experiments have exhibited a pattern that a large fraction of TADs has “peaks” in their corner [21], meaning the contact frequency between the endpoints of such domains is higher than those of their surrounding neighborhood. The configuration suggests that the boundaries of such domains form a chromatin loop. We investigated if a similar conclusion could be drawn from the TADs called by MrTADFinder using a set of significant long-range promoter contacts identified by capture Hi-C [22]. Based on the Hi-C data of GM12878 in [21], we found that there are indeed potential promoter-enhancer linkages connecting the endpoints of domains. Moreover, by increasing the resolution parameters, the boundaries of the smaller TADs further capture the potential promoter-enhancer linkages in shorter length scales (Fig 6). It is worthwhile to point out that the linkages connecting the endpoints of domains form a small fraction as compared to the total number of significant interactions identified by capture Hi-C. Therefore, identifying the domain borders is not a direct method to predict potential enhancer-gene linkages. On the other hand, though the increase in the number of boundaries can capture a higher number of potential interactions, the same analysis for an ensemble of randomly reshuffled TADs shows the observation in TADs called by MrTADFinder is significant (Fig 6). In other words, TADs in a higher resolution are potential subTADs that mediate long-range interactions in a finer length scale [23].
We have examined the interplay between domains organization and chromatin features. Recently, it has been reported that epigenomic features shape the mutational landscape of cancer [24]. Motivated by this linkage, we further investigated the occurrence of somatic mutations near the boundaries. More specifically, we mapped the somatic mutations obtained from breast cancer samples to the TAD boundaries we identified in MCF7 cells (see Methods). In a given resolution, there are 85 boundary regions identified on chromosome 10. The regions can be clustered into 3 groups based on the positional distribution of somatic mutations. As in shown in Fig 7, two of the clusters exhibit a step-function behavior (blue and red) in which the abrupt transition essentially happens at the boundary. For boundary regions in the remaining cluster, the mutational burden exhibits no difference across the TAD boundaries. Because of the close relationship between TADs and replication-timing domains [25], the observation resonates with a well-known observation that genomic regions with a high mutational burden are replicated at a later stage during DNA-replication [26]. As shown in the inset, using Repli-seq data in S1 phase, the upstream regions of the boundaries found in the blue cluster have a high mutation rate but a low Repli-seq signal, meaning they are indeed replicated at a later stage during replication. On the contrary, the upstream regions of the boundaries found in the red cluster are replicated at an early stage and therefore exhibit a low mutation rate.
Motivated by the relationship between TADs and DNA replication, we overlaid TADs in different resolutions with data from Repli-seq experiment (S6 Fig). We observed that TADs identified in different resolutions match with the Repli-seq data in different stages of a cell cycle. For instance, while a TAD identified in a low resolution does not replicate at an early phase, say S1, its sub-structures identified in a higher resolution correspond to two separate peaks at later stages, say S2 and S3 (S7 Fig). Nevertheless, it is worthwhile to point out that mapping Hi-C reads from cancer cell lines like MCF7 to the reference genome is not perfect because quite some reads may come from copy number variations. Computational approaches have recently been developed to perform specific normalization [27] as well as to infer those large scale genomic alterations from Hi-C data [28].
There are quite a few existing methods on identifying TADs using Hi-C data. Dixon et al. identified TADs based on the so-called directionality index using Hi-C data in hES cell and found an enrichment of CTCF binding sites at the boundary regions [8]. Since then the enrichment of chromatin features has been used as a benchmark for various TAD calling algorithms [29][30][31]. As a comparison, we performed the same analysis using TADs based on MrTADFinder. As shown in Fig 8, both methods exhibit a similar pattern. In fact, as reported in [29][30][31], the enrichment pattern of CTCF binding peaks is qualitatively the same for all the proposed methods. By repeating the analysis in different resolutions, we observed that the level of enrichment depends on the resolution (Fig 8, S7 Fig). At a low resolution, i.e. for larger TADs, the enrichment signal is stronger, and the signal tends to extend over a longer distance from the boundary. At a higher resolution, the signal is weaker and confined to near the boundary. In general, Fig 8 suggests that boundaries identified in lower resolutions are more likely to be bound by CTCFs. From a biological standpoint, as a boundary identified in a lower resolution separates two large domains, the results may bring insights on how to mediate chromatin loops at different length scales via an important architectural protein [32][33]. As the level of CTCF enrichment might be the consequence of different chromatin length scales, it might not be fair to use it directly for benchmarking the performance of different algorithms.
Because of the stochastic nature of the modified Louvain algorithm, we explored the robustness of MrTADFinder. In the current setting based on multiple runs of the modified Louvain procedure, we found the results of two independent callings highly robust. In fact, the normalized mutual information is 0.99 (see S8 Fig). We further investigated the reproducibility of MrTADFinder in two aspects. First, we studied the agreement of TADs called in biological replicates. Using Hi-C data released by the ENCODE consortium, we found that TADs called in a pair of biological replicates agree reasonably well, with normalized mutual information about 0.85 (see S9 Fig and Methods). Secondly, we explored the effects of sequencing depth to our algorithm. Specifically, we applied MrTADFinder to identify TADs from a deeply sequenced Hi-C data of GM12878 [21]. We then reduced the number of reads included and called TADs again. We found that the TADs identified using a subset of reads are slightly different from the original, and in general, the discrepancy increases as fewer reads were used (S10 Fig and Methods). Despite a certain level of discrepancy, nevertheless, the resultant TADs agree well. For instance, in the extreme case, by comparing using contact maps constructed from 2.4 billion reads and 480 million reads respectively, the mean normalized mutual information of various pairs of chromosomes is about 0.88. If we compare the TADs called from 2.4 billion reads to the TADs called from 1 billion reads, the normalized mutual information is higher than 0.95.
MrTADFinder is implemented in Julia. Julia programmers can import MrTADFinder as a library for calling various functions. It can also be run in command line if Julia and the required packages are installed. The performance of MrTADFinder, in general, depends on the size of the input contact map. We have tested the performance using the contact maps of GM12878 cell generated by the Aiden lab [21]. The performance is reasonable. For instance, for chromosome 10, in a bin-size of 25kb (i.e. a contact map 5400 by 5400), the time required to arrive at all TADs with 10 runs of Louvain algorithm is about 20 minutes on a laptop with 2.8GHz Intel Core i7 and 16Gb of RAM. The time required is only 6 minutes if the bin size is 50kb. We have made the source code available on GitHub (see software availability).
Despite the similarity between Eqs (1) and (2), network modules are rather arbitrary collections of nodes, but domains are continuous segments along the chromosome. In fact, the total number of possible partitions for a chromosome is much smaller than the total number of ways to divide a network into modules. As a result, while the optimization of Eq (1) is an NP-hard problem, the optimization of (2) can be quite efficiently solved using a dynamic programming inspired method (see Methods and S11 Fig). It is instructive to explore this avenue because quite some algorithms for identifying TADs are based on a similar approach but with different objective functions [29][30][31]. Moreover, by enumerating all possible ways to decompose a chromosome into TADs, one could write down the partition function and define a probability of occurrence for each of the possible partition in a statistical mechanics’ manner.
The time complexity of this algorithm is in order of O(n3), where n is the size of the contact map. Given the time complexity, finding the optimal partition using a bin size of 40kb is quite impractical. For instance, the calculation takes about an hour for chromosome 21, as compared to seconds by using the heuristic. Therefore, though the connection between identifying TADs and problems like finding RNA secondary structure is of theoretical interest, MrTADFinder is developed based on the modified Louvain algorithm. Nevertheless, we have implemented the approach based on recurrence relation and performed a comparison with the heuristic. Using a contact map of hES cell (chromosome 1) with a bin size of 500kb, we found the sub-optimal partitions based on our modified Louvain algorithm are very close to the optimal partition. The normalized mutual information between optimal and sub-optimal values is 0.977±0.007.
In this paper, we have introduced an algorithm to identify TADs from Hi-C data and performed several analyses to show the biological significance of the TADs identified. In particular, by introducing a single continuous parameter γ, we can further examine domains organization and its interplay with a variety of chromatin features in multiple resolutions. It is important to emphasize that the idea of resolution we introduced in MrTADFinder is different from some other usages of the same term in Hi-C analysis. From an experimental standpoint, the resolution of a Hi-C experiment refers to the average fragment size as digested by restriction enzymes (~4kb to ~1kb) [5][21] or more recently by micrococcal nuclease (~150bp) [34]. Regarding the construction of contact maps, the term resolution has been used to refer to the bin size, where the proper choice usually depends on the number of reads in the stage of data processing. Both usages are primarily technical. What we mean by resolution, however, refers to the multiple length scales built inside the organization of the genome. It is well known that there are structures in different length scales such as compartment, domains, and sub-domains [35], and chromatin features like histone marks exhibit multiple length scales [36]. The concept of resolution introduced here points to the integration of these structures and enables one to explore the rich structures hidden in contact maps. From a practical point of view, γ = 1 seems to be the natural starting point. One could increase or decrease the value of γ in order to explore the intrinsic structure. Nevertheless, because of the different contact maps might have various differences like the read coverage, one should be cautious to directly compare the resolution parameters between different contact maps.
A novel contribution of this work is the derivation of an expected model for any intra-chromosomal contact map by solving a system of matrix equations. The null model preserves the coverage of each genomic bin as well as the distance dependence of contact frequencies in the observed map. As such features of contact maps are involved in most computational analysis of Hi-C data, apart from the identification of TADs, the expected model can be used for applications like finding compartments [5] and identifying potential enhancer-target linkages [37]. Mathematically, the expected matrix is solved by an iterative procedure. The procedure can be regarded as a generalization of a class of matrix balancing methods used for normalizing Hi-C matrices [38], as the later is merely a different set of matrix equations. However, it is important to emphasize that the so-called ICE algorithm aims to remove bias in the contact map, whereas our method aims to generate a background model. While MrTADFinder focuses on intra-chromosomal interactions, recent studies employ various clustering methods to identify inter-chromosomal clusters using Hi-C contact frequency [39][40]. It is worthwhile to point out that similar expected models used in this study can also be derived for inter-chromosomal interactions to better separate signal and noise.
Several methods have been developed for identifying TADs from Hi-C data [41]. One of the earliest methods is based on the so-called directionality index, a 1D statistic measuring whether the contacts have an upstream or downstream bias [8], and later the bias is exploited by the so-called arrowhead algorithm [21]. Later algorithms exploit the block diagonal nature of TADs in a contact map [29] [30][42]. Though some of these algorithms do take the distance dependence into the background, but they do not take into account both the genomic distance and the effects of coverage in a compact mathematical formalism. The algorithm TADtree [30], and more recent efforts, namely Matryoshka [31] and metaTAD [43] aim to investigate the hierarchical organization of TADs based on a tree structure. Indeed, merging smaller TADs at the lower level of the hierarchy results at larger TADs similar to the TADs obtained by MrTADFinder at a low resolution. Nevertheless, MrTADFinder does not impose a hierarchical organization. The probabilistic nature of Louvain algorithm enables the definition of TAD boundaries in a probabilistic fashion, and therefore a possibility to define overlapping TADs. To a certain extent, the idea of continuous resolution used in MrTADFinder is distinct in comparison with algorithms based on a bottom-up approach, but similar in spirit to Ref. [29].
MrTADFinder is motivated by the community detection problem in network studies. Although a network perspective of chromosomal interactions has previously been proposed [44][45], a lot of widely studied concepts in networks have rarely been explored in the context of chromosomal organization. A network representation is arguably more flexible than a simple matrix representation, for instance, transcription factors binding and histone modifications can be easily incorporated into the network, forming a decorated network. Moreover, one could extend the framework by concatenating multiple Hi-C contact maps to form a multi-layer network. The same idea has been used for cross-species transcriptomic analysis [46]. By facilitating the application of a variety of graph-theoretical tools, we believe that network algorithms will be useful for future studies on the spatial organization of the genome.
The Hi-C data of human ES cells and IMR90 cells were reported in Ref. [8]. Raw reads were processed using Hi-C Pro [47], arriving at contact matrices in various bin sizes. In all analysis, the whole-genome contact map was iteratively corrected for uniform coverage [38]. Intra-chromosomal contact maps were then extracted from the whole-genome contact map of bin size 40kb for downstream analysis. Hi-C data and contact maps in MCF7 cells were reported in Ref. [48]. The whole-genome contact map provided was binned with 40kb bin size and was normalized by the ICE algorithm. Data in GM12878 were reported in [21]. The bin size of the contact maps used for the analysis related to the number of promoter-enhancer linkages was 25kb. The analysis on the effect of sequencing depth was performed by selectively combing the raw contact maps constructed from individual Hi-C libraries of the same replicates [21]. The bin size was chosen to be 50kb. The ENCODE Hi-C data were released by the ENCODE consortium. Altogether 8 cell lines with a relatively higher coverage were used in the reproducibility analysis including T47D, A549, Caki2, G401, NCI-H460, Panc1, RPMI-7951 and SK-MEL-5. For each cell line, two replicates were separately used. The ENCODE Hi-C data were processed by the tool cworld (https://github.com/dekkerlab/cworld-dekker). Capture Hi-C data were reported in Ref. [22]. Only 1618000 significant interactions linking promoters and non-promoter regions were included in the analysis of Fig 8. Visualization of contact maps were all generated by the tool HiCPlotter [49].
All chromatin data, including histone modifications, transcription factors binding, expression, replication timing, were downloaded from the ENCODE data portal.
The average number of contacts as a function of genomic distance can be estimated by considering all elements in matrix W. A local smoothing approach similar to the method used in [50] was employed. The window size equals to 1% of the data.
Eqs (3) and (4) can be rewritten in the form
∑jκi*κj*f(|i−j|)=ci∀i.
(5)
The system of non-linear equation is similar to the matrix balance approach used in [38]. As the aim of [38] is to remove bias, the coverage ci is the same for all bin i and f is replaced by the original empirical map. Nevertheless, the unknowns κi* can be solved by a similar iterative procedure as proposed in [38].
To optimize the objective function Q, we employ a modified version of Louvain algorithm [15], which is widely used in identifying modules in networks (see Fig 1). In a nutshell, the algorithm consists of two steps. The algorithm starts as every bin has its own label, and the label will end up as an identifier for the module it belongs. In the first step, for each bin, we update its label by either choosing the label of one of its two neighboring bins or by remaining unchanged based on whether or not the value of Q will be increased. There will be multiple rounds of updates in this step. For each round of update, we go through all the bins once, but the order is random. The updating procedure will be repeated for multiple rounds until no more update is possible. We will then perform the second step such that the bins with the same labels will be locked together, in a sense their labels will only be updated in a synchronized fashion. It is worthwhile to mention that the updating procedure in the first step makes sure bins with the same labels form a continuous segment. Once the bins are locked to form super-bins, the first step will be performed again but in the level of super-bins. The two steps will be repeated iteratively until no increase of modularity is possible.
The output of the modified Louvain algorithm is essentially a particular partition of the entire chromosome. As the result of the algorithm, in general, depends on the order of updates, multiple runs are performed to probe the fuzziness of the assignment. As the chromosome is binned into n equally sized bins, we examine, say after 10 trials, how likely the border between bin i and bin i + 1 is indeed a domain boundary, i.e. bin i and bin i + 1 are called to belong to two different TADs by the modified Louvain algorithm. We then naturally define a boundary score for each of the n+1 borders as the fraction of trials in which a border is called as a boundary. To define a set of consensus boundaries, we choose a cut-off of 0.9. In other words, the border between two adjacent bins is defined as a confident boundary only if they are called to belong to two different domains in at least 9 out of 10 trials. The final output of MrTADFinder is a set of consensus TADs defined as regions between the consensus domains.
The boundary score assigned to each border is not merely an immediate but serves as a proxy of the degree of insulation. A border with a high boundary score is more effective in forbidding the contacts between its left and right regions.
Given two sets of TADs, say in different cell lines, or called by different algorithms, we employ the so-called normalized mutual information to quantify the consistency. Suppose X and Y are two random variables whose values xi and yi represent the corresponding domain labels of bin i. The normalized mutual information MInorm is defined as
MInorm=2I(X;Y)H(X)+H(Y),
(6)
here H(X),H(Y) are the entropy of X and Y, and I(X;Y) is the mutual information quantifying to what extent the domain labels in X predict the labels in Y. A normalized form of mutual information is used here to make sure the value lies between 0 and 1 for comparison. To have a fair comparison, bins that are not assigned to any TADs in both sets of partitions are not counted. If two sets of partitions are identical, the value of normalized mutual information is 1.
Given the location of binding peaks of a transcription factor or a histone mark, the peak density near TAD boundaries was estimated by considering for all boundaries the region from upstream 600kb to downstream 600kb. The regions were aligned, and the number of peaks was summed accordingly. To calculate the enrichment, the number of peaks was normalized by the expected number of peaks in a particular region under a null model that peaks are randomly distributed in the genome.
The influence of individual transcription factors on the formation of domain borders was formulated as a classification problem. For a particular resolution, the set of boundaries called by MrTADFinder was used as a positive set whereas a set of random boundaries obtained by swapping the TADs along the genome was chosen as the negative set. The signal values of 60 transcription factors are used as features for classification. The combined effect of all features was modeled the logistic function
f(X,(β0,β))=11+exp(−β0+βX),
(7)
here X represents all features; β is a vector determining the coefficients of influence for all features and βo is a bias parameter. Given a training set, a likelihood function was defined. An optimal β was inferred by optimizing the likelihood function using gradient descent with L1-regularization. The inferred logistic function was used to predict the test set. To have a more accurate estimate, 10-fold cross-validation was performed, and the error bars were estimated by multiple negative training sets.
The set of somatic mutations were downloaded from the data portal of the International Cancer Genome Consortium (ICGC). The mutations were called the breast cancer samples of 676 donors. The samples were sequenced in a whole-genome level. Breast cancer samples were used in this analysis to match the Hi-C data of MCF7 cell.
The idea is to extensively enumerate all the possible partitions of the chromosome. In a nutshell, a binned chromosome can be considered as a sequence (1,2,⋯,n − 1,n). Rather than partitioning the whole sequence at a first place, we look for the optimal partition for all the possible sub-sequences starting from sub-sequences with length 1. Let us denote the optimal value of modularity Q for a sequence a1a2 … al−1al as optQ(a1a2 … al−1al). The value is the maximum of the following l possibilities:
optO(a1)+optO(a2…al−1al),optO(a1a2)+optO(a3…al−1al),⋮optO(a1a2a3…al−1)+optO(al),
∑ijQij.
(8)
Suppose the maximum is the sum optO(a1a2 … ar) + optO(ar+1 … al−1al), where 1 ≤ r < l. The sum corresponds to the case that the optimal partition of a1a2 ⋯ al is a combination of the optimal partitions of a1a2 ⋯ ar and ar+1 … al−1al (see S11 Fig). It is not necessary that a1a2 ⋯ ar forms a single domain. The key is that the expression optQ(a1a2 … al−1al) can be found recursively because all possibilities depend on the optimal values of sub-sequences shorter than l. The last summation in (4) sums Q over all positions from a1 to al, meaning the l bins belong to the same domain. Once the value of optQ(a1a2 … an−1an) is found, we can trace back the actual partition for the whole chromosome. As shown in the source code, it takes three loops to enumerate all possible partitions. The procedure is analogous to the Nussinov algorithm in finding the optimal secondary structure of RNA [51].
The source code can be downloaded at https://github.com/gersteinlab/MrTADFinder.
|
10.1371/journal.pntd.0001760 | Refining the Global Spatial Limits of Dengue Virus Transmission by Evidence-Based Consensus | Dengue is a growing problem both in its geographical spread and in its intensity, and yet current global distribution remains highly uncertain. Challenges in diagnosis and diagnostic methods as well as highly variable national health systems mean no single data source can reliably estimate the distribution of this disease. As such, there is a lack of agreement on national dengue status among international health organisations. Here we bring together all available information on dengue occurrence using a novel approach to produce an evidence consensus map of the disease range that highlights nations with an uncertain dengue status.
A baseline methodology was used to assess a range of evidence for each country. In regions where dengue status was uncertain, additional evidence types were included to either clarify dengue status or confirm that it is unknown at this time. An algorithm was developed that assesses evidence quality and consistency, giving each country an evidence consensus score. Using this approach, we were able to generate a contemporary global map of national-level dengue status that assigns a relative measure of certainty and identifies gaps in the available evidence.
The map produced here provides a list of 128 countries for which there is good evidence of dengue occurrence, including 36 countries that have previously been classified as dengue-free by the World Health Organization and/or the US Centers for Disease Control. It also identifies disease surveillance needs, which we list in full. The disease extents and limits determined here using evidence consensus, marks the beginning of a five-year study to advance the mapping of dengue virus transmission and disease risk. Completion of this first step has allowed us to produce a preliminary estimate of population at risk with an upper bound of 3.97 billion people. This figure will be refined in future work.
| Previous attempts to map the current global distribution of dengue virus transmission have produced variable results, particularly in Africa, reflecting the lack of accuracy in both diagnostic and locational information of reported dengue cases. In this study, instead of excluding these less informed points we included them with appropriate uncertainty alongside other diverse evidence forms. After assembling a comprehensive database of different evidence types, a weighted scoring system calculated “evidence consensus” for each country a continuous measure of the certainty of dengue presence or absence when considering the full aggregate of evidence. The resulting map and analysis helped highlight important evidence gaps that underlie uncertainties in the current distribution of dengue. We also show the importance of local knowledge through incorporating questionnairebased responses that can help add clarity in uncertain regions. This analysis showed that presence/absence maps do not sufficiently highlight the uncertainties in the evidence base used to construct them. Mapping by evidence consensus not only encourages greater data inclusion, but it also better illustrates the current global distribution of dengue. Consensus mapping is thus ideal for a range of neglected tropical diseases where the evidence base is incomplete or less diagnostically reliable.
| Despite increased interest in dengue in recent years, the global distribution of dengue remains highly uncertain. Estimates for the population at risk range from 30% [1] to 54.7% [2] of the world's population (2.05–3.74 billion) while the Centers for Disease Control (CDC) and the World Health Organization (WHO) currently disagree on dengue presence in 34 countries across five continents (Table S1). Clinical features of dengue virus infection include fever, rash and joint pain [3], which ensure the disease's misdiagnosis and mis-reporting among many other febrile illnesses. The diagnostic methods available also have limitations and a full complement of tests is not feasible in many healthcare settings. There is consensus, however, that dengue is a growing problem both geographically and in its intensity [4], [5], [6].
There is an urgent need to compile more extensive occurrence records of dengue virus transmission and assess them for contemporariness and accuracy. Evidence on dengue transmission comes in a wide variety of forms, with varying levels of spatial coverage and reliability. A global audit of dengue distribution therefore requires a transparent methodology to compile these disparate data types and synthesise an output map summarising the current consensus for each country. Such a methodology for compiling and assessing evidence must be robust, repeatable, able to evaluate a large variety of evidence types and incorporate expert opinion. An ideal output metric is a summary statistic (hereafter referred to as evidence consensus) that quantifies certainty on dengue virus transmission presence or absence given the accuracy and contemporariness of the evidence available. An evidence-based map of the current distribution of dengue virus transmission will have direct implications for design and implementation of dengue surveillance and, by showing gaps in contemporary knowledge, provide an advocacy platform for improved data.
Existing approaches to mapping the global limits of vector-borne diseases have used estimates of biological suitability of local environments, which have proved informative in the cases of some pathogens, such as Plasmodium falciparum [7], [8] and P. vivax [9]. Several approaches have been used to map biological suitability for dengue using non-dengue-specific variables such as temperature, rainfall and satellite-derived environmental variables [1], [10], [11]. Although successive attempts have each increased predictive capacity and resolution, this approach produces variable results in Africa due to a scarcity of confirmed occurrence points across extensive geographic areas. An alternative approach has been to map evidence of dengue occurrence making no assumptions about biological suitability, as in Van Kleef et al., who reviewed published literature to contrast historic, current and future limits of dengue [5]. To date dengue mapping has focussed on future scenarios, yet understanding of the current distribution of dengue virus transmission is far from complete and needs to be better evaluated before we can make predictions about forthcoming patterns and trends. In this study we combine evidence from large occurrence-point style databases used in biological suitability mapping approaches with a wider systematic review of various sources of evidence to create a more comprehensive dengue database. Using this database we then use the novel method of defining evidence consensus to evaluate the current level of certainty on dengue virus transmission presence or absence at national (and some sub-national) levels using a weighted evidence scoring system. Finally, we present these results as a series of global maps that explicitly identify surveillance gaps.
This study is the initial part of a five year project to collect, analyse and publicise global dengue virus transmission data. While the map presented here is the most extensive display of current dengue evidence available, we hope that continual data acquisition will result in more evidence from uncertain areas, increasing the resolution at which we can map evidence consensus in future advances.
Evidence for indigenous dengue virus transmission was obtained from four evidence categories: health organisations, peer-reviewed evidence, case data and supplementary evidence (Figure 1). The first three categories were used for all countries. For countries where some of these categories were not available and/or did not provide good consensus, the fourth category of supplementary evidence was used. Evidence was initially collected at a country level (Admin0), but resolution was improved to a state/province level (Admin1) or district level (Admin2) at the fringes of the distribution of detectable virus transmission when sufficient data were available.
Country dengue status as defined by health organisations was determined by consulting the WHO [12] and CDC [13] dengue distribution maps as well as the Global Infectious Diseases and Epidemiology Online Network (GIDEON) database [14]. GIDEON provides a collection of literature and case reports for a range of tropical and infectious diseases in 224 countries. Dengue status by country was recorded as present or absent.
The peer-reviewed evidence category contained evidence of dengue occurrence as determined by peer-reviewed sources where details of diagnostic techniques were given. Peer-reviewed journal (Google Scholar, PubMed, ISI Web of Science) and disease surveillance network (ProMED archives, Eurosurveillance archives) searches were conducted with search terms “country” or “Admin1/2” and “dengue”. Sources were included for the period 1960–2012 and only if cases were confirmed as resulting from indigenous (i.e. not imported) transmission. The specialist regional journal collections African Journals Online (http://www.ajol.info/) and China National Knowledge Infrastructure (http://en.cnki.com.cn/) were also searched. Extra publications were found by searching using the location term in Genbank nucleotide records for dengue viruses isolated from human hosts. The search of peer-reviewed sources of evidence resulted in a total of 285 articles being selected for 123 countries where positive dengue occurrence records were identified. This included evidence from returning travellers who were diagnosed upon return to their often non-endemic home countries as opposed to the transmission setting. For these cases, evidence was attributed to the place to which they had travelled. The added value of returning traveller reports is that the travellers are often more immunologically naïve to dengue infections, and also that diagnosis is often pursued more rigorously. Therefore, the sensitivity of detecting an infection is increased. The results of our search were then cross-referenced against a dengue occurrence-point database compiled internally, in a separate exercise. Unlike our country-specific searches, this database of 2836 articles results from searches simply for “dengue”, which were then geo-referenced using the article text. Full details are available in Protocol S1 and the geographic location of the occurrence points are displayed in Figures S1, S2, S3, S4, S5, S6. This cross-referencing resulted in the inclusion of an additional 16 articles in the current analysis and also provided increased justification for our choice of countries to evaluate at Admin1 level.
The case data category contained evidence of dengue outbreaks (minimum 50 infections) where evidence contained less diagnostic detail, but was more informative about the magnitude of dengue transmission occurring. Case data from the most recent outbreak were obtained from the Program for Monitoring Emerging Diseases (ProMED) archive search, WHO DengueNet data query [15] and from GIDEON which holds a detailed record of government-reported case numbers. This resulted in 100 countries with useful dengue case data.
In many resource-poor countries, both surveillance and researcher-generated reports are rare. Therefore, in countries where other evidence categories were sparse, we looked for supplemental evidence that suggested possible dengue virus presence. Supplemental evidence types included: presence of an established mosquito vector population of public health significance (Aedes aegypti, Ae. albopictus or Ae. polynesiensis) as documented by peer-reviewed literature, confirmed presence of multiple other rarely diagnosed arboviral diseases as documented by peer-reviewed literature, news reports of dengue epidemics found using GoogleNews archives (http://news.google.co.uk/archivesearch) and travel advisories from the National Travel Health Network and Centre (http://www.nathnac.org/ds/map_world.aspx) issued at a country-level. We included evidence of multiple other rarely diagnosed arboviral diseases, as these are informative about the ability of a country to detect any possible dengue infection. If other arboviral diseases are poorly reported, but documented by peer-reviewed literature as present, then it is possible that dengue is also underreported. In addition to this, we cross-referenced our dataset with the HealthMap database (www.healthmap.org/dengue/). This website-based application automatically geo-positions cases from websites with news reports and outbreak alerts related to dengue and contains data from a wide variety of sources dating back to 2007 [16], [17]. This extensive database contributed important evidence especially at smaller spatial scales and in areas where translated articles are not so easily obtained. Supplementary evidence was used in evaluating dengue consensus in 45 countries.
While the categories are clearly defined here and in Figure 1, some overlap of evidence sources did occur, depending on the information content of each source. This meant evidence sources such as ProMED reports could be included twice, in both the peer-reviewed evidence and case data categories, if they contained information about diagnostic tests used for confirmation as well as overall outbreak case numbers. In this section we outline the main sources used for each category, but it should be noted that if evidence from a particular source fitted the criteria for a different evidence category, it was not excluded, but rather included in that category.
In order to quantify evidence consensus, a weighted scoring system was developed that attributed positive values to evidence of presence and negative values to evidence of a lack of presence. The aim here was to use an optimal subset of evidence to accurately assess dengue status within a given area. By scoring the evidence categories mentioned above individually and then combining their respective scores, we were able to calculate “evidence consensus,” a measure of how strongly the combined evidence collection supports a dengue-present or dengue-absent status (Figure 2). We defined a country as having “complete consensus” on dengue presence when the evidence base was comprised of contemporary forms of most or all of the following evidence types: 1) unanimous health organisations agreement, 2) a seroprevalence survey, 3) Polymerase Chain Reaction (PCR) typing of dengue virus or dengue viral RNA, 4) a foreign visitor to the area with a confirmed dengue infection upon returning to their home country, and 5) records of an epidemic of greater than 50 infections. Such a country has a consensus score of between 80% and 100%. A country with a complete consensus on dengue virus absence is characterised by all health organisations agreeing on dengue absence and high healthcare expenditure (as an approximate proxy for surveillance capability), therefore accounting for both the observed absence of dengue and the minimised possibility of any undetected dengue infections. Such a country scores between −80% and −100% on our scale. A country with no consensus on dengue virus status is characterised by conflicting evidence from different categories and scores close to 0%. Each evidence category was scored independently and category weights applied to reflect the level of detail each category provides: health organisation status (maximum score 6), peer-reviewed evidence (maximum 9), case data (maximum 9) and supplementary evidence (maximum 6). To support the choice of assigned category weights we performed a sensitivity analysis in which two alternative evidence weighting scenarios were applied to the same sources of data: 1) neutral (all categories hold the same weight) and 2) reversed (health organisation status and supplementary evidence hold weight 9, peer-reviewed evidence and case data hold weight 6). We then checked for any major deviations in overall country score resulting from such alternative scenarios.
In countries where evidence consensus was at best moderate, we attempted to increase consensus through targeted questionnaires. The questionnaire asked about endogenous surveillance and data collection. If available, diagnostic method(s) and summary results were requested. Any returned data or reports were then entered into their relevant evidence categories and scored in combination with existing evidence. Questionnaires were distributed to healthcare officials in the country of interest as well as selected offices of the Institut Pasteur. Questionnaire responses and expert comments are part of an on-going process that will lead to future modifications of this map.
To map public awareness of dengue worldwide, we searched the ministry of health websites of each of the 128 countries identified as dengue-present (evidence consensus positive but not indeterminate). A country was indicated as publicly displaying dengue data if national dengue case numbers were displayed annually or during epidemic years at a minimum.
To calculate the maximum possible population at risk for dengue virus transmission we obtained total population counts from the Global Rural Urban Mapping Project (GRUMP) for the 128 countries identified as dengue-present. The GRUMP beta version provides gridded population count estimates at a 1×1 km spatial resolution for the year 2000 [26], [27]. Population counts for the year 2000 were projected to 2010 by applying country-specific urban and rural national growth rates [28] using methods described previously [29]. As 2010 forms a landmark year for many national censuses, we were able to adjust these expanded population counts using the United Nations 2010 population estimates [30].
The global distribution of dengue virus transmission as defined by evidence consensus is shown in Figures 3–7. The mapped colour scale ranges from complete consensus on dengue presence (dark red) to indeterminate consensus on dengue status (yellow) then through to complete consensus on dengue absence (dark green). A full list of the evidence used for each area and their scoring is available in Table S1 and Figure S7. In total we identified 128 countries as dengue-present (i.e. positive values outside the indeterminate range), compared to 100 from the WHO, 104 from the CDC and 118 from GIDEON. Compared to the lists produced by the WHO and CDC, we identified 41 additional countries where evidence consensus for presence was outside the indeterminate range yet dengue-absent status was assigned by at least one of these health organisations.
Even after performing the sensitivity analysis described earlier, the number of countries defined by our methodology as dengue-present but defined by WHO/CDC as absent never dropped below 36 (Table 1). We therefore suggest that this list of 36 countries be subject to a review regarding their current health organisation dengue-absent classification. Of these countries, 31 had at least moderate consensus on dengue presence in our final analysis.
The majority of these newly identified dengue-present countries were in Africa and the evidence type that allowed greatest identification was returning traveller reports. These sporadic reports established preliminary evidence, which we improved with supplementary evidence and questionnaire retrieval to clarify dengue status if possible (Table 2). Outside of Africa, the remaining newly identified countries were almost exclusively islands in the Indian and Pacific Oceans and in the Caribbean. The reason for a lack of dengue presence identification by health organisations here is likely the longer interval between epidemics in small isolated nations, resulting in sparse data which different health organisations have interpreted inconsistently. Inclusion of less official surveillance evidence, such as ProMED reports, that detected background case loads alongside officially reported outbreaks allowed our distinction of these areas as in fact dengue-present.
A total of 3.97 billion people live in these 128 countries outside the indeterminate consensus class. Of these, 824 million live in urban and 763 million in peri-urban areas. These numbers therefore constitute plausible preliminary estimates for the maximum possible population at any risk of dengue transmission. We expect more comprehensive population at risk calculations to refine this figure and quantify levels of risk in our future work, allowing us to give a more accurate estimate.
Public display of dengue data varied by continent (Figure 8). In total, 46 of 128 dengue-present countries displayed annual dengue case numbers. Of these, the highest reporting coverage was observed in Asia and the Americas where 55% and 57% of countries respectively reported dengue publically. This figure was comparably worse in the Pacific (29%) and Africa, Saudi Arabia, Yemen and the western Indian Ocean islands (Africa+) where just 7% of dengue-present countries publicly report dengue and none on mainland continental Africa. There were no regional patterns in the level of dengue case data provided, although the publicising of epidemiological weeks in some Central and South American countries tended to provide higher levels of detail. Deaths due to DHF/DSS/severe dengue were far less commonly reported, although the data are available for some Central American countries. Even allowing for variable internet usage and endogenous public health systems, we highlight the magnitude of disparity in countries' provision of freely available dengue data.
Dengue presence is well documented in the Americas with a continuous set of good- or complete- consensus countries from southern Brazil to the Mexico-U.S.A. border (Figure 3). However, a general regional classification was not producible as in some cases such as Montserrat and Saint Vincent and the Grenadines, where moderate rather than good consensus was found. With only 22% of dengue-present Caribbean countries displaying dengue data publically, dengue status in these small island nations that are characterised by longer inter-epidemic periods proved considerably more heterogeneous. This was mainly due to a lack of confirmed indigenous cases during recent epidemics.
Other regions of uncertainty reflect dynamic dengue status at the limits of the disease distribution. Lower consensus estimates in areas of Florida and Argentina result from reliance on smaller amounts of evidence from recent epidemics. Although the disease extent is better described in Florida (both in terms of resolution and consensus) due to greater data availability, uncertainty is still present due to the unknown persistence of recent events. A similar pattern of uncertainty exists in Texas but for different reasons, being that the occurrence evidence is older and six of seven counties have no record of occurrence since the late 1980s.
A total of 58% of Africa+ countries had a good consensus or better but Africa still showed the highest levels of uncertainty in countries with poor consensus. Concentrations of higher consensus were identified in East and West Africa (Figure 4). Multiple seroprevalence surveys over several years [31], [32], [33], [34], [35] made the most significant contribution in defining East Africa's higher-consensus cluster which ranges from Sudan to Tanzania with only Uganda, Rwanda and Burundi exhibiting poor or worse evidence consensus. In addition to this, evidence of outbreaks in coastal areas of Yemen, Saudi Arabia and some evidence of spill-over into Egypt added certainty to the definition of the East Africa high-consensus cluster. Although not as contiguous a tract of countries, a higher-consensus region also exists in West Africa from Senegal to Gabon. Inclusion of reported dengue cases in travellers and soldiers returning from West Africa was available for 13 countries and proved the most useful information in this region.
Outside of these higher-consensus regions, evidence consensus is low and a series of countries with moderate or worse consensus can be identified from Chad to Mozambique with only the Democratic Republic of Congo exhibiting good evidence consensus. For many of these countries, there are sporadic reports of dengue occurrence combined with poor disease surveillance and a general lack of data. Dated seroprevalence surveys in areas where many other arboviruses are circulating did little to increase certainty. These factors result in a positive evidence consensus that is nevertheless highly uncertain in large portions of Africa. Even where evidence was available from contemporary epidemics, such as in the case of the western Indian Ocean islands, it was often devalued because there was a lack of clinical differentiation between dengue and chikungunya despite epidemics coinciding. The lack of clear clinical distinction between the two diseases [36] makes the scale of dengue here difficult to identify and as a result, some countries (such as Reunion) were identified as having low consensus.
Despite the widespread uncertainty in dengue status in many African countries, we were able to differentiate multiple levels of uncertainty. Angola and Mozambique both show lower consensus due to dated evidence forms, yet they are still distinguishable from countries with no evidence or just sporadic occurrences such as Zambia or Congo.
A wide variety of contemporary evidence allowed us to display a near continuous distribution of good or complete evidence consensus countries from Indonesia to as far north as Pakistan and Zhejiang, China (Figure 5). Within this dengue-present area, 58% of countries publicly displayed dengue data (Figure 8) and many reported dengue case data with a high spatial resolution. Minor exceptions to this continuous distribution occur in southern China and North-East India largely due to a lack of contemporary evidence. In Gunagxi and Hainan there is little research interest or case data in recent years despite occurrences in urban centres further along the Chinese coast [37], [38], [39]. In North-East India, lower consensus was observed due to a lack of reported cases in recent years combined with the arrival of chikungunya in the area which complicates any potential dengue reporting [40].
Evidence consensus in Asia is lowest in central Asia where contemporary dengue occurrence records combined with low surveillance capacity results in an unclear boundary to the disease. While evidence for dengue presence in the lowland urban centres of Pakistan is accurate and contemporary, reports from the more remote north-west provinces are contemporary, but not accurate [41], [42], [43]. This makes determining the extent further north into remote and data-deficient areas of Afghanistan and central Asia difficult to assess. We also found serologic evidence consistent with dengue presence in Turkey [44] and Kuwait [45], reducing evidence consensus for absence in these countries despite not belonging to any known cluster of dengue-present countries.
Although no countries in Europe were defined as dengue-present, sporadic indigenous transmission events have lowered consensus in some countries (Figure 6). Since the invasion and spread of Ae. albopictus along the Mediterranean coast [46], indigenous dengue transmission has been detected in Marseilles, France and Korčula, Croatia (both regions have moderate consensus on dengue absence) and chikungunya has been found in Italy (having good consensus on dengue absence) [47], [48], [49]. These isolated events do not in themselves confer dengue presence, but increased surveillance will be required in light of the Ae. albopictus invasion to maintain this status. This, combined with the lower levels of healthcare expenditure, has led to an observed greater uncertainty in some eastern European states.
In general, consensus on dengue presence and absence was well defined across Australia and the Pacific islands, with 85% of countries showing good or complete evidence consensus (Figure 7). Where low consensus was observed, it was largely due to a lack of contemporary evidence despite Pacific-wide dengue epidemics such as in Niue, Nauru, Tuvalu and Papua New Guinea. The duration between epidemics is typically longer in the Pacific and consensus is subject to continual change; for example, in the Marshall islands evidence consensus was upgraded from moderate to complete in the wake of the December 2011 epidemic, which came two decades after the last reported epidemic [50]. Such fluctuation is not entirely unexpected from remote, isolated communities, however. Even though evidence consensus decreases with time, it still remains positive, allowing for potential re-occurrence.
Lower evidence consensus was observed for Papua New Guinea due to a lack of reported case data since the 1980's, yet multiple literature sources suggest that dengue is still widespread [51], [52], [53]. While dengue occurrence is closely documented in some counties on the Australian coast, the serologic results from Charters Towers has contributed to uncertainty over the inland extent of the disease in Queensland [54]. Only the governments of Australia, New Caledonia and the Solomon Islands report dengue case numbers publicly. Considering the long intervals between epidemics in the Pacific, it is perhaps unsurprising that this is not a priority.
Here we present the distribution of dengue virus transmission as assessed by evidence-based consensus. By emphasising the need for accurate, contemporary evidence through a weighted scoring system, we were able to identify areas where dengue status was more uncertain, particularly in Africa and Central Asia, and identify evidence gaps where surveillance might be better targeted to more accurately assess dengue status. By including a wide variety of evidence we were able to cast doubt on dengue status in countries previously described by health organisations as dengue-absent.
While many studies have focussed on the future threat of dengue as a result of range expansion or climate change, this is the first to assess the entirety of knowledge regarding the extent of current virus transmission. We have found that evidence of dengue virus transmission is temporally dynamic and that a contemporary map must emphasise evidence by weighting it appropriately. By increasing temporal resolution to one inter-epidemic period, we have extended the approach of Van Kleef et al. [5] who used evidence from literature searches to produce distribution maps pre- and post- 1975. Focussing on a higher resolution timescale for dengue evidence is necessary if we are to infer changes in the evidence-based distribution of dengue.
The suggestion that dengue is an under-recognised problem in Africa is not a new one [55], [56], [57], but here we present a detailed summary of the specific gaps in evidence that exist in different regions. We show that consensus mapping is flexible to regional differences in evidence availability and as such can produce meaningful outputs in resource-high and low settings. The evidence that dengue is widespread in Africa implies that the continent is underrepresented by occurrence points in the model-based approaches that have been used to investigate the distribution of dengue so far [1], [10], [11]. If we are to estimate the burden of dengue in Africa with any fidelity, available data and their underlying assumptions need to be reassessed.
Evidence consensus maps provide a more informative alternative to existing country-level maps, such as those provided by the WHO [12] and CDC [58]. As presence or absence exists on a continuous scale of certainty, evidence consensus approaches are more adaptable to incorporating diverse forms of dengue evidence ignored by these organisations in producing their estimates. While we show that different evidence weightings in our scoring system do not significantly alter the result, we were unable to formalise a statistical validation of these weightings due to lack of a training dataset. Our results provide the best estimate thus far of where such data are most needed and comparisons with higher-consensus countries in similar settings should form the first step in directing regional surveillance. Development of methodologies to make approaches such as consensus mapping more reliable is needed as dengue status will increasingly rely on harder-to-quantify evidence types, such as internet search engine terms [59] and multi-language internet text-mining systems [60], [61]. The success of automated disease surveillance systems such as HealthMap [16], [17], and Biocaster [60], [63] have already been demonstrated. We believe evidence consensus provides the best platform for integrating these diverse forms of information now available for disease occurrence to create an up-to-date, high-resolution map of dengue evidence, whilst retaining important assessments of certainty. We also intend to extend our own data collection and accessibility with a new website linked to the Global Health Network (http://globalhealthtrials.tghn.org/) that will allow evidence contribution from members and will provide a key platform for display of dengue data and consensus maps. Although the current approach was used to map the distribution of dengue, minor modifications to the scoring system would allow it to be utilised for a variety of diseases for which the quality of presence evidence is spatially variable.
In this work, our aim was to produce a standardised methodology that used the largest variety of evidence to assess country dengue status, whilst still being applicable in diverse healthcare settings and suitable at multiple spatial scales. We considered the stark contrast in evidence available in Africa as compared to the rest of the world. Our results show that the inclusion of supplementary evidence (used in 44% of African countries but only 11% of the rest), healthcare expenditure information (for case data absences) and questionnaires increased evidence consensus in these countries without impacting the methodology applied to the rest of the world. Similarly, we are aware that increasing resolution to Admin1 or Admin2 level may well reduce the evidence available for calculating evidence consensus in each area compared to country-level calculations. As a result, we carefully chose which countries should have increased spatial resolution based on whether sufficient evidence was available in smaller administrative units. We also limited the selection of these countries to those at the limits of the disease's distribution, as data deficiencies in these regions more accurately represent the uncertainty on dengue status given the dynamic nature of global dengue spread. Here we present the most flexible methodology available, to date, for overcoming these problems. We have demonstrated that a systemic approach with relevant optional categories has allowed us to utilise the maximum variety of evidence available for assessing dengue status in the widest variety of situations.
We also openly provide a full list of evidence for each country by category (Table S1). We intend to continue data acquisition by including more endogenous, local evidence through questionnaires and local language search methods, which we expect will allow us to further customise our methodology and assess dengue status in places where we are currently uncertain.
Mapping by evidence consensus is a useful approach to quantifying contemporary disease evidence and can be further integrated with geo-spatial modelling to produce worldwide continuous surfaces of dengue risk [64]. Current mapping approaches use presence/absence expert opinion maps to sample pseudo-presence or pseudo-absence points to increase the number of data points on which to base their prediction [65], [66], [67], [68]. Pseudo-sampling could be improved by using the continuous scale of evidence consensus to either affect sample number or point weight within the geo-spatial model. This will lead to more robust, higher resolution dengue maps which are currently in progress [69]. By combining uncertainty assessment from consensus mapping with high-resolution predictions using geo-spatial modelling, we will be able to make more accurate predictions of disease burden with associated confidence intervals made explicit. This will then provide a series of up-to-date assessments of global dengue distribution, thus providing key information to assess dengue spread and the impact of control measures.
|
10.1371/journal.pntd.0006997 | Potential of Aedes albopictus to cause the emergence of arboviruses in Morocco | In 2015, the mosquito Aedes albopictus was detected in Rabat, Morocco. This invasive species can be involved in the transmission of more than 25 arboviruses. It is known that each combination of mosquito population and virus genotype leads to a specific interaction that can shape the outcome of infection. Testing the vector competence of local mosquitoes is therefore a prerequisite to assess the risks of emergence. A field-collected strain of Ae. albopictus from Morocco was experimentally infected with dengue (DENV), chikungunya (CHIKV), zika (ZIKV) and yellow fever (YFV) viruses. We found that this species can highly transmit CHIKV and to a lesser extent, DENV, ZIKV and YFV. Viruses can be detected in mosquito saliva at day 3 (CHIKV), day 14 (DENV and YFV), and day 21 (ZIKV) post-infection. These results suggest that the local transmission of these four arboviruses by Ae. albopictus newly introduced in Morocco is a likely scenario.
Trial registration: ClinicalTrials.gov APAFIS#6573-201606l412077987v2.
| The Asian tiger mosquito Aedes albopictus is responsible for the transmission of several arboviruses such as dengue and chikungunya viruses. In 30 to 40 years, it has extended its geographical distribution in both tropical and temperate regions of all continents. The species was first detected in September 2015, in Rabat, Morocco. Using experimental infections, we demonstrated that Ae. albopictus Morocco are competent to transmit zika and yellow fever viruses in addition to the transmission of dengue and chikungunya viruses. Our results are central to suggest developing the most effective national surveillance program and to designing the most suitable control strategy to avoid the mosquito spreading beyond its point of entry in Morocco.
| Over the past decades, arboviruses caused acute emergences leading to global pandemics. Dengue viruses (DENV; family Flaviviridae, genus Flavivirus) are responsible for 390 million infections per year including 96 million symptomatic cases [1]. In 2005, chikungunya virus (CHIKV; family Togaviridae, genus Alphavirus) emerged outside Africa producing devastated outbreaks in all continents [2]. While its importance was underestimated, zika virus (ZIKV; family Flaviviridae, genus Flavivirus) hit Brazil in 2015 causing several million cases in the Americas [3] and severe unusual symptoms such as Guillain-Barré syndrome and congenital microcephaly. Despite the availability of an efficient vaccine 17D, yellow fever virus (YFV; family Flaviviridae, genus Flavivirus) continues to cause human fatalities in South America and Sub-Saharan Africa.
All four arboviruses share the same mosquito vectors: Aedes aegypti and Aedes albopictus. Ae. aegypti is an urban mosquito feeding exclusively on humans [4] and Ae. albopictus colonizes a larger range of sites and feeds on both animals and humans [5]. While Ae. aegypti took several centuries to invade most countries in the world [6], Ae. albopictus took only a few decades to establish stable colonies worldwide [7]. Native to Southeast Asia, Ae. albopictus has invaded America, Africa and Europe during the last 40 years [8]. In Europe, it was introduced in 1979 in Albania and then in Italy in 1990. It is now present in 20 European countries [9]. In Africa, Ae. albopictus was first reported in the early 1990s in South Africa [10] and Nigeria [11]. Thereafter, it was described in several West and Central African countries: Cameroon in 2000 [12], Equatorial Guinea in 2003 [13], Gabon in 2007 [14], Central African Republic in 2009 [15], and Republic of Congo in 2011 [16]. More recently, it was detected in Mali [17], Mozambique [18] and São Tomé and Príncipe [19]. In North Africa, Ae. albopictus was detected in Algeria in 2010 [20] then in Morocco in 2015 [21].
Morocco is considered a low prevalent country for mosquito-borne diseases [22]. However, since 1996, the country has faced West Nile virus (WNV) with three epizootic episodes: 1996, 2003 and 2010 [23, 24]. In 2008, a serosurvey of wild birds confirmed the circulation of WNV in native birds [25]. Other arboviruses like Usutu virus and Rift valley fever virus (RVFV) have never been reported despite serological evidence of RVFV antibodies in camels at the border between Morocco and Mauritania [25–27]. Morocco is considered by several reports of the Intergovernmental Panel on Climate Change (IPCC) as a hotspot for climate change with its significant impact for several infectious diseases [28]. The introduction of an invasive species such as Ae. albopictus will likely cause a new public health problem. Moreover, Morocco is a tourist destination with more than 11 million visitors reported in 2017 {http://www.tourisme.gov.ma/fr/tourisme-en-chiffres/chiffres-cles}, increasing the risk of importing arboviral pathogens.
In this work, we evaluate the ability of Ae. albopictus recently introduced in Morocco to transmit CHIKV, DENV, ZIKV and YFV, where the outcome of vector infection depends on specific genotype-by-genotype (G x G) interactions between a vector population and a pathogen lineage [29]. This measure of the vector competence of field-collected mosquitoes helps to assess the risk of arbovirus emergence.
Animals were housed in the Institut Pasteur animal facilities accredited by the French Ministry of Agriculture for performing experiments on live rodents. Work on animals was performed in compliance with French and European regulations on care and protection of laboratory animals (EC Directive 2010/63, French Law 2013–118, February 6th, 2013). All experiments were approved by the Ethics Committee #89 and registered under the reference APAFIS#6573-201606l412077987 v2.
During the national surveillance plan implemented in 2016 to establish the geographical distribution of Ae. albopictus in Morocco, five ovitraps less than 500 m apart were placed on a street of the Agdal neighborhood in Rabat (33°59'20.9′′ N, 6°51′07.9′′W). Ovitraps were checked for eggs once a week from May to November 2016 and were brought back to the laboratory to be stored in humid chambers (relative humidity of 80%) before being sent to Institut Pasteur in Paris to perform the vector competence studies.
After hatching, larvae were split into pans of 200 individuals and supplied every 2 days with a yeast tablet dissolved in 1L of dechlorinated tap water. All immature stages were reared at 26±1°C. Emerging adults were maintained at 28±1°C with a 16L:8D cycle, 80% relative humidity and supplied with a 10% sucrose solution. Females were fed twice a week on anaesthetized mice (OF1 mice, Charles River laboratories, France). Resulting F2 adults were used for vector competence assays. It should be noted that variations of oral susceptibility to an arbovirus can be considered negligible in fewer than five laboratory generations [30].
CHIKV strain (06.21) was isolated from a patient on La Reunion Island in 2005 [31]. After isolation on Ae. albopictus C6/36 cells, this strain was passaged twice on C6/36 cells and the viral stocks produced were stored at -80°C prior to their use for mosquito oral infections. DENV-2 strain provided by Prof. Leon Rosen, was isolated from a human serum collected in Bangkok (Thailand) in 1974 [32] and had been passed only in different mosquito species (2 times in Ae. albopictus, 2 times in Toxorhynchites amboinensis, and one time in Ae. aegypti) by intrathoracic inoculation. Viral stocks were obtained by inoculating C6/36 cells. ZIKV strain (NC-2014-5132) originally isolated from a patient in April 2014 in New Caledonia was passaged five times on Vero cells; this strain belongs to the same genotype than the ZIKV strains circulating in Brazil in 2015 [33]. Lastly, a YFV strain (S79) belonging to the West African lineage, was isolated from a human case in Senegal in 1979 [34]. YFV-S79 was passaged twice on newborn mice and two times on C6/36 cells.
Six to eight batches of 60 7–10 day old females were exposed to an infectious blood meal containing 1.4 mL of washed rabbit erythrocytes and 700 μL of viral suspension. The blood meal was supplemented with ATP as a phagostimulant at a final concentration of 1 mM and provided to mosquitoes at a titer of 107.2 plaque-forming unit (pfu)/mL for ZIKV, 106.5 focus-forming unit (ffu)/mL for YFV and 107 ffu/mL for CHIKV and DENV, using a Hemotek membrane feeding system. Mosquitoes were allowed to feed for 15 min through a piece of pork intestine covering the base of a Hemotek feeder maintained at 37°C. Fully engorged females were transferred in cardboard containers and maintained with 10% sucrose under controlled conditions (28±1°C, relative humidity of 80%, light:dark cycle of 16 h:8 h) for up to 21 days with mosquito analysis at 3, 7, 14 and 21 days post-infection (dpi). For each virus, 21–30 mosquitoes were examined at each dpi.
For each mosquito examined, body (abdomen and thorax) and head were tested respectively for infection and dissemination rates at 3, 7, 14 and 21 dpi. For this, each part was ground in 250 μL of Leibovitz L15 medium (Invitrogen, CA, USA) supplemented with 3% FBS, and centrifuged at 10,000×g for 5 min at +4°C. The supernatant was processed for viral titration.
Mosquitoes examined previously were also tested for viral transmission by collecting saliva using the forced salivation technique [35]. Mosquitoes were anesthetized on ice and legs and wings were removed. The proboscis was then inserted into a pipette tip containing 5 μL of fetal bovine serum (FBS). After 30 min, the tip content was transferred in 45 μL of L15 medium. Saliva was then titrated to estimate the transmission rate.
CHIKV, DENV and YFV were titrated by focus fluorescent assay and ZIKV by plaque forming assay as ZIKV cannot produce distinct viral foci on mosquito cells.
For mosquitoes challenged with CHIKV, DENV or YFV, saliva, head and body homogenates were titrated by focus fluorescent assay on Ae. albopictus C6/36 cells [36]. Samples were serially diluted and inoculated onto C6/36 cells in 96-well plates. After an incubation of 3 days for CHIKV, and 5 days for YFV and DENV-2 at 28°C, cells were stained using hyper-immune ascetic fluid specific to each virus as the primary antibody (CHIKV: provided by the French National Reference Center for Arbovirus at the Institut Pasteur, YFV: OG5 NB100-64510; Novusbio, CO, USA, and DENV: Ms X Dengue complex MAB 8705, Millipore, MA, USA) and Alexa Fluor 488 goat anti-mouse IgG (Life Technologies, CA, USA) as the secondary antibody. Saliva titers were expressed as ffu/saliva.
For ZIKV, body and head suspensions were serially diluted and inoculated onto monolayers of Vero cells in 96-well plates. Cells were incubated for 7 days at 37°C then stained with a solution of crystal violet (0.2% in 10% formaldehyde and 20% ethanol). Presence of viral particles was assessed by CPE detection. Saliva was titrated on monolayers of Vero cells in 6-well plates incubated 7 days under an agarose overlay. Saliva titers were expressed as pfu/saliva.
Means, standard deviations, 95% confidence interval were calculated and statistical analyses were performed using the Stata software (StataCorp LP, Texas, and USA). The effect of virus and dpi on infection, dissemination and transmission rates was evaluated using Fisher’s exact test. The titer of viral particles in mosquito saliva was compared across groups using a Kruskall-Wallis non parametric test. P-values<0.05 were considered statistically significant. Heatmaps were built under R (v 3.3.1) (https://www.R-project.org).
Mosquito females were exposed to four separate infectious blood meals containing CHIKV, DENV, ZIKV or YFV. The first step after the ingestion of the infectious blood meal is the infection of the midgut which is appraised by calculating the infection rate (IR) corresponding to the proportion of mosquitoes with an infected midgut. At 3 dpi, Ae. albopictus Morocco were more infected with CHIKV (Fig 1; Fisher’s exact test: p<10−4, df = 3) with an IR reaching 93% (N = 30) whereas with the 3 other viruses, IRs were lower than 20% (N = 30). At 7 dpi, the IR with CHIKV reached 100% (N = 30) and remained significantly lower with DENV (60%; N = 30), ZIKV (60%; N = 30) and YFV (26.7; N = 30) (Fisher’s exact test: p<10−4, df = 3). At 14 dpi, mosquitoes become more infected with DENV reaching 90% (N = 30) close to CHIKV (86.7%, N = 30) (Fisher’s exact test: p = 0.69, df = 3) but significantly higher than with ZIKV (66.7%, N = 30), and YFV (20%, N = 30) (Fisher’s exact test: p<10−4, df = 30). At 21 dpi, the same pattern was observed: IRs were higher with CHIKV (90%, N = 30) and DENV (100%, N = 21) than with ZIKV (69.6%, N = 23) and YFV (53.3%, N = 30) (Fisher’s exact test: p<10−4, df = 3). IRs with all viruses increased along with dpi except with CHIKV which remained high (>86%) very early from 3 dpi. The lowest IRs were obtained with YFV fluctuating from 6.7% at 3 dpi to 53.3% at 21 dpi.
Once the midgut is infected, viral particles can disseminate from the midgut to internal organs and tissues. The dissemination rate (DR) gives the number of mosquitoes with infected heads among mosquitoes with infected midgut. At 3 dpi, only CHIKV was detected in mosquito heads (Fig 2; 28.6%, N = 28). At 7 dpi, DR with CHIKV reached 53.3% (N = 30) and only 5.5% (N = 18) with DENV (Fisher’s exact test: p<10−4, df = 3). At 14 dpi, DRs with CHIKV (65.4%, N = 26) and DENV (59.2%, N = 27) were higher and similar (Fisher’s exact test: p = 0.65, df = 1) compared to YFV (33.3%, N = 6) and ZIKV (25%, N = 20) which were both lower and comparable (Fisher’s exact test: p = 0.69, df = 1). At 21 dpi, DRs for each virus were significantly different (Fisher’s exact test: p<10−4, df = 3) and slightly higher than the DRs at 14 dpi. Viral dissemination started earlier with CHIKV at 3 dpi while it was only at 7 dpi with DENV and 14 dpi with YFV and ZIKV. The lowest DRs were obtained with ZIKV maintained at 25% at 14 and 21 dpi.
After the virus has spread into the general cavity of the mosquito and infected the salivary glands, the virus must be excreted in saliva for subsequent transmission. The transmission rate (TR) is defined as the proportion of mosquitoes delivering infectious saliva among mosquitoes having disseminated the virus (Fig 3A). At 3 and 7 dpi, viral particles could be detected in saliva of mosquitoes infected with CHIKV, with TRs of 37.5% (N = 8) and 68.7% (N = 16) respectively. At 14 dpi, TR with YFV (50%, N = 2) predominated over TRs with CHIKV (35.3%, N = 17) and DENV (11.1%, N = 18), TR with ZIKV remaining at 0%; no significant difference was observed among all TRs (Fisher’s exact test: p = 0.14, df = 3). At 21 dpi, transmission with ZIKV became detectable with a TR of 50% (N = 4), not significantly different from TRs with DENV (26.3%, N = 19), CHIKV (17.4%, N = 23), and YFV (10%, N = 10) (Fisher’s exact test: p = 0.36, df = 3). Transmission started early at 3 dpi with CHIKV, at 14 dpi with DENV and YFV, and at 21 dpi with ZIKV with respectively, a mean of 2.06±0.60 Log10 ffu/saliva (N = 3), 0.87±0.38 Log10 ffu/saliva (N = 2), 1.53 Log10 ffu/saliva (N = 1), and 2.71±0.01 Log10 pfu/saliva (N = 2) (Fig 3B). No significant difference was detected between all viruses at 14 dpi (Kruskal-Wallis test: p = 0.47, df = 2) and 21 dpi (Kruskal-Wallis test: p = 0.10, df = 3). The highest number of viral particles was detected in saliva of mosquitoes infected with YFV and examined at 21 dpi: TR of 50% (2 among 4 mosquitoes with viral dissemination), 2 females delivering 2.70 Log10 pfu (500) and 2.72 Log10 pfu (530) infectious particles.
Whereas IR, DR and TR measure the efficiency of the midgut and salivary glands barriers to modulate, respectively, viral dissemination and transmission, the transmission efficiency (TE) gives an overview of transmission potential of mosquitoes tested; it corresponds to the proportion of mosquitoes with infectious saliva among all mosquitoes examined (presenting or not a viral dissemination with infected heads). Fig 4 shows that, the highest TE was detected at 7 dpi with CHIKV, at 21 dpi with DENV, at 14/21 dpi with YFV, and at 21 dpi with ZIKV. Collectively, Ae. albopictus Morocco were more susceptible to CHIKV and secondarily, to DENV, ZIKV and YFV.
To summarize the vector competence corresponding to the overall ability of a mosquito population to be infected, to ensure the viral dissemination and to transmit the virus, heatmaps were built (Fig 5). Ae. albopictus Morocco were better infected with CHIKV from 3 dpi than with DENV and ZIKV (Fig 5A). Mosquitoes ensured an early dissemination (Fig 5B) and transmission (Fig 5C) with CHIKV (from 3 dpi) than with DENV and ZIKV. The species was less susceptible to YFV. Altogether, vector competence of Ae. albopictus Morocco depends on the virus and the dpi: it is more susceptible to CHIKV and susceptibility increases along with the dpi.
Using experimental infections, we show that the recently-introduced Ae. albopictus in Morocco were susceptible to all four viruses tested, CHIKV, DENV, YFV and ZIKV. Viral transmission was detected at 3 dpi with CHIKV, 14 dpi with DENV and YFV, and only 21 dpi with ZIKV.
Even if DENV, YFV and ZIKV belong to the same genus, they behave differently in Ae. albopictus mosquitoes. Infection of the midgut increases gradually from 3 dpi: DENV infects more efficiently mosquitoes than YFV and ZIKV, YFV remaining the less successful. Dissemination of DENV from the midgut to the mosquito general cavity started at 7 dpi as observed with most populations of Ae. albopictus [37]; it takes a shorter time with Ae. aegypti, i.e. from 5 dpi [38]. DENV dissemination is more strongly inhibited at early dpi than later meaning that the role of midgut as a barrier is diminished with dpi. Transmission of DENV was observed from 14 dpi suggesting an intrinsic incubation period higher than 7 dpi, likely around 10 dpi [37]. With ZIKV and YFV, dissemination was observed only at 14 dpi, YFV spreading at a higher rate than ZIKV suggestive of a stronger role of the midgut barrier with YFV. Transmission was detected at 14 dpi with YFV as observed with other Ae. albopictus populations [39] and 21 days with ZIKV which is longer than expected [40].
CHIKV presents a different profile. This alphavirus infects, disseminates and is transmitted more intensively and more quickly than the three other viruses. This viral strain presents an amino acid substitution (A226V) in the envelope glycoprotein E1 [31] favoring the viral transmission by Ae. albopictus [41, 42]. Importantly, exposure of infected mosquitoes to lower temperatures (lower than 25°C) compatible to values recorded in Morocco can modulate transmission [37]. It has been demonstrated that Ae. albopictus were able to better transmit CHIKV at a temperature lower than 28°C [43].
These assessments of vector competence of Ae. albopictus from Morocco to CHIKV, DENV, ZIKV and YFV are important for appraising the risk of local transmission. ZIKV shows the longer extrinsic incubation period (EIP) which refers to the time between the uptake of the virus during the blood feeding and the delivery of the virus by vector bite after successful infection and dissemination in the mosquito. If the EIP is longer than the daily survival rate of the mosquito, the risk of transmission is low. By shortening mosquito lifespan, vector control measures reduce disease transmission [44]. However, other factors such as environmental factors, e.g. the temperature, may influence the vector competence [43]. The vector competence and the EIP both contribute to estimating the vector capacity which describes the basic reproductive rate of a pathogen by a vector [44]. A high abundance of the vector [45], increased contacts between the vector and humans (i.e. anthropophily of mosquitoes) [5] and a high proportion of immunologically naïve humans, are also factors that should be considered in estimating the risk of emergence. Introductions of viremic travelers from endemic countries for all these viruses may initiate local transmission and outbreaks. Therefore surveillance of travelers must be reinforced.
|
10.1371/journal.pmed.1002495 | The cost-effectiveness of alternative vaccination strategies for polyvalent meningococcal vaccines in Burkina Faso: A transmission dynamic modeling study | The introduction of a conjugate vaccine for serogroup A Neisseria meningitidis has dramatically reduced disease in the African meningitis belt. In this context, important questions remain about the performance of different vaccine policies that target remaining serogroups. Here, we estimate the health impact and cost associated with several alternative vaccination policies in Burkina Faso.
We developed and calibrated a mathematical model of meningococcal transmission to project the disability-adjusted life years (DALYs) averted and costs associated with the current Base policy (serogroup A conjugate vaccination at 9 months, as part of the Expanded Program on Immunization [EPI], plus district-specific reactive vaccination campaigns using polyvalent meningococcal polysaccharide [PMP] vaccine in response to outbreaks) and three alternative policies: (1) Base Prime: novel polyvalent meningococcal conjugate (PMC) vaccine replaces the serogroup A conjugate in EPI and is also used in reactive campaigns; (2) Prevention 1: PMC used in EPI and in a nationwide catch-up campaign for 1–18-year-olds; and (3) Prevention 2: Prevention 1, except the nationwide campaign includes individuals up to 29 years old.
Over a 30-year simulation period, Prevention 2 would avert 78% of the meningococcal cases (95% prediction interval: 63%–90%) expected under the Base policy if serogroup A is not replaced by remaining serogroups after elimination, and would avert 87% (77%–93%) of meningococcal cases if complete strain replacement occurs. Compared to the Base policy and at the PMC vaccine price of US$4 per dose, strategies that use PMC vaccine (i.e., Base Prime and Preventions 1 and 2) are expected to be cost saving if strain replacement occurs, and would cost US$51 (−US$236, US$490), US$188 (−US$97, US$626), and US$246 (−US$53, US$703) per DALY averted, respectively, if strain replacement does not occur.
An important potential limitation of our study is the simplifying assumption that all circulating meningococcal serogroups can be aggregated into a single group; while this assumption is critical for model tractability, it would compromise the insights derived from our model if the effectiveness of the vaccine differs markedly between serogroups or if there are complex between-serogroup interactions that influence the frequency and magnitude of future meningitis epidemics.
Our results suggest that a vaccination strategy that includes a catch-up nationwide immunization campaign in young adults with a PMC vaccine and the addition of this new vaccine into EPI is cost-effective and would avert a substantial portion of meningococcal cases expected under the current World Health Organization–recommended strategy of reactive vaccination. This analysis is limited to Burkina Faso and assumes that polyvalent vaccines offer equal protection against all meningococcal serogroups; further studies are needed to evaluate the robustness of this assumption and applicability for other countries in the meningitis belt.
| Meningococcal epidemics remain a major cause of morbidity and mortality in the meningitis belt, a region in sub-Saharan Africa with an estimated population of 400 million people.
The advent of affordable polyvalent meningococcal vaccines offers the opportunity for more effective control of Neisseria meningitidis in this region.
It is not yet clear how best to use these novel polyvalent meningococcal vaccines to control meningococcal epidemics in the African meningitis belt.
We developed a mathematical model that describes the key characteristics of meningococcal epidemics within districts of Burkina Faso.
Our model estimates the health impact and costs of different meningitis vaccines and vaccination strategies (for example, in routine, reactive, or catch-up mass immunization campaigns).
We project that a nationwide immunization campaign in young adults with a polyvalent meningococcal conjugate vaccine and the addition of this new vaccine into the Expanded Program on Immunization would be cost-effective.
The estimated incremental cost of this new strategy is less than US$1,980 (3 times per capita gross domestic product of Burkina Faso in 2015) per disability-adjusted life year averted, compared with the current WHO-recommended strategy.
This work suggests an opportunity to revisit the current WHO strategy for meningitis control in sub-Saharan Africa once affordable polyvalent meningococcal conjugate vaccines become available.
Our results indicate that these novel vaccines can be used in a cost-effective manner to control meningococcal epidemics if adopted within the Expanded Program on Immunization and used in a catch-up nationwide vaccination program in Burkina Faso, but further studies are needed to evaluate the effectiveness and cost-effectiveness of this policy in other countries.
| N. meningitidis remains a major cause of morbidity and mortality in the meningitis belt, a region in sub-Saharan Africa extending from Senegal to Ethiopia, with an estimated population of 400 million people [1]. In this region, meningitis epidemics occur sporadically, resulting in tens of thousands of cases and imposing substantial economic costs to affected communities. Since the late 1970s, control of meningitis epidemics in the meningitis belt has relied on reactive vaccination campaigns using polysaccharide vaccines. These reactive campaigns are triggered once an outbreak surpasses an epidemic threshold defined by the World Health Organization (WHO). While timely implementation of this reactive strategy may blunt the severity of meningococcal epidemics [2–4], in many settings the impact of reactive vaccination campaigns is limited by delays in the diagnosis and reporting of meningitis cases and subsequent delays in the launch of these vaccination activities [5].
In December 2010, a new group A meningococcal conjugate vaccine, PsA-TT (MenAfriVac) [6], was introduced in Burkina Faso, Mali, and Niger, and within one month, almost 20 million individuals, ages 1–29 years, were vaccinated [7–9]. The introduction of MenAfriVac across the meningitis belt has been associated with a dramatic reduction of meningitis A cases and carriage during subsequent seasons [10–12]. While MenAfriVac has been successful in controlling group A disease and transmission, persistent threat from other meningococcal serogroups [13,14] has spurred development of polyvalent vaccines that target non-A serogroups, including C, Y, W, and X [15–17].
The advent of affordable polyvalent meningococcal vaccines offers the opportunity for more effective control and potential elimination of N. meningitidis epidemics in the meningitis belt, but the health impact and costs of alternative vaccination strategies are not yet clear [18]. While available polyvalent meningococcal polysaccharide (PMP) vaccines can only be used in reactive campaigns because they are poorly immunogenic in children under 2 years old [19,20] and induce a short period of immunity in adults [1,21], polyvalent meningococcal conjugate (PMC) vaccines are immunogenic among infants and can induce longer-term protection. Therefore, PMC vaccines can potentially replace MenAfriVac in the Expanded Program on Immunization (EPI) and can also be used in reactive and/or mass preventive vaccination campaigns. Consensus on how to best use these novel polyvalent meningococcal vaccines has not yet been achieved.
Here, we describe a mathematical transmission model that captures the key characteristics of meningococcal epidemics in Burkina Faso as well as the costs associated with routine, reactive, and preventive vaccination campaigns. We utilize this model to investigate the health effects and costs associated with alternative vaccination policies that can inform the use of PMP and PMC vaccines.
We developed a stochastic compartmental model of meningococcal transmission to capture the essential characteristics of meningococcal epidemics within 55 districts of Burkina Faso. This level of spatial disaggregation by district is necessary to model reactive vaccination campaigns that are triggered within each district upon passing the WHO epidemic threshold of 10 cases per 100,000 population per week [22,23]. To allow for age-specific mixing patterns and targeting of vaccinations, the simulated population was stratified into relevant age groups. We used a gravity model to describe the mixing pattern of individuals residing in different districts (see S1 Text).
The adoption of MenAfriVac as part of the EPI is expected to eliminate serogroup A meningococcal epidemics in the meningitis belt by 2020 [10,11,24–26]. Our model therefore assumes that there is no circulation of serogroup A and aggregates all remaining serogroups covered in PMC and PMP vaccines (Fig 1). Upon infection, individuals become asymptomatic carriers, but only a small portion of these infections lead to meningitis. Individuals with active disease and individuals with asymptomatic carriage contribute to the force of infection. Because the duration of carriage is relatively short [27], we assume that the probability of superinfection during carriage is negligible. Individuals who recover from active disease and carriers who clear carriage will remain temporarily immune to reinfection [27,28]. Details of our modeling approach are provided in S1 Text.
Because the immunity induced by MenAfriVac is serogroup specific, the possibility of strain replacement due to the rollout of MenAfriVac across the meningitis belt cannot yet be excluded [29,30]. Therefore, to account for the impact of MenAfriVac on future epidemics, we consider two extreme scenarios: (1) “with strain replacement” reflects the pessimistic assumption of essentially complete strain replacement and assumes that future epidemics will occur with similar frequency and magnitude after the introduction of MenAfriVac (Fig 2A) and (2) “without strain replacement” assumes that future serotype incidence will be similar to the past but with serogroup A excluded (i.e., potentially lower frequency and/or magnitude of epidemics). For the latter scenario (without strain replacement), we determined the weekly meningitis incidence using estimates for the proportion of confirmed meningitis cases during 2002–2015 due to non-serogroup A N. meningitidis (Fig 2B). Evaluation of vaccine strategies under these two extreme scenarios allows us to identify crude bounds on the performance of vaccine policies, given existing uncertainty about the degree to which strain replacement can be expected.
The age-specific mortality rate and life expectancy are informed by population data (S1 Text) [32,33]. The weekly clinical meningitis notification data are provided by Burkina Faso’s Ministry of Health (Fig 2A). Our model is calibrated against the age distribution of meningococcal meningitis incidence (Fig 3A) and average age-specific meningococcal carriage prevalence (Fig 3B). Because reactive campaigns are launched when the number of clinical meningitis cases exceeds the WHO epidemic threshold [22,23], the model must reproduce observed trends in clinical diagnoses (Fig 2A). To this end, we used the average, standard deviation, and periodicity (as characterized by Fourier analysis) of the time series of clinical diagnoses as additional calibration targets (Fig 3C–3E). Finally, Fig 3F demonstrates that the number of districts in which the epidemic threshold of 10 meningitis cases per 100,000 persons has been exceeded in our simulations is consistent with the epidemics observed between 2002 and 2015 in Burkina Faso.
Fig 4 displays the clinical meningitis time series from three simulated trajectories over 30 years produced by the calibrated model, in comparison with the clinical meningitis time series observed in Burkina Faso during 2002–2015. We emphasize that our goal is not to fit to the timing of past epidemics but instead to calibrate the model against the periodicity of past epidemics, in addition to calibration targets depicted in Fig 3. Details of our calibration approach are described in S1 Text.
Even after the introduction of MenAfriVac in 2010, reactive strategies using PMP vaccine, alongside the EPI with MenAfriVac, remain important for responding to epidemics caused by non-A meningococci (Base strategy in Table 1) [35]. Because PMP vaccine is poorly immunogenic in infants and children younger than 2 years old [15], an alternative strategy is to use PMC vaccine in place of both MenAfriVac in EPI and PMP vaccines in reactive campaigns (Base Prime strategy in Table 1). Expanding PMC vaccine coverage to older age groups through preventive vaccination campaigns can both mitigate the risk of meningococcal epidemics by reducing the size of the at-risk population and can potentially lead to the elimination of meningitis (Prevention 1 and 2 strategies in Table 1). While the Base Prime strategy attempts to contain local outbreaks by implementing district-level reactive campaigns, the two Prevention strategies seek to reduce the population risk of infection through preventive campaigns implemented at a national level. We note that Base Prime can potentially lead to the same level of herd immunity as Prevention strategies over the long term, but this occurs only after the population has aged sufficiently such that all young adults were vaccinated through the EPI program or after outbreaks have occurred in a sufficient number of districts.
Our model assumes that vaccinated individuals are protected against progression to invasive disease for a limited time (Fig 1), but the level and duration of protection differ for PMP and PMC vaccines. We assume that PMP- and PMC-vaccinated carriers still remain infectious and contribute to the force of infection. Because the duration of immunity provided by PMP vaccination is rather short (Table 2), we assume that PMP vaccination does not impact carriage- or disease-induced immunity. Upon vaccination with PMP, individuals with carriage- or disease-induced immunity move to PMP-Vaccinated compartments to prevent revaccination but will either return to the corresponding Immune compartments at the beginning of the following epidemic season or lose their carriage- or disease-induced immunity and join the Susceptible compartment.
For the Base and Base Prime strategies, once an epidemic is declared in a district, a vaccine procurement order will be placed. For each district, we assume that the delay between exceeding the epidemic threshold and the initiation of a reactive vaccination campaign follows a discrete Uniform distribution [2, 10] weeks (Fig F in S1 Text). For Prevention strategies, a catch-up vaccination campaign is launched in November of the first simulation year and is completed before the start of the next epidemic season. During this period, PMC vaccine will be available to all individuals who are eligible for this catch-up vaccination.
The costs of vaccination programs in Burkina Faso are borne by the government and donors; hence, demonstration of affordability is essential for these programs to be considered in practice. We therefore take the payer’s perspective in estimating the costs associated with vaccine strategies. Our model accounts for costs incurred because of meningitis case management; care for patients who experience sequelae; and the operation costs of routine, reactive, and preventive vaccination campaigns. We assumed US$0.49 per MenAfriVac dose and US$4 per PMP vaccine dose [42]; as the price of the PMC vaccine is not yet determined, we allow this price to vary from US$4 to US$10 in sensitivity analyses. To measure the health outcome associated with each vaccine strategy, we use disability-adjusted life years (DALYs) [43]. Both costs (presented in the US dollars) and health outcomes are discounted at an annual rate of 3% to 2016. The details of the cost and DALY calculations are provided in S1 Text.
We followed the Consolidated Health Economic Evaluation Reporting Standards (CHEERS) [44] to report the results of our cost-effectiveness analysis study (see S11 in S1 Text). All estimates from the model are presented as the average and 95% prediction intervals (the 2.5th and 97.5th percentiles) of 500 epidemic trajectories simulated over a 30-year period. The number of simulated trajectories was chosen such that the resulting prediction intervals were stable (i.e., the values of the 2.5th and 97.5th percentiles of estimates did not vary appreciably when additional trajectories were used). Costs and health effects of vaccinations strategies are presented with respect to the current WHO-recommended strategy of reactive vaccination campaigns using PMP vaccines (Base strategy in Table 1).
Over a 30-year simulation period in Burkina Faso using the Base strategy of the EPI with MenAfriVac and reactive immunization with PMP vaccines, we project an annual average of 5,412 meningococcal cases (95% prediction interval: 105–16,550) with strain replacement and 1,642 (32–5,794) meningococcal cases without strain replacement. Compared to a counterfactual scenario in which reactive vaccination is not used, this represents an expected reduction of 45% (26%–62%) and 43% (22%–59%) in meningococcal incidence. The relatively modest impact of this strategy is attributable to (1) delays in the launch of reactive campaigns within districts upon crossing the epidemic threshold and (2) the short duration of immunity and lack of effect on carriage offered by PMP vaccines.
Vaccination strategies that use PMC vaccines could markedly reduce the expected annual number of meningitis cases (Fig 5), but they do not eliminate the possibility of meningitis outbreaks (Fig 6). The Prevention 2 strategy results in the most dramatic impact, averting 78% (63%–90%) of cases expected to occur under the Base strategy if strain replacement occurs and averting 87% (77%–93%) if no strain replacement occurs. Our model suggests that under strategies that utilize PMC vaccines, meningitis outbreaks may recur 10–15 years after the implementation of the first mass preventive campaign, when the immunity induced by PMC vaccines begins to wane in the adult population (Fig G in S1 Text).
Fig 7 displays the expected number of vaccines required for each vaccination strategy. Under either strain-replacement assumption, the Base strategy has the highest expected annual consumption of total vaccine doses and the Base Prime strategy has the lowest expected annual consumption of vaccine doses. The wide prediction intervals for the estimated number of PMP vaccines used under the Base strategy are the result of the sporadic occurrence of meningococcal outbreaks that may trigger district-wide reactive campaigns. We also note that extending the projection horizon does not impact the estimated annual consumption of MenAfriVac, PMP, or PMC vaccines in routine programs, but it reduces the estimated annual consumption of PMC vaccines in reactive and mass preventive campaigns. This is because preventive campaigns are implemented a single time at the beginning of the projection period, and the outbreaks under the Base Prime strategy occur only in early years, when there is a pool of children and young adults who were born too early to receive PMC vaccine in their routine infant vaccination schedules.
For the scenario with strain replacement, the Base strategy is dominated by Base Prime, as the latter strategy is expected to cost less and reduce the population’s DALYs (Fig 8A and Table 3). If strain replacement does not occur, we estimate the incremental cost-effectiveness ratios (ICERs) for DALYs averted by Base Prime compared with Base at US$51 (−US$233–US$476). Per WHO recommendations, strategies that avert one DALY for less than the per capita gross domestic product (GDP) are considered “very cost-effective” and one DALY for less than three times the per capita GDP as “cost-effective” [45]. Hence, at the cost-effectiveness threshold of US$660, the per capita GDP of Burkina Faso in 2015, the Base Prime strategy is considered cost-effective with respect to the Base strategy under either strain-replacement scenario.
The ICER of Prevention 1 compared to Base Prime is estimated at US$188 (−US$6–US$402) and US$870 (US$483–US$1,599) per DALY averted for with and without strain replacement, respectively. These ICER estimates are below the cost-effectiveness threshold of US$1,980, three times the per capita GDP of Burkina Faso in 2015. Compared with the Base policy, strategies that use PMC vaccine (i.e., Base Prime and Preventions 1 and 2) are expected to be cost saving if strain replacement occurs and would cost US$51 (−US$236, US$490), US$188 (−US$97–US$626), and US$246 (−US$53–US$703) per DALY averted if strain replacement does not occur.
We also compare the performance of vaccination strategies in terms of their impact on the population’s net monetary benefit (NMB) [46,47] for varying values of cost-effectiveness threshold (ω). The expected gain in NMB of a strategy is calculated with respect to the Base strategy as ω × (additional DALYs averted by the strategy)–(additional cost of the strategy). Fig 8C and 8D confirms that strategies that use the PMC vaccine dominate Base and that Prevention 1 and 2 strategies demonstrate similar performance under both strain-replacement scenarios. As expected, the cost-effectiveness of the vaccine strategies varies between strain-replacement scenarios and all strategies—Base Prime, Prevention 1, and Prevention 2—present larger incremental benefit when strain A elimination is followed by complete strain replacement (Fig 8 and Table 3).
Our sensitivity analysis shows that reducing PMP vaccine price from US$4 to US$2 per dose does not change the conclusions about the comparative performance of these vaccination strategies (Fig H in S1 Text). While increasing the price of PMC vaccine from US$4 to US$10 per dose diminishes the cost-effectiveness of vaccination strategies that involve PMC vaccines, these strategies maintain their relative performance with respect to the Base policy (Fig I in S1 Text). If PMC vaccine price is US$10 per dose, we estimate an average cost per DALY averted by Base Prime and Prevention 1 and 2 strategies with respect to the Base strategy at US$257 (US$57–US$558), US$286 (US$84–US$546), and US$326 (US$117–US$602) when strain replacement occurs and at US$1,246 (US$631–US$2,336), US$1,369 (US$759–US$2,431), and US$1,488 (US$833–US$2,619) without strain replacement.
While the currently recommended strategy for meningitis control in sub-Saharan Africa relies on reactive vaccination campaigns using PMP vaccines in districts where the epidemic threshold is passed, our model suggests that this approach will be outperformed by alternative policies using affordable PMC vaccines. The use of PMC vaccines in the EPI and in reactive vaccination programs could markedly reduce the public health burden of meningococcal epidemics but still leaves districts at substantial risk of sporadic outbreaks. The addition of nationwide catch-up vaccination campaigns to immunize 1–18-year-olds with PMC vaccines could prevent the majority of meningococcal cases. Our results suggest that this strategy is likely to be cost-effective (and potentially cost saving) with respect to the current WHO-recommended meningitis control strategy in sub-Saharan Africa once affordable PMC vaccine becomes available.
The introduction of MenAfriVac is expected to eliminate serogroup A meningitis in the meningitis belt [10,11,24–26], but little is known about the impact of MenAfriVac on the future non-A epidemics. As expected, benefits of additional vaccination interventions are highest when the elimination of serogroup A is followed by replacement by other circulating serogroups. While we do not know the likelihood or extent of serogroup replacement, our analysis shows that the comparative performance of the vaccination strategies we considered are not meaningfully altered by this source of uncertainty. Nevertheless, sustaining strong case-based surveillance in the post-MenAfriVac will facilitate more accurate estimates of the health impact and costs of these competing strategies.
Meningococcal outbreaks in the meningitis belt are sporadic and caused by different serogroups (mainly A, C, W, and X) [11,49], and the accurate prediction of future meningitis epidemics is challenged by the absence of data to characterize competition between these serogroups [50]. Most meningitis transmission models either describe the circulation of a single serogroup or two serogroups (e.g., vaccine type and non-vaccine type) [27,28,50–53]. In our study, we assume that polyvalent vaccines will offer protection against all meningococcal serogroups that can circulate in this setting, and therefore our model aggregates all serogroups into a single vaccine type serogroup. This simplification improves tractability but would compromise the insights derived from this model if there is differential effectiveness of the vaccine by serogroup or if there are complex between-serogroup interactions that influence the frequency and magnitude of future meningitis epidemics.
While our analysis is limited to Burkina Faso, the conclusions may well apply to other hyperendemic areas in the meningitis belt, including Mali, Niger, Chad, and Northern Nigeria, as the key characteristics of meningococcal epidemics in these regions (e.g., frequency of epidemics, age distribution of cases, and age-specific carriage prevalence) are similar to those in Burkina Faso [1,54]. However, additional studies are needed to confirm the generalizability of our conclusions to other settings.
This work suggests that there is a need to revisit the current WHO strategy for meningitis control in sub-Saharan Africa once affordable PMC vaccines become available. Our model-based results indicate that PMC vaccines can be used in a cost-effective manner to control meningococcal epidemics if adopted within the EPI and used in a catch-up preventive vaccination program.
|
10.1371/journal.pbio.1000563 | HER2 Phosphorylation Is Maintained by a PKB Negative Feedback Loop in Response to Anti-HER2 Herceptin in Breast Cancer | Herceptin (trastuzumab) is used in patients with breast cancer who have HER2 (ErbB2)–positive tumours. However, its mechanisms of action and how acquired resistance to Herceptin occurs are still poorly understood. It was previously thought that the anti-HER2 monoclonal antibody Herceptin inhibits HER2 signalling, but recent studies have shown that Herceptin does not decrease HER2 phosphorylation. Its failure to abolish HER2 phosphorylation may be a key to why acquired resistance inevitably occurs for all responders if Herceptin is given as monotherapy. To date, no studies have explained why Herceptin does not abolish HER2 phosphorylation. The objective of this study was to investigate why Herceptin did not decrease HER2 phosphorylation despite being an anti-HER2 monoclonal antibody. We also investigated the effects of acute and chronic Herceptin treatment on HER3 and PKB phosphorylation in HER2-positive breast cancer cells. Using both Förster resonance energy transfer (FRET) methodology and conventional Western blot, we have found the molecular mechanisms whereby Herceptin fails to abolish HER2 phosphorylation. HER2 phosphorylation is maintained by ligand-mediated activation of EGFR, HER3, and HER4 receptors, resulting in their dimerisation with HER2. The release of HER ligands was mediated by ADAM17 through a PKB negative feedback loop. The feedback loop was activated because of the inhibition of PKB by Herceptin treatment since up-regulation of HER ligands and ADAM17 also occurred when PKB phosphorylation was inhibited by a PKB inhibitor (Akt inhibitor VIII, Akti-1/2). The combination of Herceptin with ADAM17 inhibitors or the panHER inhibitor JNJ-26483327 was able to abrogate the feedback loop and decrease HER2 phosphorylation. Furthermore, the combination of Herceptin with JNJ-26483327 was synergistic in tumour inhibition in a BT474 xenograft model. We have determined that a PKB negative feedback loop links ADAM17 and HER ligands in maintaining HER2 phosphorylation during Herceptin treatment. The activation of other HER receptors via ADAM17 may mediate acquired resistance to Herceptin in HER2-overexpressing breast cancer. This finding offers treatment opportunities for overcoming resistance in these patients. We propose that Herceptin should be combined with a panHER inhibitor or an ADAM inhibitor to overcome the acquired drug resistance for patients with HER2-positive breast cancer. Our results may also have implications for resistance to other therapies targeting HER receptors.
| HER2 (ErbB2) is a surface protein and member of the epidermal growth factor receptor (EGFR) family that is overexpressed in approximately one-fifth of breast cancers. HER2-positive breast tumours tend to be very aggressive, and patients with this type of tumour have a poor prognosis. A therapeutic monoclonal antibody called trastuzumab (Herceptin) has been designed to block HER2 signalling and is used as a treatment for patients with HER2-positive breast cancer. However, recent studies have shown that Herceptin does not decrease HER2 activation. This may be why patients invariably develop resistance if treated with Herceptin monotherapy. To date, no study has explained why Herceptin cannot abolish HER2 signalling despite being an anti-HER2 monoclonal antibody. We have found that Herceptin switches on a feedback loop that increases the production of the ADAM17 protein, a protease that in turn releases the growth factors that activate HER (ErbB) receptors. These growth factors activate HER2 and also the other members of the HER receptor family—EGFR, HER3 and HER4—in such a way as to maintain HER2 activation and cell survival in HER2-positive breast cancer cells. We have found that when Herceptin is provided in combination with ADAM17 inhibitors, the feedback loop is abrogated in cells. Furthermore, a pan-HER inhibitor that decreases the activation of other HER receptors can also inhibit the feedback loop and decrease HER2 activation when used in combination with Herceptin. We further demonstrated that the combination therapy of Herceptin with a pan-HER inhibitor is more effective than Herceptin alone in an animal model of breast cancer. We believe our results offer treatment strategies that may help overcome acquired Herceptin resistance in patients with HER2-positive breast cancer.
| Dysregulation of human epidermal growth factor (HER/ErbB) receptors is implicated in various epithelial cancers [1]. The four HER receptors are capable of dimerising with each other (homodimerisation) or with different HER receptors (heterodimerisation) upon ligand binding [2]. The homo- or heterodimerisation of the receptors results in the activation of the intrinsic tyrosine kinase domain and autophosphorylation of specific tyrosine residues in the C-terminal tail [2]. The ligand-induced HER receptor dimerisation follows a strict hierarchy, and HER2 has been shown to be the preferred dimerisation partner [3]. The crystal structure explains why HER2 is ligandless, since its extracellular domain is always in the “open” conformation, with the projection of domain II ready for dimerisation even when monomeric [4]. This fixed “open” conformation of HER2 in the absence of ligand binding (mimicking the ligand-bound form in the EGFR structure) may account for why it is the preferred dimerisation partner [3].
Herceptin (trastuzumab) is a humanised mouse monoclonal antibody 4D5 and binds to the juxtamembrane region of HER2 of domain IV [4]. However, the precise mechanisms of its action and its acquired resistance are still poorly understood. Around 15%–20% of patients with breast cancer have HER2-positive tumours, and the amplification or overexpression of HER2 has been shown to be a significant predictor for both overall survival and time to relapse in these patients [5]. Herceptin has been shown to induce tumour regression in about a third of patients with metastatic HER2-positive breast cancer, but the response is rarely sustained if Herceptin is given as a single agent [6]. Therefore, understanding the mechanisms of its acquired resistance is of paramount importance.
The current proposed primary mechanisms of action for Herceptin include HER2 receptor down-regulation and inhibition of aberrant receptor tyrosine kinase activity [7],[8]. There is strong evidence of an immune-mediated mechanism in which the interaction of Herceptin's human Fc region with immune effector cells results in the stimulation of natural killer cells and activation of antibody-dependent cellular cytotoxicity [9],[10]. Other proposed mechanisms of Herceptin's action include inhibition of basal and activated HER2 ectodomain cleavage in breast cancer cells [11], the induction of G1 arrest and cyclin-dependent kinase inhibitor p27Kip1 levels [12], or activation of PTEN [13].
Although Herceptin was developed to target the HER2 receptor, recent studies have shown that Herceptin does not decrease HER2 phosphorylation [14],[15]. Its failure to abolish HER2 phosphorylation may be a key to why acquired resistance inevitably occurs for all responders if Herceptin is given as monotherapy. To date, no studies have explained why Herceptin does not abolish HER2 phosphorylation. The objective of our study was to investigate why Herceptin did not decrease HER2 phosphorylation despite being an anti-HER2 monoclonal antibody. We also investigated the effects of acute and chronic Herceptin treatment on HER3 and PKB phosphorylation in HER2-positive breast cancer cells.
We showed that HER2 phosphorylation was maintained by the activation of other HER receptors during Herceptin treatment through an ADAM17-mediated ligand release. Although Herceptin initially decreased HER3 phosphorylation, reactivation of HER3 occurred in prolonged Herceptin treatment through a PKB negative feedback loop. The reactivation of HER3 and failure of Herceptin to abolish HER2 phosphorylation may be responsible for acquired resistance to Herceptin in HER2-overexpressing breast cancer.
We investigated how binding of HER2 receptors by the anti-HER2 monoclonal antibody Herceptin affects HER2 receptors. Although Herceptin was initially thought to inhibit aberrant HER2 receptor tyrosine kinase activity, recent studies have shown that Herceptin does not decrease HER2 phosphorylation [14],[15]. However, the mechanisms of why Herceptin does not inhibit HER2 phosphorylation have not been elucidated. Furthermore, studies that have investigated the effect of Herceptin on HER2 phosphorylation have typically been based on classical Western blot analysis, which cannot detect phosphorylation status in individual cells. We proceeded to monitor the effect of Herceptin on HER2 phosphorylation in HER2-overexpressing cells using classical biochemical methods in combination with an established Förster resonance energy transfer (FRET) methodology that can assess HER2 phosphorylation in individual cells [16].
Using the classical biochemical methods, we confirmed that Herceptin down-regulated HER2 receptors in sensitive SKBR3 cells after 10 d of treatment (Figure 1A). We then assessed the effect of Herceptin on HER2 phosphorylation. Herceptin did not decrease nor abolish HER2 phosphorylation (Figure 1A). Paradoxically, it increased HER2 phosphorylation in SKBR3 cells. However, despite an increase in HER2 phosphorylation, there was a decrease in cell viability in SKBR3 cells after 10 d of Herceptin treatment compared to untreated cells (p = 0.02) (Figure 1B). Since a Western blot is unable to assess HER2 phosphorylation in individual cells or assess heterogeneity between cells, we proceeded to use FRET to assess HER2 phosphorylation in individual cells.
Using an established method to assess HER2 phosphorylation by FRET [16],[17], we conjugated an anti-HER2 antibody to a Cy3b fluorophore (HER2-Cy3b) and an anti-phospho-HER2 antibody to Cy5 (pHER2-Cy5) to assess HER2 phosphorylation in fixed SKBR3 cells with or without 40 µg/ml Herceptin (see Materials and Methods). The median donor lifetime of Cy3b was 2.15 ns (Figure 1C). HER2 phosphorylation would bring the donor and acceptor fluorophores into close proximity, resulting in a decrease of donor lifetime. We first monitored the basal phosphorylation in SKBR3 cells (without Herceptin treatment) and found a decrease in the average lifetime of HER2-Cy3b when pHER2-Cy5 was present (from 2.15 ns to 1.4 ns) (Figure 1C). Following Herceptin treatment, there was considerable heterogeneity between the cells, with suppression of HER2 phosphorylation in a few cells, although the phosphorylation of HER2 was maintained in the majority of cells (Figure 1C). After 10 d of Herceptin treatment, the remaining treated cells still had persistent HER2 phosphorylation (Figure 1C), and this represented approximately 50% of the cell number compared to untreated cells (p = 0.02) (Figure 1B). Herceptin has been shown to target ALDH-positive stem cells in HER2-overexpressing breast cancer cells [18],[19]. We proceeded to show that after 6 d of Herceptin treatment, there was a decreased proportion of cells that were ALDH positive compared to untreated cells, correlated with a decrease in HER2 receptors (Figure S1).
As control, the effect of Herceptin on HER2 phosphorylation in the normal breast epithelial cell line MCF12F was also assessed. Even though acute Herceptin treatment could not inhibit HER2 phosphorylation in SKBR3 cells, it was able to decrease HER2 phosphorylation (shown by increase of lifetime) in MCF12F cells (Figure S2A). The inability of Herceptin to inhibit HER2 phosphorylation in SKBR3 cells was not due to the degradation of Herceptin (Figure S2B).
We also observed similar results in another HER2-overexpressing cell line, BT474 (Figure 1D). These cells were also sensitive to Herceptin treatment after several days of treatment, with decreased cell viability compared to control (Figure S3A, upper panel). As in SKBR3 cells, acute Herceptin exposure did not decrease HER2 phosphorylation in these cells (Figure 1D). HER2 phosphorylation increased in BT474 cells after 1 h of Herceptin treatment (Figure 1D). After treating these cells with Herceptin for 8 mo (with replacement of Herceptin every week), the cells became resistant to 40 µg/ml Herceptin (Figure S3A, lower panel). Herceptin was able to decrease but not eliminate ALDH-positive stem cells in long-term Herceptin-treated BT474 cells compared to untreated cells (Figure S3B). The decrease in ALDH-positive cells correlated with down-regulation of HER2 receptors. However, HER2 phosphorylation and cell viability remained in these resistant cells treated for a prolonged period with Herceptin (Figure 1D).
We found that the down-regulation of HER2 receptors was detectable after 1 h of Herceptin treatment in BT474 cells, and was associated with an increase in HER2 phosphorylation (Figure 1E). Lee-Hoeflich et al. [20] showed that knockdown of HER2 receptors but not EGFR caused a significant decrease of HER3 phosphorylation in HER2-positive breast cell lines. We investigated whether this occurred in our experiments. After 1 h of Herceptin treatment in SKBR3 and BT474 cells, there was a decrease in HER3 phosphorylation correlating with a down-regulation of HER2 receptors (Figure 1F). HER2-overexpressing cells have been shown to constitutively suppress PTEN activity with increased PKB activity, and it has been shown that acute Herceptin exposure decreased PKB phosphorylation through PTEN activation [13]. We found that after 1 h of Herceptin treatment, there was a decrease in PKB phosphorylation, and this correlated with a decrease in HER3 phosphorylation in both SKBR3 and BT474 cells (Figure 1F). Thus, acute Herceptin treatment down-regulated HER2 receptors, resulting in a decrease of HER3 phosphorylation and PKB phosphorylation.
The amplification of HER2 results in constitutive activation of HER2 in a human mammary epithelial cell system [21]. We found that although Herceptin down-regulated HER2 receptors, the remaining cells had persistent and increased HER2 phosphorylation in both SKBR3 and BT474 cells (Figure 1A and 1D). Since HER2 is the preferred dimerisation partner, we postulated that HER2 phosphorylation was maintained by the other HER receptors via their dimerisation with HER2. We proceeded to show, using the streptavidine-biotin immunoprecipitation method (see Materials and Methods), that acute Herceptin treatment increased EGFR/HER2 dimerisation in BT474 cells (Figure 2A, left two panels). This effect was specifically induced by Herceptin, since 1 h of IgG treatment did not increase EGFR/HER2 dimerisation (data not shown). There was also an increase in HER2/HER3 dimerisation in both SKBR3 and BT474 cells after 1 h of Herceptin treatment (Figure 2B, left upper and lower panels). Furthermore, there was increased HER2 dimerisation with the phosphorylated EGFR and HER3 receptors in BT474 cells treated with Herceptin (which was demonstrated using two immunoprecipitation methods; see Materials and Methods) (Figure 2B, middle two upper and lower panels). There was also increased HER2 dimerisation with the phosphorylated HER4 receptor (Figure 2B, right upper and lower panels). Because of a low level of HER4 expression in these cells, the quality of the Western blot was not optimal using the streptavidin-biotin immunoprecipitation method, despite repeated attempts (Figure 2B, right upper panels). However, the quality of the blot was better using the immunoprecipitation method with Herceptin (Figure 2D, right lower panels).
We postulated that the increased dimerisation of EGFR, HER3, and HER4 with HER2 was due to activation by their respective ligands. We proceeded to assess the levels of endogenous ligands, using heregulin (ligand for HER3 and HER4) and betacellulin (ligand for EGFR and HER4) as examples. Herceptin-treated cells were lysed, and endogenous ligand levels were detected using ELISA. We found that Herceptin induced a statistically significant up-regulation of heregulin and betacellulin (p = 0.0152 and p = 0.0286, respectively) after 1 h of Herceptin treatment compared to untreated cells in both SKBR3 and BT474 cells (data on BT474 cells are shown in Figure 2C). There was also increased secretion of these ligands in the conditioned medium of these cells (Figure 2D). Thus, Herceptin increased the dimerisation of EGFR, HER3, and HER4 with HER2 as a result of activation by their ligands.
We showed that Herceptin induced an up-regulation of HER ligands, including betacellulin and heregulin (Figure 2C and 2D). This resulted in an increased phosphorylation of EGFR and HER4 (Figure 2E) and an increase in their dimerisation with HER2 (Figure 2A and 2B). Herceptin, however, decreased HER3 phosphorylation initially after 1 h of Herceptin treatment (Figure 1F). We postulated that the increased heregulin release (Figure 2C and 2D) with Herceptin treatment would have an effect on HER3 phosphorylation. We showed that with prolonged Herceptin treatment, HER3 phosphorylation was reactivated in SKBR3 cells (Figure 2E). Reactivation of HER3 phosphorylation also occurred in BT474 cells that became resistant to Herceptin (Figure 2F). The total expression of HER3 and HER4 increased in SKBR3 cells treated with Herceptin, but the total EGFR expression decreased (Figure 2E). Thus, Herceptin induced ligand activation of EGFR and HER4 as well as reactivation of HER3 phosphorylation during prolonged Herceptin treatment.
We analysed the effects of Herceptin on the downstream signalling pathways in HER2-positive breast cancer cells. It was found that the effects of acute Herceptin treatment on phosphorylation of PKB and ERK1/2 were not concordant (Figures 1F, 2A, and 2E), in contrast to acute tyrosine kinase inhibitor (TKI) treatment, which decreased both PKB and ERK phosphorylation [17]. Acute Herceptin exposure increased ERK phosphorylation (Figure 2A and 2E) but decreased PKB phosphorylation in BT474 and SKBR3 cells (Figure 1F).
Acute Herceptin exposure increased EGFR/HER2 and HER2/HER4 dimerisation, correlating with an increase in ERK phosphorylation (Figure 2A and 2E). In contrast, acute Herceptin treatment decreased PKB phosphorylation (Figures 1F and 2E); this decrease has been shown to be due to activation of PTEN [13], correlating with a decrease in HER3 phosphorylation (Figure 1F). With prolonged Herceptin treatment, reactivation of PKB and HER3 occurred (Figure 2E). The increased ERK phosphorylation was transient (Figure 2A and 2E), mimicking the effect of exogenous ligand stimulation.
Therefore, Herceptin treatment decreased PKB phosphorylation because of a decrease in HER3 phosphorylation induced by HER2 down-regulation. However, 1 h of Herceptin treatment increased ERK phosphorylation as a result of ligand-dependent EGFR and HER4 activation.
To further show that HER ligands play a role in the acquired resistance to Herceptin, we stimulated BT474 cells with 100 ng/ml EGF, heregulin, or betacellulin while they were treated with 40 µg/ml Herceptin. After 5 d, we assessed their cell viability. For Herceptin treatment without exogenous ligands, there was a decreased cell viability of BT474 cells, which was statistically significant (p<0.001) (Figure 2G). However, when Herceptin treatment was given in BT474 cells with concurrent stimulation of exogenous HER ligands, the decrease in cell viability was reversed. The reverse in cell viability in these conditions was statistically significant compared to Herceptin treatment alone (p = 0.0001 for EGF, p = 0.002 for heregulin, p = 0.003 for betacellulin, respectively, compared to Herceptin alone) (Figure 2G).
We investigated the role of ADAM proteases since they mediate shedding of pro-HER ligands including HB-EGF, epiregulin, heregulin, and betacellulin [22]. As ADAM17 is one of the most important ADAM proteases for HER ligands, we studied the role of this ADAM protease in Herceptin treatment. SKBR3 cells were transfected with small interfering RNA (siRNA) against ADAM17, and the knockdown was validated by Western blot. There was a decrease in both pro and active forms of ADAM17 in transfected cells compared to the control (Figure 3A, left panels). We also showed that heregulin production increased in response to acute Herceptin exposure in cells transiently transfected with control siRNA, but this production was inhibited by siRNA against ADAM17 (Figure 3A, right panel). We also assessed the effect of Herceptin on the expression of ADAM17 protease. We demonstrated that after treating the cells with Herceptin for 1 h, ADAM17 protease mRNA was increased by 2.2-fold (n = 4, p = 0.0008) (Figure 3B). To assess whether an increase in mRNA production of ADAM17 was translated to protein, we proceeded to assess ADAM17 protein level in response to Herceptin treatment. We showed that Herceptin increased the protein levels of ADAM17 in a dose-dependent manner after 1 h of treatment in both BT474 (Figure 3C, left two panels) and SKBR3 (Figure 3C, right two panels) cells. Furthermore, the increase of ADAM17 protease was shown to correlate with the suppression of PKB phosphorylation by Herceptin (Figure 3C). As a control, we treated MCF12F cells with Herceptin for comparison. Herceptin did not cause significant pPKB inhibition nor affect ERK phosphorylation in these cells after 1 h of treatment (Figure S2C, right panels). In addition, Herceptin did not induce a significant increase in ADAM17 level nor HER ligand levels (heregulin and betacellulin) in the conditioned medium of the normal epithelial breast cell line MCF12F treated with Herceptin (Figure S2C).
In summary, we showed that ADAM17 is involved in the up-regulation of heregulin in response to Herceptin treatment. We found increased levels of mRNA and protein levels of ADAM17 protease in response to acute Herceptin exposure in HER2-positive cells but not in normal epithelial MCF12F cells. The up-regulation of ADAM17 was shown to correlate with the suppression of PKB phosphorylation in HER2-overexpressing cells.
We observed earlier that the up-regulation of ADAM17 correlated with the suppression of PKB phosphorylation by Herceptin treatment, suggesting the existence of a negative PKB feedback loop involving ADAM17 in acute Herceptin treatment. We hypothesized that if there was a negative PKB feedback loop, a PKB inhibitor should initiate the same response as Herceptin treatment, inducing an up-regulation of HER ligands and ADAM17 levels. To assess the role of a PKB feedback loop induced by Herceptin treatment, we treated BT474 cells with a PKB/Akt inhibitor (Akt inhibitor VIII, Akti-1/2), which can decrease PKB phosphorylation via a mechanism different from that of Herceptin.
Using the quantitative Meso Scale Discovery (MSD) method (see Materials and Methods), we showed that the PKB inhibitor decreased PKB phosphorylation after 1 h of treatment, which was statistically significant in comparison to DMSO control treatment (p<0.0001) (Figure 4A, left panel). Herceptin also decreased PKB phosphorylation in comparison with untreated cells (p = 0.03 compared to untreated) (Figure 4A, left panel). Neither PKB inhibitor nor Herceptin decreased total PKB levels in comparison to control cells treated with IgG or DMSO (Figure 4A, right panel).
We also assessed the effects of the PKB inhibitor on PKB and ERK phosphorylation in BT474 cells using Western blot. As expected, 1 h of treatment with 2.5 µM PKB inhibitor, but not DMSO, decreased PKB phosphorylation. However, it also increased ERK phosphorylation (Figure 4B), just like acute Herceptin treatment (Figure 2A and 2E). More importantly, the decrease in PKB phosphorylation by the PKB inhibitor was also associated with an increase in heregulin production (p = 0.012) (Figure 4C) and an up-regulation of ADAM17 mRNA levels (p = 0.016) (Figure 4D). Thus, the decrease in PKB phosphorylation by a PKB inhibitor initiated the same feedback loop as that seen in Herceptin treatment, which reduces PKB phosphorylation via a different mechanism.
In order to further prove the existence of a PKB negative feedback loop involving ADAM17 during Herceptin treatment, we needed to demonstrate that we can abrogate the loop and suppress HER2 phosphorylation by inhibiting ADAM17.
HER2-overexpressing cells express autocrine ligands, including heregulin, resulting in HER2 activation and basal phosphorylation. We showed that 1 h of treatment with TAPI-1 (an ADAM17 and metalloprotease inhibitor) or with specific ADAM17 inhibitor and ADAM10/17 inhibitor was able to inhibit basal HER2 phosphorylation in SKBR3 cells (Figure 5A). Whereas acute Herceptin exposure increased HER2 phosphorylation in SKBR3 cells (Figure 1A), combination treatment with Herceptin and TAPI-1 decreased HER2 phosphorylation (Figure 5A). We also investigated the effect of the combination of Herceptin and TAPI-1 in individual SKBR3 cells using FRET. There was a basal phosphorylation of HER2 in SKBR3 cells, as shown by a decrease in the average lifetime of HER2-Cy3b with pHER2-Cy5 from about 2.05 ns to 1.6 ns (Figure 5B). Acute Herceptin treatment did not decrease HER2 phosphorylation (Figure 1A and 1C), but with concurrent TAPI-1 inhibitor treatment there was suppression of HER2 phosphorylation (an increase in the average lifetime) (p = 0.008) (Figure 5B).
To further prove the role of ADAM17 in the negative feedback loop, we transiently transfected SKBR3 cells with siRNA against ADAM17. We showed that Herceptin was unable to decrease HER2 phoshorylation in control cells. However, in cells that were transfected with specific siRNA against ADAM17, HER2 phosphorylation was decreased after Herceptin treatment (Figure 5C).
We proceeded to assess the effect of various ADAM17 inhibitors on cell viability with or without concurrent Herceptin treatment. We hypothesized that the combination of Herceptin with ADAM17 inhibitors would exert greater inhibition than Herceptin alone. We showed that combination of Herceptin with either TAPI-1 or specific ADAM17 inhibitor exerted greater inhibition of cell viability in BT474 cells after 2, 3, or 6 d of treatment (Figure 5D).
Thus, our data prove that inhibition of ADAM17 is able to abrogate the feedback loop that maintains HER2 phospshorylation during Herceptin treatment.
We demonstrated earlier that ADAM17 inhibitors were able to abrogate the PKB negative feedback loop and inhibit HER2 phosphorylation during Herceptin treatment. Since an up-regulation of ADAM17 and HER ligands resulted in activation of all HER receptors, we hypothesized that a panHER inhibitor should also be able to reverse the effect of the PKB negative feedback loop induced by Herceptin treatment.
We investigated whether a panHER inhibitor, which inhibits the activation of all HER receptors, could decrease HER2 phosphorylation and be synergistic in tumour growth inhibition with Herceptin treatment. JNJ-26483327 is a potent multi-kinase inhibitor against EGFR (half-maximal inhibitory concentration [IC50] = 9.6 nM), HER2 (IC50 = 18 nM), and HER4 (IC50 = 40.3 nM) [23] (Figure S4A). It is also known as a panHER inhibitor since its IC50 against these receptors is comparable with that of other panHER inhibitors [24]. We found that 1 h of treatment with either 40 µg/ml Herceptin or 10 µM JNJ-26483327 was not able to decrease HER2 phosphorylation in SKBR3 cells (Figure 6A). However, the combination of Herceptin treatment with either 5 µM or 10 µM JNJ-26483327 for 1 h was able to decrease HER2 phosphorylation in these cells (Figure 6A). Furthermore, whereas Herceptin treatment alone increased ADAM17 (p = 0.011) and heregulin (p = 0.008) mRNA levels, neither JNJ-26483327 treatment nor the combined treatment of Herceptin with JNJ-26483327 increased their levels compared to the untreated cells (Figure 6B). Therefore, the combined treatment of a panHER inhibitor with Herceptin could abrogate the PKB feedback loop involving ADAM17 and heregulin. We also showed that the combination of Herceptin and panHER inhibitor JNJ-26483327 exerted greater inhibition of cell viability after 3, 6, or 8 d of treatment compared to Herceptin or JNJ-26483327 alone (Figure 6C). We investigated whether a panHER inhibitor in combination with an ADAM17 inhibitor without Herceptin treatment could exert similar synergistic effect. However, JNJ-26483327 with TAPI-1 exerted less cell viability inhibition than Herceptin alone in both SKBR3 and BT474 cells (Figure S4C). This may be because neither JNJ-26483327 alone nor JNJ-26483327 with TAPI-1 could decrease pHER3 and pAKT after 1 h of treatment, in contrast to Herceptin or Herceptin with JNJ-26483327 (Figure S4B).
To test the in vivo relevance of the interaction between a panHER inhibitor and Herceptin, BT474 xenografts were treated with an empty vehicle, Herceptin alone, the panHER inhibitor JNJ-26483327 alone, or the combination of drugs for 21 d (Figure 6D). As seen in Figure 6D, Herceptin alone or the panHER inhibitor alone could only delay xenograft tumour growth compared to vehicle treatment, but the combination of the two drugs caused an almost complete inhibition of tumour growth in these HER2-positive xenografts.
In summary, Herceptin in combination with a panHER inhibitor was able to decrease HER2 phosphorylation and abrogated the up-regulation of ADAM17 and heregulin in response to Herceptin treatment. Whereas either a panHER inhibitor or Herceptin treatment alone delayed BT474 xenograft tumour growth, the combination treatment was synergistic in tumour inhibition.
It was previously thought that Herceptin inhibits HER2 receptor tyrosine kinase activity, but recent studies have shown that this is not the case [14],[15]. The mechanisms whereby Herceptin fails to decrease HER2 phosphorylation remain unclear. Our results confirmed that Herceptin did not decrease HER2 phosphorylation although it down-regulated HER2 receptors in HER2-positive SKBR3 and BT474 breast cell lines. We showed that HER2 phosphorylation was maintained and increased by the ligand-induced activation of EGFR, HER3, and HER4 receptors, which preferentially dimerise with HER2. This is consistent with reports that Herceptin does not prevent the dimerisation of HER2 with other receptors [25].
It was previously demonstrated that primitive mammary stem cells are enriched in vitro in non-adhering spherical colonies called mammospheres [26]. These cells have stem-cell-like properties, with the ability to undergo multi-lineage differentiation [26]. The proportion of these stem cells in normal mammary epithelial cells is increased by HER2 overexpression, as demonstrated by in vitro mammosphere assays and the expression of the stem cell marker ALDH [18]. One of the clinical benefits of Herceptin is thought to be its ability to target the cancer stem cell population in HER2-amplified tumours [19]. We confirmed in our study that Herceptin decreased ALDH-positive stem cells after a prolonged treatment, in correlation with a decrease in HER2 receptors. However, there was significant heterogeneity in the inhibition of HER2 phosphorylation by Herceptin between different cells, especially during the first week of treatment. Thus, Herceptin was able to decrease HER2-mediated signalling in some of these cells, resulting in decreased HER2 receptors and phosphorylation, but Herceptin monotherapy could not eliminate all the stem cells. The surviving cells had decreased HER2 receptors and fewer ALDH-positive cells compared to untreated cells but had maintained HER2 phosphorylation via activation of other HER receptors as a result of ADAM17-mediated ligand release.
HER3 is kinase-defective, and its phosphorylation depends on other HER receptors. The effects of Herceptin on HER3 phosphorylation have been controversial. Yakes et al. [27] showed that 1 h of Herceptin treatment induced a transient increase of pHER3 in BT474 cells, whereas Junttila et al. [15] reported a decrease of pHER3 with acute Herceptin treatment. We showed that acute Herceptin treatment initially decreased HER3 phosphorylation. This decrease is thought to be due to HER2 down-regulation, since loss of HER2 induced by siRNA decreased pHER3 levels in HER2-positive breast cancer cells [20]. It is possible that the difference in observed pHER3 is due to several factors that are competing to affect pHER3 levels. The dominant effect of acute Herceptin treatment is a decrease in HER2 levels, but this is in competition with the increased HER3 phosphorylation as a result of increased ligand production induced by Herceptin treatment. This would account for the subsequent reactivation of HER3 with prolonged Herceptin treatment. Furthermore, variable cell lines and experimental conditions, as well as different treatment durations and doses of Herceptin used by different investigators, may account for differences in pHER3 levels seen [15],[27].
We found that Herceptin had a discordant effect on PKB and ERK signalling. HER2-overexpressing cells have been shown to activate Src with constitutively suppressed PTEN activity and increased PKB activity [13],[28]. Herceptin increased PTEN membrane localisation and phosphatase activity by reducing PTEN tyrosine phosphorylation via Src inhibition, leading to a decreased PKB phosphorylation [13]. The decrease in PKB activity is not due to PI3K inhibition, since there was no decreased PI3K activity after Herceptin treatment [13]. However, acute Herceptin activates ERK1/2 pathways, correlating with an increase in EGFR and HER4 dimerisation with HER2. Since EGFR, HER3, and HER4 have binding sites for Shc and Grb2 [29], their ligand-dependent activation would account for increased activation of ERK by acute Herceptin treatment. The sudden increase in ERK phosphorylation induced by Herceptin treatment is very much similar to MAPK/ERK activation in cells stimulated with exogenous HER ligands [30]. This observation supports our data that Herceptin induces the up-regulation of HER ligands through ADAM proteases.
Sergina et al. [31] showed that TKI treatment failed to suppress HER3 phosphorylation for a sustained duration because of a PKB-mediated feedback loop. However, they did not link the PKB-mediated feedback loop with the HER ligands and ADAM proteases. We have previously shown that reactivation of HER3 and PKB in response to TKI is due to HER ligand release [16]. We have now shown that acute Herceptin treatment decreases the phosphorylation of HER3 and PKB, which in turn induces the activation of a feedback loop involving HER ligands and ADAM proteases. We postulated that Herceptin treatment initiated a PKB-mediated negative feedback loop. If such a negative loop exists, we predicted that an inhibitor that decreases PKB phosphorylation should also induce the up-regulation of HER ligands and ADAM17 protease. Indeed, we demonstrated that a PKB inhibitor, which decreases PKB phosphorylation via a mechanism different from that of Herceptin, could also initiate the same feedback loop induced by Herceptin treatment. Thus, it is a Herceptin-induced decrease in PKB phosphorylation that results in the activation of a feedback loop involving ADAM proteases and HER ligands. This PKB feedback loop activates the other HER receptors and maintains HER2 phosphorylation, which is a key to acquired resistance to Herceptin (Figure 7).
There have been several reports of positive and negative feedback loops linking the complex cross talk of MAPK and PI3K signalling pathways with scaffolding protein Grb2-associated binders 1 and 2 (Gab1 and Gab2) [30],[32]. mTOR inhibition has also been shown to lead to MAPK activation through a PI3K-dependent feedback loop in human cancer [33]. The exact mechanism of how these feedback loops link to each other is likely to be very complex. It is likely that FoxO proteins, which are downstream targets of PKB, are central players in this PKB feedback loop [34]. The phosphorylation of FoxO transcription factors by PKB promotes the cytoplasmic sequestration of these transcription factors, including FoxO3a [34]. It is likely that FoxO transcriptional factors can modify Akt/PKB phosphorylation indirectly by modifying the expression of kinases or phosphatases [35]. It is also possible that FoxO proteins regulate the transcription of ADAM17, since PKB inhibition increases mRNA production of ADAM17. The interactions of FoxO transcriptional factors with ADAM proteases and phosphatase, as well as how they affect the phosphorylation and dephosphorylation of PKB and HER receptors, are likely to be very complicated. The mechanisms are currently being investigated in our lab.
Our xenograft experiment showed that neither panHER inhibitor nor Herceptin treatment alone was adequate to control tumour growth in a HER2-oncogene-driven tumour. In HER2-positive breast cancer, the underlying problem is HER2 overexpression, which results in increased HER2-related signalling. Herceptin is able to down-regulate HER2 receptors, whereas TKI-like lapatinib induces HER2 accumulation at the cell surface [14]. We have shown that a panHER inhibitor with TAPI-1 (ADAM17 inhibitor) resulted in less inhibition of cell viability than Herceptin alone in both SKBR3 and BT474 cells. This confirms that TKI treatment without Herceptin is not as effective as treatment with either Herceptin or Herceptin with TKI in HER2-positive breast cancer cells. However, the combination of Herceptin with a panHER inhibitor, which inhibits the activation of all HER receptors, was able to abrogate the feedback loop during Herceptin treatment and was synergistic in tumour inhibition in a HER2-positive BT474 xenograft model. Our data would support the rationale of combining a panHER inhibitor with Herceptin treatment in patients with HER2-positive breast cancer. Our results would also explain why pertuzumab and Herceptin are synergistic in tumour inhibition in breast cancer cells and xenograft models [20]. Furthermore, it would account for the effectiveness of a combination treatment of pertuzumab with Herceptin in some patients whose disease progressed while on Herceptin alone [36].
Our results demonstrate why it is inadequate to consider individual HER receptors alone for anti-HER therapies as these receptors are intrinsically linked together with a close network of feedback loop(s). We have demonstrated that a PKB feedback loop is activated upon Herceptin treatment, resulting in the activation of all HER receptors and maintenance of HER2 phosphorylation. We have not attempted to look at all the positive and negative feedback loops linking MAPK and PI3K [30],[32]. It is likely that more feedback loops are involved in the acquired resistance to Herceptin. Future research should identify the exact candidates, other than ADAM17 and HER ligands, involved in the feedback loops. Such candidates are best identified through a systems biology approach, which may help to further dissect the mechanisms of acquired resistance to Herceptin in HER2-overexpressing tumours and to improve the survival for patients with this type of tumour.
SKBR3 and BT474 cells were obtained from cell services at Cancer Research UK (Lincoln's Inn Fields laboratory). The human cell lines BT474 and SKBR3 are HER2-overexpressing breast cancer cell lines. SKBR3 cells were cultured in Dulbecco's Modified Eagle Medium (DMEM) supplemented with 10% fetal bovine serum and the antibiotic penicillin-streptomycin. BT474 cells were cultured in RMPI supplemented with 10% fetal bovine serum and the antibiotics penicillin-streptomycin. For these cells, 10 µg/ml insulin was added to the medium when cells were split or medium was refreshed. Normal breast epithelial MCF12F cells were purchased from the ATCC and cultured in a 1∶1 mixture of DMEM and Ham's F12 medium supplemented with 20 ng/ml EGF, 100 ng/ml cholera toxin, 0.001 mg/ml insulin, 500 ng/ml hydrocortisone, and 5% horse serum.
Anti-HER2 (recognising the intracellular residues), anti-phospho-HER2 (Tyr1221/1222), anti-phospho-HER3 (Tyr1289), anti-HER4 (recognising the intracellular residues near the carboxyl-terminus of human HER4), and anti-phospho HER4 (Tyr1284) antibodies were obtained from Cell Signalling Technology. The polyclonal anti-phospho-EGFR (Thr992) was obtained from Invitrogen. F4-IgG1 mouse monoclonal antibody (against the EGFR cytoplasmic domain) was obtained from the monoclonal antibody laboratory of Cancer Research UK (Lincoln's Inn Fields laboratory). Antibodies recognising PKB, phospho-PKB (Ser473), p44/42 MAP kinase (Erk1/Erk2), and phospho-Erk1/Erk2 (Thr202/Tyr204) were from Cell Signalling Technology. Anti-ADAM17 was purchased from Abcam. The secondary antibodies, goat anti-mouse IgG and goat anti-rabbit IgG, were purchased from GE Healthcare. The mono-conjugated fluorophores Cy3B and Cy5 were from GE Healthcare. PKB inhibitor (Akt inhibitor VIII, isozyme-selective, Akti-1/2) was obtained from Calbiochem. Herceptin was initially a gift from Roche, but subsequent supply was obtained from the pharmacy department of Oxford Radcliffe Hospitals, National Health Services Trust. Human IgG control was purchased from R&D Systems. EGF, heregulin, and betacellulin were purchased from Sigma Aldrich. ADAM and metalloprotease inhibitor (TAPI-1) was purchased from Calbiochem. Incyte kindly provided ADAM17 inhibitor INCB4298 and ADAM10/17 inhibitor INCB3619. Janssen (Johnson & Johnson) kindly provided panHER inhibitor JNJ-26483327.
For Western blotting confluent six-well plates of cells were placed on ice and washed with PBS. Cells were scraped off the plates and incubated for 10 min in lysis buffer (10 mM EDTA, 20 mM Tris [pH 7.5], 150 mM NaCl, 10 mM Na2P2O7, and 100 mM NaF with 1% Triton and protease inhibitor cocktail [Roche]). Samples were centrifuged at 4°C to remove the insoluble cell pellets, and a protein assay was performed to check protein quantity. Equal amounts of protein sample were prepared in 4× SDS with 10% beta-mercaptoethanol and boiled for 10 min at 95°C. Then samples were loaded on a NuPage 4%–12% gel (Invitrogen) and run at 130 V. The proteins were semi-dry-transferred to a membrane for 2 h at 12 V. The membrane was blocked in 3% BSA in PBS-Tween (0.2%) for a minimum of 1 h. Then the blot was incubated with primary antibody in the same solution for 3 h at room temperature. The membrane was washed four times with 1% milk in PBS-Tween (0.2%) before secondary antibody was added in 5% milk in PBS-Tween (0.2%). The membrane was incubated at room temperature for 1 h before it was washed four times again with 1% milk in PBS-Tween (0.2%). Antibodies were visualised with an enhanced chemiluminescent (ECL) system (GE Healthcare).
BT474 and SKBR3 cells were grown to near confluency before they were lysed as described above. The cell lysate was centrifuged for 10 min at maximum speed before transferring the supernatant to a new reaction vial. A protein assay was performed to check protein quantity, and equal amounts of protein were used for immunoprecipitation. In order to look at the interaction between HER2 and other HER receptors after Herceptin treatment, we needed a technique that would specifically pull down our protein of interest. We could not use magnetic protein G beads because they would bind Herceptin as well as our receptor-specific antibody (data not shown). Therefore, streptavidin-coated magnetic beads (Bio-Nobile) were absorbed with biotin-conjugated HER antibodies (1∶100) (conjugated using kit from Innova Biosciences) for 1 h, as illustrated in the figures. After the bead-antibody complex was washed with PBS-Tween (0.2%), it was incubated with the supernatant for at least 1 h. After that, the complex was thoroughly washed with PBS-Tween and transferred to a new reaction vial. Wash buffer was taken off and 50 µl of 4× SDS with 10% beta-mercaptoethanol was added to the beads. Samples were boiled for 10 min at 95°C to elude protein off the beads. Twenty microlitres was loaded per lane in a SDS gel for Western blot analysis, as described above. For further confirmation, we also used protein G beads labelled with Herceptin (anti-HER2) to pull down HER2, and looked at the levels of phosphorylated HER receptors.
SKBR3 and BT474 cells were grown in 24-well plates after seeding approximately 20,000 cells per well. The cells were grown for at least 24 h before treatment with 40 µg/ml Herceptin, 5 µM JNJ-26483327, 10 µM TAPI-1, 10 µM ADAM17 inhibitor, or a combination of these drugs for different durations, as illustrated in the figures. For the exogenous ligand experiments, 100 ng/ml EGF, heregulin, or betacellulin was added to the cells in addition to Herceptin (40 µg/ml) for a total of 5 d in BT474 cells. On the day of the experiment, the cells were trypsinized and diluted with PBS. The viable cells were counted using a cell counter.
For all transient transfections, the cells were plated in 10-cm2 dishes (50% confluent) in medium and given the opportunity to settle overnight. The next day, medium was replaced by 7 ml of fresh normal medium. Transfection mix containing 10 nM siRNA in 300 µl of OptiMEM (Invitrogen) was incubated with 15 µl of NeoFX (Ambion) in 300 µl of OptiMEM for 15 min at room temperature. This mix was added drop-wise to the cells, and cells were placed back into the incubator. After 24 h, transfected cells were re-plated into six-well plates for further experiments. Validated siRNAs were obtained from Ambion.
Cells treated with Herceptin (40 µg/ml), PKB inhibitor (2.5 µM), or panHER inhibitor JNJ-26483327 (5 µM) were analysed for mRNA levels of ADAM17 and heregulin. RNA was purified from cells using the Aurum Total RNA mini kit (Bio-Rad) according to manufacturer's instructions. RNA purity and quantity were determined using Nanodrop (Nanodrop Technologies). For synthesis of the cDNA, a high-capacity cDNA reverse transcription kit was used (Qiagen).
Quantitative PCR (QPCR) reactions were performed using the following primers, together with FAM-labelled probes from the Universal ProbeLibrary (Roche) or Sybergeen: Beta-actin primers, 5′-ATTGGCAATGAGCGGTTC-3′ and 5′-GGATGCCACAGGACTCCAT-3′, and universal probe 11; heregulin-β1 primers, 5′- CTTGTGGTCGGCATCATGT-3′ and 5′-CAGCTTTTTCCGCTGTTTCT-′3, and probe 49; ADAM17 primers, 5′-CCTTTCTGCGAGAGGGAAC-3′ and 5′- CACCTTGCAGGAGTTGTCAG-3′ and probe 78.
cDNA samples were assayed in triplicate using a detection system (Chromo4; GRI), and gene expression levels for each individual sample were normalised relative to Beta-actin. Levels of mRNA in untreated samples were set to 1.
We used immunoprecipitation to look at the levels of ligands (betacellulin and heregulin) in the media of untreated or Herceptin-treated cells. Cells were treated with Herceptin in serum-free medium for 1 h. The medium was then collected and incubated with magnetic protein G beads (Invitrogen) linked to antibodies for heregulin or betacellulin for at least 1 h. The samples were prepared for Western blot analysis as described above.
Cells treated with Herceptin (40 µg/ml) or PKB inhibitor (2.5 µM) were analysed for HER ligands. Levels of the HER ligands betacellulin and heregulin were measured in cell lysate using ELISA (R&D Systems). Cells were lysed in 20 mM Hepes buffer containing 1.5 mM EDTA and protease inhibitor (Roche). Cells were then scraped off and homogenised by putting the lysate through a syringe with needle. After spinning the lysate down, the supernatant was used for ELISA analysis.
To detect heregulin or betacellulin, the R&D Systems ELISA kits were used as the protocol prescribes. Briefly, the ELISA plate was coated overnight with coating buffer. After blocking the plate for 2 h, 100-µl samples and standards were loaded and the plate was kept at room temperature for 2 h. Samples were washed and incubated with detection antibody for 2 h, then washed and incubated with streptavidin labelled with horseradish peroxidase for 20 min. With the use of a substrate solution that reacts with horseradish peroxidase, the levels of ligands could be detected. After 20 min, the reaction was stopped using stop solution, and absorption was measured using an ELISA reader.
BT474 cells were plated on a six-well plate and left to settle overnight. The next day, the cells were treated for 1 h with 40 µg/ml Herceptin or 2.5 µM PKB inhibitor at 37°C. After 1 h, the cells were lysed in a lysis buffer provided by MSD. The assay was then performed following manufacturer's protocol. In brief, MSD provided the plates pre-coated with pPKB (Ser473) and PKB spots. Multiple spots could be placed in one well of a 96-well plate, making it possible to look at several proteins in one sample. The plates provided were blocked for 1 h with 3% BSA at room temperature. The plates were then washed three times with wash buffer before 25 µl (20 µg) of sample was loaded onto the plates (in duplicate). After incubation for 2 h at room temperature, the plates were washed again, and 25 µl of detection antibody in 1% BSA was added. After 1 h of incubation, 150 µl of read buffer was added, and the plate was analyzed using a SECTOR imager (MSD).
FRET experiments were performed as described previously [16],[17]. In short, 30,000 cells were seeded onto cover slips and left to attach overnight. The next day, the cells were treated with 40 µg/ml Herceptin or the ADAM inhibitor TAPI-1 or the combination of these drugs (for different durations as illustrated) in the medium at 37°C. The medium was washed off with PBS, and the cells were fixed with 4% paraformaldehyde (Pierce) in PBS for 10 min at room temperature. The cells were then permeabilized with 0.2% Triton X-100 (T8532, Sigma-Aldrich) in PBS for 5 min at room temperature before being treated for 10 min with 1 mg/mg sodium borohydrate (Sigma-Aldrich) in PBS to quench the background fluorescence. The cells were washed twice with PBS between steps. Following the above steps, the cells were treated for 1 h with in 1% BSA (Sigma-Aldrich) in PBS at room temperature to block unspecific binding of the antibodies. After that, antibodies conjugated to the fluorescent dyes Cy3b or Cy5 were added to the samples sequentially starting with the Cy3b-conjugated antibody. Cells were washed twice with PBS and twice with sterile water, after which they were mounted onto a microscope slide using Fluoromount-G (Southern Biotech). All images were taken using a Zeiss Plan-APOCHROMAT 6100/1.4 NA phase three-oil objective. Images were recorded at a modulation frequency of 80 MHz. The donor (Cy3b) was excited using the 514-nm line of an argon/krypton laser, and the resultant fluorescence was separated using a combination of dichroic beam splitter (Q565 LP, Chroma Technology) and narrow band emitter filter (BP 610/75, Lys and Optik).
To detect ALDH activity in cells, we used the ALDEFLUOR kit from Aldagen. The manufacturer's protocol was followed carefully. In brief, cells were trypsinized and, per treatment protocol, 1×106 cells were taken up into 1 ml of assay buffer provided in the kit. This tube was the sample tube. To a separate control tube, 5 µl of DEAB solution was added. ALDEFLUOR substrate (5 µl) was added to the sample tube, and immediately after mixing, 500 µl of the sample solution was transferred to the control tube with DEAB. The procedure was repeated for all samples. Both control and sample tubes were then placed at 37°C for 45 min. After the cells were spun down for 5 min at 250 g, they were taken up in new assay buffer and analyzed using a FACS Cyan. Control samples were used to set a gate for ALDH positivity, after which the test sample was analysed.
The Mann-Whitney test was used to compare the medians of the protein levels of heregulin and betacellulin as well as mRNA levels of heregulin and ADAM17 between the untreated samples and those treated with drugs. For each figure, at least three experiments were done, and both technical and biological replicates were used in the calculation, using a confidence interval of 95%. Data were analysed against the untreated samples unless otherwise stated.
|
10.1371/journal.pgen.1007294 | Phosphorelay through the bifunctional phosphotransferase PhyT controls the general stress response in an alphaproteobacterium | Two-component systems constitute phosphotransfer signaling pathways and enable adaptation to environmental changes, an essential feature for bacterial survival. The general stress response (GSR) in the plant-protecting alphaproteobacterium Sphingomonas melonis Fr1 involves a two-component system consisting of multiple stress-sensing histidine kinases (Paks) and the response regulator PhyR; PhyR in turn regulates the alternative sigma factor EcfG, which controls expression of the GSR regulon. While Paks had been shown to phosphorylate PhyR in vitro, it remained unclear if and under which conditions direct phosphorylation happens in the cell, as Paks also phosphorylate the single domain response regulator SdrG, an essential yet enigmatic component of the GSR signaling pathway. Here, we analyze the role of SdrG and investigate an alternative function of the membrane-bound PhyP (here re-designated PhyT), previously assumed to act as a PhyR phosphatase. In vitro assays show that PhyT transfers a phosphoryl group from SdrG to PhyR via phosphoryl transfer on a conserved His residue. This finding, as well as complementary GSR reporter assays, indicate the participation of SdrG and PhyT in a Pak-SdrG-PhyT-PhyR phosphorelay. Furthermore, we demonstrate complex formation between PhyT and PhyR. This finding is substantiated by PhyT-dependent membrane association of PhyR in unstressed cells, while the response regulator is released from the membrane upon stress induction. Our data support a model in which PhyT sequesters PhyR, thereby favoring Pak-dependent phosphorylation of SdrG. In addition, PhyT assumes the role of the SdrG-phosphotransferase to activate PhyR. Our results place SdrG into the GSR signaling cascade and uncover a dual role of PhyT in the GSR.
| The general stress response (GSR) in alphaproteobacteria represents an essential feature for survival in stressful, constantly changing habitats. A variety of stresses are sensed by GSR-activating histidine kinases and lead to multiple stress resistance via the response regulator PhyR. Here, we describe the essential bifunctional regulator PhyT, which tunes GSR activation in the plant-protecting strain Sphingomonas melonis. It prevents lethal over activation of the GSR by impeding inappropriate phosphorylation of the response regulator PhyR via protein-protein interaction. In addition, PhyT acts as a phosphotransferase in a GSR activating phosphorelay. Our data suggest a model according to which histidine kinases are induced by environmental cues, which results in phosphorylation of the single domain response regulator SdrG. The phosphoryl group of SdrG is then transferred by the phosphotransferase PhyT to the master regulator of the general stress response PhyR. Histidine kinase derived PhyT-type regulators are found also in other alphaproteobacteria, implying that the identified regulatory strategy might be conserved.
| Two-component regulatory pathways enable bacteria to react to changing environmental conditions. Classical two-component systems consist of a sensor histidine kinase and a response regulator; the histidine kinase autophosphorylates upon sensing appropriate environmental signals and subsequently transfers the phosphoryl group to the response regulator, which in turn triggers an adaptation response [1–4]. Multicomponent phosphorelays represent more complex two-component systems involving either a so-called hybrid histidine kinase or, alternatively, a single domain response regulator (SDRR) which can be phosphorylated by multiple histidine kinases. Either way, a phosphotransferase subsequently transfers the phosphoryl group to an output response regulator. Therefore, by increasing the number of checkpoints in a phosphorylation pathway, phosphorelays allow for more precise regulation, e.g. [1–8].
The general stress response (GSR) is pivotal for alphaproteobacteria for environmental adaption and host microbe interactions [9, 10]. Notably, it connects a two-component system to alternative sigma-factor regulation [11, 12]. The GSR can be induced by a variety of different stresses and results in multiple stress resistance [9, 10]. Studied systems involve for example plant-associated bacteria [13] such as Sphingomonas melonis Fr1 [14], Methylobacterium extorquens [15], Bradyrhizobium diazoefficiens [16], Sinorhizobium meliloti [17], intracellular pathogens like Brucella abortus [11], or free-living species like Caulobacter crescentus [18]. The anti-sigma factor antagonist PhyR is phosphorylated under stressful conditions [14, 19] and acts via a mechanism termed "sigma factor mimicry" [20, 21]. A phosphorylation-induced conformational change of PhyR results in the release of its sigma-factor like domain that subsequently binds to the anti-sigma factor NepR. Thereby, NepR liberates the alternative sigma-factor EcfG, which can then bind to RNA polymerase and re-direct transcription towards the GSR regulon [20, 21]. In the plant-protecting S. melonis Fr1 [22, 23], seven cytoplasmic histidine kinases (Paks) are involved in the GSR [19]. They belong to the HWE/HisKA_2 family [19, 24], encoding the HRxxN motif in their DHp domain. Paks integrate multiple stress signals [19]. Individual stresses are sensed by one or more Paks, and some of the Paks are able to sense more than one stress. In vitro, the Paks do not only phosphorylate PhyR (in the presence of NepR), but also the SDRR SdrG [19].
The dual specificity of the Paks for PhyR and SdrG is supported at the structural level; in fact, the receiver domain of PhyR is the best structural homolog of SdrG [25]. Like Sma0144 from S. meliloti [26], SdrG belongs to the FAT GUY family of response regulators [25]. The SDRR is a key regulator of the GSR in S. melonis Fr1 [19] and M. extorquens [27]; however, its function can be bypassed by overexpression of either PhyR or the Paks [19]. The exact role of SdrG in GSR remains enigmatic. Potential mechanisms of SdrG activity might involve protein-protein interaction, comparable e.g. to CheY from Escherichia coli, which is involved in chemotaxis regulation [28, 29] or to DivK from C. crescentus, in which the regulator is involved in cell cycle control [30]. Alternatively, it is conceivable that SdrG participates in phosphotransfer, e.g. as the SDRR Spo0F, which is involved in sporulation initiation in Bacillus subtilis [5]; however, no phosphotransferase corresponding to Spo0B of B. subtilis has been identified for SdrG.
While in S. melonis Fr1 the genes encoding SdrG and the Paks are scattered throughout the genome, PhyP, an additional regulator of the GSR, is encoded at the phyR locus. The gene encodes a predicted membrane-anchored histidine kinase with a putative periplasmic sensor domain [14]. Closer inspection revealed that PhyP contains the HRxxN motif in its DHp domain, but harbors a degenerated ATPase domain [9, 14]. Due to its observed ability to disrupt the PhyR/NepR complex in vitro in a phosphorylatable His-341-dependent fashion, it has been proposed that PhyP acts as a PhyR phosphatase [14]. A phyP knockout was only obtained in GSR impaired strains, that is, in phyR, ecfG, or multiple pak knockout mutants, confirming PhyP as an essential negative regulator of the GSR in S. melonis Fr1 [14, 19].
The identification of PhyP-type proteins in other alphaproteobacteria, e.g. Bab1_1673 in B. abortus [11], RpaI_4707 in Rhodopseudomonas palustris TIE-1 [31], RsiC in S. meliloti [17], and LovK in C. crescentus [18] suggests a conserved key role of the protein in the GSR.
Here, we investigate the role of PhyP (re-designated PhyT in this study, see below) in S. melonis Fr1 and identify a direct molecular link to SdrG. Our results support a GSR regulatory model in which PhyT acts as a phosphotransferase mediating phosphotransfer between Pak-phosphorylated SdrG and PhyR. Furthermore, our results suggest that PhyT simultaneously prevents lethal over activation of the GSR by inhibiting direct PhyR phosphorylation by Paks via complex formation. The dual function of PhyT explains its essentiality and assigns a key function to SdrG.
In S. melonis Fr1, the phyR locus encodes the membrane-anchored protein PhyP, previously proposed to be a PhyR phosphatase, which is essential in strains with a functional GSR [14]. Here, we studied the function of this central regulator further. First, we set out to test in vitro PhyR dephosphorylation by PhyP using PhyR 32P-labeled at the Asp-194 residue [20, 21], which was generated in presence of NepR using one of the PhyR-activating kinases (here PakF) [19]. For dephosphorylation, we used E. coli membrane particles containing heterologously produced PhyP or the PhyP (H341A) derivative as a control [14]. However, no phosphatase activity was detected (S1 Fig).
We tested an alternative hypothesis regarding the function of PhyP that integrates the central, but not yet understood, role of the positive GSR regulator SdrG [19]. We speculated that PhyP could participate in the GSR-activating phosphotransfer and would assume the role of a PhyR-activating phosphotransferase rather than that of a PhyR phosphatase. According to our working model, Paks act as the primary phosphoryl group sources for SdrG [19]. The phosphoryl group could then be shuttled to PhyP and subsequently transferred to PhyR. To test such a putative phosphorelay, we performed time course in vitro phosphotransfer assays. To guarantee SdrG~P as the sole phosphodonor in the reaction mixture, we used an on-column phosphorylation approach, in which SdrG was phosphorylated by Ni-NTA-bound 32P-labeled PakF. The addition of SdrG~P to E. coli membrane particles harboring heterologously produced PhyP resulted in increasing PhyP phosphorylation and a concomitant decrease of SdrG~P (Fig 1). When SdrG~P was added to a mixture of PhyR and PhyP, SdrG~P dephosphorylation could be observed over time, while the level of PhyR~P increased. This transfer was enhanced in presence of the anti-sigma factor NepR, which is required for efficient direct phosphotransfer from Paks to PhyR in vitro [19]. No phosphotransfer from SdrG to PhyR was detectable in the absence of PhyP. Moreover, no decrease in SdrG~P and thus no phosphotransfer was observed for the inactive PhyP (H341A) derivative (Fig 1). Our in vitro data thus indicate that PhyP acts as a phosphotransferase that shuttles phosphoryl groups from SdrG~P to PhyR. Therefore, we have chosen to rename PhyP as PhyT and refer to it accordingly from now on.
Next, we aimed to confirm that SdrG takes part in PhyR phosphorylation in presence of PhyT in vivo by testing whether the phosphorylatable Asp residue of SdrG is required to fulfill its role as a positive GSR regulator. Therefore, we performed EcfG-dependent β-galactosidase assays and compared sdrG knockout strains producing either wild-type SdrG or the SdrG (D56E) derivative encoding a Glu substitution of the phosphorylatable Asp-56 [25], a substitution that renders many response regulators active by mimicking the phosphorylated state [32–34]. Only wild-type SdrG rescued the impaired GSR observed for the sdrG knockout mutant. The SdrG (D56A) derivative was used as a negative control (Fig 2). This observation confirms previous work that demonstrated the inability of the SdrG (D56E) derivative to complement the salt-sensitive phenotype of an sdrG knockout strain [19] and suggests that SdrG phosphorylation is essential for its positive regulatory function in GSR.
Taken together, our data indicate that PhyT acts as a phosphotransferase in the Pak-SdrG-PhyT-PhyR phosphorelay of the GSR activating pathway in S. melonis Fr1.
The PhyT- and SdrG-dependent PhyR phosphorylation shown above implies that the Paks' main function in vivo is to phosphorylate SdrG rather than PhyR, although Paks phosphorylate PhyR in vitro in the presence of NepR [19]. To further investigate PhyT- and SdrG-dependent PhyR phosphorylation in stress-induced GSR activation in vivo, we determined EcfG activity using β-galactosidase reporter assays. We analyzed a pakB-G deletion mutant (thus leaving pakA intact), with and without an additional sdrG knockout pre- and post-induction with a chemical stress mixture. Although we observed residual GSR in the absence of SdrG, no increased GSR activation could be observed upon addition of the chemical stress mixture (Fig 3). Next, we tested GSR induction upon overexpression of pakA in an sdrG knockout background and a pakA-G deletion mutant with and without an additional sdrG knockout. We observed that overexpression of pakA bypasses the sdrG knockout (Fig 3), which is congruent with previous work [19]. This implies that PhyT- and SdrG-independent PhyR phosphorylation by the Paks is possible in vivo, but plays only a minor role under physiological expression levels of the paks. Additionally, considering that an sdrG knockout may lead to an artificial increase of direct PhyR phosphorylation by Paks, the observed residual GSR activation is likely to be even lower in wild-type cells.
We also tested stress-dependent GSR induction in a pakA-G deletion mutant. In this strain, the GSR was still inducible upon stress application (Fig 3), which points to an additional kinase present in S. melonis Fr1 leading to PhyR phosphorylation. In the light of this finding, we performed EcfG-dependent β-galactosidase assays emphasizing the dependency of GSR induction on both SdrG and PhyT. In a pakA-G deletion mutant, an additional phyT knockout abolished GSR induction under stress conditions, indicating that the so far unidentified kinase strictly depends on the phosphorelay involving PhyT and SdrG (S3A Fig). GSR inducibility was rescued upon overexpression of phyT in this background. However, overexpression of phyT did not enable GSR induction under stress conditions in a pakA-G deletion mutant when an additional sdrG knockout was introduced (S3A Fig). Importantly, overexpression of sdrG in a pakA-G deletion mutant with an additional phyT knockout did not increase GSR (S3B Fig), which is in-line with the dependency of SdrG and PhyT on each other and their positive regulatory role in GSR activation.
Because PhyT functions as a phosphotransferase (Fig 1), one would predict that it plays a role as a positive regulator of the GSR, which is puzzling in light of its described function as a negative regulator [14]. The latter function was deduced from the ability of PhyT to disrupt the NepR/PhyR complex in vitro—interpreted as phosphatase activity—and on the finding that its knockout is only possible in a GSR-impaired background, implying a lethal over activation of GSR in the absence of PhyT [14]. Based on these findings, we aimed to further characterize the negative regulatory role of PhyT. First, we performed EcfG-dependent β-galactosidase assays using a pakB-G deletion mutant with only pakA present. The knockout of phyT in this background resulted in increased GSR pre- and post-induction with a chemical stress mixture (Fig 4), which was caused by direct PakA-dependent PhyR phosphorylation. Overexpression of phyT complemented the observed phenotype (Fig 4). This result confirms that PhyT performs an additional negative GSR regulatory function in vivo.
We speculated that PhyT controls the amount of phosphorylated PhyR by binding the response regulator and thereby preventing its direct phosphorylation by Paks. To test this working model, we first performed a bacterial two-hybrid (BACTH) assay. We used wild-type PhyR and a derivative in which the phosphorylatable Asp-194 residue of the receiver domain was mutated. In addition, we also examined a PhyR Glu-235 mutant, which contains a mutation located at the interface of the sigma factor-like and the receiver domain and results in a constitutively active PhyR [20]. Binding of the response regulator derivatives to the anti-sigma factor NepR was validated as controls. We showed that, indeed, PhyR and PhyT interact in the BACTH assay and that PhyT forms a dimer using plate assays (Fig 5A and S4A Fig) as well as β-galactosidase assays (Fig 5B and S4B Fig). Interestingly, the D194A mutation in PhyR seems to weaken its interaction with PhyT, while no effect on PhyT interaction could be observed for the E235A derivative (Fig 5). However, we did not observe interactions between either PhyR and any of the Paks or between SdrG and PhyT (S4 Fig).
The protein-protein interaction of PhyR and PhyT predicts that PhyR is bound to the membrane under unstressed conditions. To test this, we studied the interaction between PhyR and PhyT by observing the association of sfGFP-tagged PhyR to the cell membrane depending on the stress condition and the presence/absence of PhyT using fluorescence microscopy. In a first set of experiments, we found that PhyR localized to the membrane under unstressed conditions, which supports binding of PhyR to the membrane-anchored PhyT (Fig 6A), while a homogeneous distribution was observed for sfGFP alone (S5 Fig). PhyR dissociated from the membrane under stress conditions, presumably due to PhyR phosphorylation and subsequent binding to NepR (Fig 6A). To confirm that PhyT is required for PhyR membrane localization, the cellular distribution of sfGFP-tagged PhyR was examined in a pakA-G deletion mutant with and without an additional phyT knockout. Supporting our previous results, membrane association of PhyR was abolished in the absence of PhyT (Fig 6B). Our data imply that PhyT binds to unphosphorylated PhyR, which is in agreement with the BACTH assay (Fig 5).
In addition, we tested the relevance of the phosphorylatable His-341 of PhyT for the interaction with PhyR (Fig 6C). We observed membrane localization of sfGFP-PhyR in a pakA-G deletion mutant with an additional phyT knockout mutant overproducing wild-type PhyT, while overproduction of the PhyT (H341A) derivative in the same strain background led to a homogeneous distribution of sfGFP-PhyR in the cells (Fig 6C).
Taken together, our results support complex formation between PhyT and PhyR. Furthermore, we showed that PhyR is released upon stress induction, indicating that only unphosphorylated PhyR is bound to PhyT, which provides a means by which PhyR is sequestered from direct phosphorylation by the Paks.
In this study, we uncovered essential functions in the regulation of the GSR in the alphaproteobacterium S. melonis Fr1, which resulted in an extended model of the core pathway involved in the activation of the crucial regulon (Fig 7). Based on our data, we defined the function of membrane-anchored PhyT, formerly described as the PhyR phosphatase PhyP, [14] and placed the SDRR SdrG in the core mechanism of GSR. We present in vivo and in vitro evidence that the negative regulatory function of PhyT, which prevents the lethal over activation of the GSR [14], relies on complex formation with unphosphorylated PhyR (Figs 5 and 6). Our results also identify PhyT as a phosphotransferase participating in a GSR-activating phosphorelay (Fig 1). The importance of the Pak-SdrG-PhyT-PhyR phosphorelay became evident when we discovered that both SdrG and PhyT are needed for appropriate GSR activation in stressful conditions in vivo (Figs 2 and 3, S3 Fig).
Altogether, we suggest a model for GSR activation according to which Paks phosphorylate SdrG upon stress induction. When the concentration of SdrG~P reaches a threshold level, it induces the phosphotransferase activity of PhyT in a competitive way, thereby temporarily replacing PhyR to transfer its phosphoryl group to PhyT. EcfG-bound NepR could then eventually prime PhyR [35] via direct interaction for facilitated phosphorylation by PhyT. Overall, our data suggest a mechanism by which a precise interplay between the bifunctional regulator PhyT, SdrG, PhyR and the seven stress-sensing Paks is essential to ensure appropriate GSR induction via "sigma factor mimicry" and thus via NepR and the general stress sigma factor EcfG.
Our in vitro data indicate that the phosphotransferase PhyT irreversibly transfers the phosphoryl group from SdrG~P to PhyR (Fig 1 and S1 Fig). This is in contrast to other described bacterial phosphotransferases such as Spo0B of B. subtilits that catalyze reversible phosphotransfer reactions [36]; however, note that unidirectional phosphotransfer has been observed for the histidine phosphotransferase YPD-1 in Saccharomyces cerevisae to the response regulator SSK1-R2 [37]. Currently, we cannot rule out that PhyT switches its activity to PhyR dephosphorylation in vivo either by turning into a phosphatase or by catalyzing the reverse phosphotransfer reaction in response to external signals e.g. via its so far uncharacterized periplasmic domain or via protein-protein interaction with so far unknown interaction partners. It thus remains to be investigated how the GSR is shut down and whether an additional phosphatase exists. Additionally, it is unknown whether the system relies on the instability of the phosphoryl-Asp of PhyR, basal PhyR protein turnover or if stress-dependent PhyR degradation plays a role, as it has been observed in B. abortus [38].
S. melonis Fr1 inhabits the phyllosphere, a stressful environment characterized by constantly changing conditions [13]. Therefore, it is likely that a strict regulatory system such as the GSR is under constant pressure to evolve in terms of sensitivity and specificity. PhyT and SdrG might have co-evolved in S. melonis Fr1 in order to ensure appropriate GSR induction while simultaneously counteracting the danger of lethal over activation.
We suspect that PhyT evolved from a histidine kinase, e.g. via degeneration of its catalytic domain [9, 14]. Precedence exists in other two-component systems that involve degenerated histidine kinases. These include the dimeric histidine phosphotransferases Spo0B from B. subtilis involved in the initiation of sporulation [5] and ChpT from C. crescentus, which is involved in cell cycle progression [6, 39, 40]. Notably, the evolution of PhyT-type regulators is not limited to the model strain S. melonis Fr1, but rather conserved among alphaproteobacteria [9]. Various studies have been conducted on PhyT-type negative GSR regulators [11, 17, 18, 31], albeit the mechanism by which they act at the molecular level remains to be elucidated. It is conceivable that this protein functions as a phosphotransferase in other alphaproteobacteria as well. Nonetheless, features and regulatory strategies of the regulators mediating the GSR seem to have diversified to different extents in the phylogenetic class, likely as a result of gene duplication events and divergent evolutionary pressures in the various linages of alphaproteobacteria. Regardless, GSR regulation must be precisely coordinated with the requirements of the corresponding bacterium, i.e. depending on the properties of the present stress-sensing kinases in terms of specificity and efficiency regarding PhyR phosphorylation. Thereby, an increasing number of partially redundant kinases as in S. melonis Fr1 [19] and M. extorquens [27] might bring the need to build in tunable negative regulators and a phosphorelay. Altogether, the characterization of PhyT together with SdrG as essential players in GSR regulation in S. melonis Fr1 shades light on an evolutionary pattern of histidine kinase derived regulators, which is conserved within and beyond the boundaries of the GSR.
S. melonis Fr1 in-frame deletion mutants were constructed with the plasmid pAK405 via double homologous recombination [41].
All plasmids constructed during this study are listed in S1 Table. The primers used for plasmid construction are listed in S2 Table.
For heterologous expression of NepR, PhyR, and SdrG in E. coli, overnight cultures of BL21(DE3)Gold pET26bII-nepR-His6, pET26bII-phyR-His6 and pET28b-sdrG-Strep (S1 Table) were grown in 5 mL LB-Lennox medium supplemented with kanamycin (50 μg/mL). Stationary cultures were diluted and incubated further at 37°C to an OD600 of 0.8. After induction with 1 mM IPTG, cells were grown for 3.5 h at 37°C prior to harvest. The pellets were washed once with 1x cold PBS before being stored at -80°C. For expression of PakF, a BL21-Gold (DE3) pDEST-His6-MBP-pakF pre-culture was inoculated in the morning in LB-Lennox medium supplemented with carbenicillin (50 μg/mL) and grown until turbidity could be observed. The main culture was subsequently inoculated and incubated at 37°C. As soon as an OD600 of 0.8 was reached, the flask was placed on ice for 30 min. Afterwards, expression was induced with 50 μM IPTG and further incubation overnight was conducted at 16°C. Cells were washed once with 1x PBS prior to harvest and stored at -80°C. NepR-, PhyR-, and PakF-containing cell pellets were thawed and resuspended in lysis buffer (20 mM HEPES, 0.5 M NaCl, 10% glycerol, 1 tablet protease inhibitor (EDTA-free, Roche), 2 mM beta-mercaptoethanol, 0.1 mg/mL DNase and 20 mM imidazole; pH 8.0). Cells were passed 3x through the French press for cell lysis, followed by centrifugation (12.000g, 30 min, 4°C) to remove cell debris. The supernatant was incubated for 1 h at 4°C with 500 μL Ni-NTA beads (Macherey-Nagel) (1 mL slurry, washed 2x with wash buffer (20 mM HEPES, 0.5 M NaCl, 10% glycerol and 20 mM imidazole; pH 8.0)). Afterwards, the beads were loaded on a polypropylene column. Proteins were eluted from the Ni-NTA resin with 2.6 mL elution buffer (wash buffer with 200 mM imidazole) after washing with 40 mL wash buffer by gravity flow. The SdrG-containing cell pellet was thawed and also resuspended in imidazole-free lysis buffer and passed 3x through the French press. After centrifugation, SdrG-containing supernatant was added to a polypropylene column loaded with 1 mL Strep-tactin beads (IBA Lifesciences) (2 mL slurry), which were equilibrated with 2 mL imidazole-free wash buffer prior to use. The beads were washed with 10 mL imidazole-free wash buffer by gravity flow, before SdrG-Strep was eluted with 3 mL wash buffer supplemented with 2.5 mM desthiobiotin. All cleaned up proteins were subjected to PD10 desalting columns for exchange to kinase buffer (10 mM HEPES, 50 mM KCl, 10% glycerol; pH 8.0). The purified proteins were concentrated with 3 kDa cutoff amicon tubes (Millipore). Protein concentrations were determined with a BCA protein assay (Life Technologies Europe B.V.). 50 μL aliquots were stored at -20°C.
Heterologous overexpression of PhyT and the PhyT (H341A) derivative was carried out as described for NepR and PhyR using IPTG-inducible pET26bII expression plasmids (S1 Table). Cell pellets were thawed and resuspended in imidazole-free lysis buffer. Cells were passed 3x through the French press. The supernatant of the first centrifugation step (12.000g, 30 min, 4°C) was subjected to ultracentrifugation (180.000g, 1 h, 4°C). The membrane pellets were resuspended in kinase buffer to a concentration of 100 mg membrane fraction/mL and 100 μL aliquots were stored at -80°C. Comparable amounts of PhyT and the PhyT (H341A) derivative in the prepared E. coli membrane particles were confirmed with standard SDS-PAGE followed by Western blot analysis using a mouse Tetra·His antibody (1:2.000) (34570 Qiagen AG) and a goat α-mouse HPR-coupled antibody (1:3.000) (BioRad) (S2A and S2B Fig).
Protein expression and purification as well as preparation of E. coli membrane particles are described above. Purified proteins were thawed on ice. First, PakF was diluted to a concentration of 10 μM in kinase buffer supplemented with 10 mM MgCl2 and 1 mM DTT. 200 μL Ni-NTA beads (400 μL slurry) were washed 2x with kinase buffer supplemented with 10 mM MgCl2 and 1 mM DTT before 320 μL PakF (10 μM) were added and incubated at room temperature (RT) for 30 min. Afterwards, autophosphorylation of PakF was initiated with 3.2 μL [γ-32P]ATP (5.000 Ci/mmol; Hartmann Analytic GmbH). After 10 min of autophosphorylation, PakF bound to Ni-NTA resin was loaded on polypropylene columns and washed with kinase buffer until radioactivity of the wash fraction decreased significantly.
S. melonis Fr1 strains carrying the reporter plasmid pAK501-nhaA2p-lacZ and an inducible vanillate or cumate expression plasmid (S1 Table) were streaked out from a cryo-stock on TYE agar plates (1% tryptone, 0.5% yeast extract, 2.4 mM Na2HPO4, 37.6 mM KH2PO4) containing tetracycline (10 μg/mL) and chloramphenicol (34 μg/mL) and incubated overnight at 28°C. The plates were sealed with parafilm and protected from light, to reduce stress exposure. Pre-cultures were inoculated from a small loop in 20 mL TYE medium supplemented with tetracycline (10 μg/mL) and chloramphenicol (34 μg/mL) in 100 mL baffled flasks and incubated at 28°C, 160 rpm during the day.
To test protein-protein interactions, C- and N-terminal translational fusions of the T18 and T25 domains of the adenylate cyclase CyaA of Bordetella pertussis [42] were used for PhyT, NepR, PhyR wild-type, PhyR derivatives and PakA-G (S1 Table). Fusions were tested in pairwise combinations in E. coli BTH101 cya- (Euromedex). For the interaction analysis the optimal pair for each protein combination is shown. For each transformation mixture, 5 μl were spotted onto LB agar plates containing 0.5 mM isopropyl 1-thio-β-D-galactopyranoside and 40 μg/ml 5-bromo-4- chloro-3-indolyl β-D-galactopyranoside (X-Gal) with selection for carbenicillin (50 μg/mL) and kanamycin resistance (50 μg/mL). Plates were incubated at 30°C for 24 h. Formation of blue colonies was scored as a positive interaction. In addition, a volume of 20 μl of each transformation mixture was plated on LB agar plates containing carbenicillin (50 μg/mL) and kanamycin (50 μg/mL) for selection. Plates were incubated at 28°C overnight. For quantitative analysis, 5 mL cultures of LB supplemented with carbenicillin (50 μg/mL) and kanamycin (50 μg/mL) for selection and 0.5 mM IPTG were inoculated from single colonies. Overnight cultures were grown at 30°C and used to measure β-galactosidase activity (Miller 1972). 500 μl of bacterial culture were spun down and pellets were resuspended in 1x Lämmli buffer so that a final OD600 of 10 was reached. The samples were subjected to standard SDS-PAGE followed by Western blot analysis. For detection of T18 fusions the mouse α-CyaA (3D1) monoclonal antibody (Santa Cruz Biotechnology) (1:2.000) and a goat α-mouse HPR-coupled antibody (1:3.000) were used (BioRad).
S. melonis Fr1 strains carrying either pQY-sfGFP or pQYD-sfGFP-phyR (S1 Table) were streaked out from cryo-stock on TYE agar plates supplemented with tetracycline (10 μg/mL) and were incubated overnight at 28°C. The plates were sealed with parafilm and protected from light to avoid stress induction. Pre-cultures were inoculated in 20 mL TYE medium supplemented with tetracycline (10 μg/mL) in 100 mL baffled flasks and incubated during the day at 180 rpm and 28°C. Overnight cultures were inoculated at an OD600 appropriate to reach exponential growth phase at the time of the experiment. Incubation conditions were the same as for the pre-cultures. PhyR-sfGFP or sfGFP expression in the S. melonis Fr1 mutants was induced in mid-exponential phase via the addition of 25 μM cumate (100 mM stock in 100% ethanol) for 12 min. In order to analyze the importance of His-341 of PhyT for binding of PhyR and therefore membrane localization of sfGFP-PhyR, the appropriate strains carrying pVH-sfGFP-phyR and either pAK200-phyT or pAK200-phyT (H341A) (S1 Table) were streaked out from cryo-stock on TYE agar plates supplemented with tetracycline (10 μg/mL) and kanamycin (50 μg/mL). Following overnight incubation at 28°C, pre-cultures were inoculated in 20 mL TYE medium supplemented with tetracycline (10 μg/mL) and kanamycin (50 μg/mL) in 100 mL baffled flasks and cultivated as described above. In addition to both antibiotics, cumate (25 μM) was added to the 20 mL overnight cultures. Production of sfGFP-PhyR was induced on the following day with 250 μM vanillate (250 mM stock in 100% ethanol) for 12 min in mid-exponential phase. The bacteria were subsequently washed with Tris-buffered saline (50 mM Tris-HCl, 150 mM NaCl; pH 7.55) and mounted on a cover slip for imaging. The live-cell imaging was performed on an inverse spinning disc microscopy system (Visitron software, Yokogawa CSU-X1 spinning-disk confocal unit) equipped with a solid state Laser Unit (Toptica) at 488 nm excitation wavelength, a 100x Oil Plan-Neofluar Objective (Zeiss, NA: 1.3), and an Evolve 512 EMCCD camera (Photometrics). The images were deconvolved using Huygens Professional version 17.04 (Scientific Volume Imaging, The Netherlands, http://www.svi.nl/HuygensSoftware).
|
10.1371/journal.pgen.1001362 | Quantitative Fitness Analysis Shows That NMD Proteins and Many Other Protein Complexes Suppress or Enhance Distinct Telomere Cap Defects | To better understand telomere biology in budding yeast, we have performed systematic suppressor/enhancer analyses on yeast strains containing a point mutation in the essential telomere capping gene CDC13 (cdc13-1) or containing a null mutation in the DNA damage response and telomere capping gene YKU70 (yku70Δ). We performed Quantitative Fitness Analysis (QFA) on thousands of yeast strains containing mutations affecting telomere-capping proteins in combination with a library of systematic gene deletion mutations. To perform QFA, we typically inoculate 384 separate cultures onto solid agar plates and monitor growth of each culture by photography over time. The data are fitted to a logistic population growth model; and growth parameters, such as maximum growth rate and maximum doubling potential, are deduced. QFA reveals that as many as 5% of systematic gene deletions, affecting numerous functional classes, strongly interact with telomere capping defects. We show that, while Cdc13 and Yku70 perform complementary roles in telomere capping, their genetic interaction profiles differ significantly. At least 19 different classes of functionally or physically related proteins can be identified as interacting with cdc13-1, yku70Δ, or both. Each specific genetic interaction informs the roles of individual gene products in telomere biology. One striking example is with genes of the nonsense-mediated RNA decay (NMD) pathway which, when disabled, suppress the conditional cdc13-1 mutation but enhance the null yku70Δ mutation. We show that the suppressing/enhancing role of the NMD pathway at uncapped telomeres is mediated through the levels of Stn1, an essential telomere capping protein, which interacts with Cdc13 and recruitment of telomerase to telomeres. We show that increased Stn1 levels affect growth of cells with telomere capping defects due to cdc13-1 and yku70Δ. QFA is a sensitive, high-throughput method that will also be useful to understand other aspects of microbial cell biology.
| Telomeres, specialized structures at the end of linear chromosomes, ensure that chromosome ends are not mistakenly treated as DNA double-strand breaks. Defects in the telomere cap contribute to ageing and cancer. In yeast, defects in telomere capping proteins can cause telomeres to behave like double-strand breaks. To better understand the telomere and responses to capping failure, we have combined a systematic yeast gene deletion library with mutations affecting important yeast telomere capping proteins, Cdc13 or Yku70. Quantitative Fitness Analysis (QFA) was used to accurately measure the fitness of thousands of different yeast strains containing telomere capping defects and additional deletion mutations. Interestingly, we find that many gene deletions suppress one type of telomere capping defect while enhancing another. Through QFA, we can begin to define the roles of different gene products in contributing to different aspects of the telomere cap. Strikingly, mutations in nonsense-mediated mRNA decay pathways, which degrade many RNA molecules, suppress the cdc13-1 defect while enhancing the yku70Δ defect. QFA is widely applicable and will be useful for understanding other aspects of yeast cell biology.
| Linear chromosome ends must be protected from the DNA damage response machinery and from shortening of chromosome ends during DNA replication [1], [2]. Chromosome ends therefore adopt specialized structures called telomeres, distinct from double-stranded DNA breaks elsewhere in the genome. Telomeric DNA is protected, or capped and replicated by a large number of different DNA-binding proteins in all eukaryotic cell types [2], [3].
In budding yeast, numerous proteins contribute to telomere capping and amongst these are two critical protein complexes, the Yku70/Yku80 (Ku) heterodimer and the Cdc13/Stn1/Ten1 (CST) heterotrimeric complex [4]. Orthologous protein complexes play roles at telomeres in other eukaryotic cell types suggesting that understanding the function of the Ku and CST protein complexes in budding yeast will be generally informative about key aspects of eukaryotic telomere structure and function.
In budding yeast Yku70 is a non-essential protein that has multiple roles in DNA repair and at telomeres, being involved in the non-homologous end-joining (NHEJ) DNA repair pathway, in the protection of telomeres and the recruitment of telomerase. The mammalian orthologue, Ku70, has similar properties [5]. In budding yeast, deletion of the YKU70 gene (yku70Δ) results in short telomeres and temperature sensitivity [6]. At high temperatures, cells lacking Yku70 accumulate ssDNA at telomeres, which activates the DNA damage response and leads to cell-cycle arrest [7], [8], [9].
Cdc13 is a constituent of the essential budding yeast Cdc13-Stn1-Ten1 (CST) protein complex which is analogous to the CST complex found recently in mammalian, plant and fission yeast cells [10], [11]. Cdc13 binds to ssDNA overhangs at telomeres and functions in telomerase recruitment and telomere capping [12], [13], [14]. Acute inactivation of Cdc13 by the temperature sensitive cdc13-1 allele induces ssDNA generation at telomeres and rapid, potent checkpoint-dependent cell cycle arrest [14].
cdc13-1 or yku70Δ mutations each cause temperature dependent disruption of telomere capping that is accompanied by ssDNA production, cell-cycle arrest and cell death [7], [15]. Interestingly, the poor growth imparted by each mutation can be suppressed by deletion of EXO1, removing the Exo1 nuclease that contributes to ssDNA production when either Cdc13 or Yku70 is defective [7]. However, cdc13-1 and yku70Δ mutations show a synthetic poor growth interaction [8] and different checkpoint pathways are activated by each mutation [7]. These latter observations, along with numerous others, show that CST and Ku complexes perform distinct roles capping budding yeast telomeres and that further clarification of their functions at the telomere is important to help understand how eukaryotic telomeres function.
Many insights into the telomere cap and the DNA damage responses induced when capping is defective were first identified as genetic interactions. For example all DNA damage checkpoint mutations suppress the temperature sensitive growth of cdc13-1 mutants [16], but only a subset of these suppress the temperature sensitive growth of yku70Δ mutants [7]. We reasoned that the roles of Cdc13 and Yku70 at telomeres could be further understood by quantitative, systematic analysis of genetic interactions between telomere capping mutations and a genome-wide collection of gene deletions.
We used standard synthetic genetic array (SGA) approaches to combine the systematic gene deletion collection with cdc13-1 and yku70Δ mutations [17], [18]. After this, strain fitnesses were measured at a number of temperatures by quantitative fitness analysis (QFA). For QFA, liquid cultures were spotted onto solid agar plates and culture growth was followed by time course photography. Images were processed and fitted to a logistic growth model to allow an accurate estimation of growth parameters, such as doubling time. In other high-throughput experiments such as SGA or EMAP approaches, culture fitness is determined from colony size [17], [18], [19]. In QFA, analysis of growth curves of cultures grown on solid agar plates allows us to measure fitness more precisely.
Through QFA we identify hundreds of gene deletions, in numerous different classes, showing genetic interactions with cdc13-1, yku70Δ or both. One particularly striking example of the type of genetic interactions we measured by QFA is between deletions affecting nonsense mediated RNA decay pathways (upf1Δ, upf2Δ, upf3Δ), cdc13-1 and yku70Δ. Additional experiments show that disabling nonsense mediated mRNA decay pathways, using upf2Δ as an example, suppresses the cdc13-1 defect but enhances the yku70Δ defect by increasing the levels of the telomere capping protein Stn1. QFA is generally applicable and will be useful for understanding other aspects of yeast cell biology or studying other microorganisms.
To systematically examine genetic interactions between a genome-wide collection of gene deletion strains (yfgΔ, your favorite gene deletion, to indicate any of ∼4200 viable systematic gene deletions) and mutations causing telomere capping defects we crossed the knockout library to cdc13-1 or yku70Δ mutations, each affecting the telomere, or to a neutral control query mutation (ura3Δ) using SGA methodology [17], [18]. Since both cdc13-1 and yku70Δ mutations cause temperature sensitive defects, we generated all double mutants at low, permissive temperatures before measuring the growth of double mutants at a number of semi-permissive or non-permissive temperatures. We cultured yku70Δ yfgΔ strains at 23°C, 30°C, 37°C and 37.5°C, cdc13-1 yfgΔ strains at 20°C, 27°C and 36°C and ura3Δ yfgΔ strains at 20°C, 27°C and 37°C and measured fitness.
Double mutant fitness was measured after spotting of dilute liquid cultures onto solid agar. We estimate approximately 100 separate cells were placed in each of 384 spots on each agar plate. Fitness of thousands of individual cultures, each derived from spotted cells, was deduced by time course photography of agar plates followed by image processing, data analysis, fitting of growth measurements to a logistic model and determination of quantitative growth parameters (Figure 1) [20], [21], [22]. We fitted logistic growth model parameters to growth curves allowing us to estimate maximum doubling rate (MDR, population doublings/day) and maximum doubling potential (MDP, population doublings) of approximately 12,000 different yeast genotypes (e.g. cdc13-1 yfg1Δ, yku70Δ yfg1Δ, etc.) at several temperatures. At least eight independent biological replicates for each strain at each temperature were cultured and repeatedly photographed, capturing more than 4 million images in total. To rank fitness we assigned equal importance to maximum doubling rate and maximum doubling potential and defined strain fitness as the product of the MDR and MDP values (Fitness, F, population doublings2/day). Other measures of fitness can be derived from the sets of logistic parameters available from Text S1.
Figure 1A shows approximately 170 example images, corresponding to eight independent time courses for each of three pair-wise combinations of yku70Δ, ura3Δ and upf2Δ mutations. These example images clearly show, qualitatively, that upf2Δ yku70Δ strains grow less well than yku70Δ ura3Δ strains, which in turn grow less well than upf2Δ ura3Δ strains at 37°C. These fitness measures are consistent with numerous earlier studies, showing that yku70Δ mutants do not grow well at high temperatures, but also demonstrate a novel observation, that the upf2Δ mutation enhances the yku70Δ defect and this is further investigated below. Images like those in Figure 1A were processed, quantified, plotted and logistic growth curves fitted to the data (Figure 1B). We applied QFA to all genotypes at each temperature, as the three examples in Figure 1C illustrate.
QFA of cdc13-1 yfgΔ, yku70Δ yfgΔ and ura3Δ yfgΔ double mutant libraries was performed at a number of temperatures and therefore a variety of informative comparisons were possible. For example to help identify gene deletions that suppress or enhance the yku70Δ temperature dependent growth defect it is useful to compare the fitness of yku70Δ yfgΔ cells incubated at 37.5°C, with that of control, ura3Δ yfgΔ, cells incubated at 37°C. In Figure 2, genes which, when deleted, suppress the yku70Δ phenotype at 37.5°C will be positioned above the linear regression line and enhancers of yku70Δ defects below the line. yfgΔ mutations that result in low fitness when combined with the neutral ura3Δ mutation will be found on the left and those with high fitness on the right of the x-axis.
The location of each gene in Figure 2 indicates the effect of each deletion on fitness of yku70Δ strains versus the effect of the deletion on fitness of ura3Δ strains. The regression line drawn through all data points (solid gray line) indicates the expected yku70Δ yfgΔ fitness given the fitness of the corresponding ura3Δ yfgΔ mutant. The line of equal growth (dashed gray line) shows the expected positions of yku70Δ yfgΔ strains if they grew similarly to ura3Δ yfgΔ strains. Comparing the linear regression with the line of equal growth, it is clear that yku70Δ mutants grow poorly relative to control ura3Δ mutants, as expected due to the temperature dependent telomere uncapping observed in yku70Δ mutants. Figure 2 also highlights large numbers of yku70Δ yfgΔ strains growing significantly better than expected, given the fitness of the equivalent ura3Δ yfgΔ mutation at 37°C (red data points, Figure 2) and these yfgΔ genes can be classified as yku70Δ suppressors. There are also large numbers of yku70Δ yfgΔ strains that grow worse than expected and these are classified as yku70Δ enhancers (green data points, Figure 2). Three further example plots comparing growth of yfgΔ cdc13-1 versus yfgΔ ura3Δ at 20°C; yfgΔ cdc13-1 versus yfgΔ ura3Δ at 27°C and yfgΔ ura3Δ at 37°C versus yfgΔ ura3Δ at 20°C are shown in Figure S1 and others can be found on our supporting information data files website.
We estimated genetic interaction strength (GIS) as the vertical displacement of each yku70Δ yfgΔ normalised mutant fitness from the expected normalised fitness, with expected fitness given by a linear regression model (see Text S1, experimental procedures). GIS is dimensionless. This method is equivalent to defining GIS as the deviation of observed fitness from that expected if a multiplicative model of genetic interaction were correct. In all, more than 30,000 genetic interaction strengths, together with their statistical significances, were calculated (Tables S1, S2, S3, S4, S5, S6). Table 1 summarizes the numbers of statistically significant genetic interactions observed under the different conditions of telomere capping. Table 1 clearly illustrates that many more genetic interactions are observed under conditions of mild telomere uncapping (cdc13-1 strains at 27°C and yku70Δ strains at 37.5°C) and that at these temperatures around 5% of gene deletions can show strong suppressing or enhancing interactions (GIS >0.5).
In order to compare the effects of gene deletions on cell fitness when combined with cdc13-1 or yku70Δ induced telomere cap defects, it was particularly useful to compare the GIS of each gene with respect to cdc13-1 or yku70Δ after induced telomere uncapping. Figure 3 summarises how different gene deletions interact with the two types of telomere capping defect, suppressing, enhancing or showing no strong interaction with each telomere cap defect. For example, genes that when deleted significantly suppress temperature sensitivity of both cdc13-1 and yku70Δ mutants appear in the top right of this plot (Figure 3, region 3). EXO1 is in this area as expected because Exo1 generates ssDNA at telomeres in both types of telomere capping mutants (Figure 3, region 2/3, arrow) [7]. Deleting components of the checkpoint sliding clamp (9-1-1 complex) and its clamp loader, suppress cdc13-1 but have minor effects on growth of yku70Δ mutants [7]. DDC1, RAD17 and RAD24 are in region 2, as expected. MEC3, encoding the third component of the sliding clamp was missing from our knock out library and was not tested. Gene deletions that disrupt the telomerase enzyme directly (est1Δ, est3Δ) enhance the temperature sensitivity of both mutations and so appear in region 7. Genes that, when deleted, suppress cdc13-1 yet enhance the yku70Δ temperature sensitivity (Figure 3, region 1) represent a novel telomere-related phenotype and interestingly include three major components of the nonsense mediated RNA decay pathways (UPF1, UPF2, UPF3). It is reassuring that the UPF genes cluster so closely in Figure 3 because this strongly suggests that positioning of genes on this plot is an accurate measure of the function of the corresponding gene products in telomere biology.
The position of YKU80 in the bottom right corner of region 8 is informative. The negative interaction of yku80Δ with cdc13-1 is expected since it is known that yku80Δ (and yku70Δ) mutations reduce fitness of cdc13-1 mutants, even at permissive temperatures [8]. However, the positive effect of yku80Δ on the growth of yku70Δ mutants appears, at first, surprising. The positive epistatic effect simply reflects the fact that yku70Δ, yku80Δ and yku70Δ yku80Δ double mutants are all similarly unfit at high temperatures. We have confirmed that in the different W303 genetic background that yku70Δ, yku80Δ and yku70Δ yku80Δ double mutants are all similarly unfit at high temperatures. According to the multiplicative model of epistasis the fitness of the yku70Δ yku80Δ double mutants is significantly higher than expected based on the fitness of the single mutants. Thus, by this criterion, yku80Δ suppresses the yku70Δ fitness defect. These data can be explained if neither single sub-unit of the Ku comlex retains a telomere capping function in the absence of the other.
It is reasonable to hypothesize, based partly on the behaviour of UPF1, UPF2 and UPF3 genes, that genes having similar genetic interactions with cdc13-1 and yku70Δ under particular conditions which are proximal in Figure 3 share similar functions in telomere biology. For example, genes that function similarly to EXO1 and for example, regulate ssDNA at uncapped telomeres might appear close to EXO1 in Figure 3. Similarly, genes with strong effects on telomerase function might appear in region 7. Consistent with this hypothesis, it is clear from Figure 3 that many genes encoding members of the same protein complex, or proteins which work together to perform a particular function, often have similar genetic interaction profiles and are located in similar positions on this plot. Examples in Figure 3 include: NMD pathway (UPF1, UPF2, UPF3, region 1); OCA complex (regions 1 & 4); clamp-loader and clamp-like complex (RAD24, DDC1, RAD17, region 2); telomerase (EST1, EST3, region 7) and dipthamide biosynthesis (JJJ3, DPH1, DPH2, DPH5, regions 8 & 9) genes, as well as the numerous other complexes highlighted by the key at the bottom of Figure 3. Table 2 shows the number of genes found in each section of Figure 3. Table 3 lists 19 different sets of genes that are functionally or physically related and that cluster in Figure 3 as well as the single genes EXO1, RIF1, RIF2 and TEL1 also found in interesting positions. EXO1 is in its expected position but it is interesting that RIF1 and RIF2 are found in different positions in Figure 3, suggesting they have different functions in telomere biology. Further experiments in the W303 genetic background confirm the different interactions of RIF1 and RIF2 with cdc13-1 (Xue, Rushton and Maringele, submitted). TEL1 encodes the ATM orthologue and is required for telomere maintenance and it clusters very near components of telomerase, in region 7.
Groupings such as these and their positioning on this type of plot help generate testable, mechanistic predictions about the roles of proteins/protein complexes on telomere capping in budding yeast. For example, we predict that NMD genes (which we examine further in this study) and dipthamide synthesis genes have opposing effects on both Cdc13-mediated and Yku70-mediated telomere capping, because they lie in opposite corners of Figure 3.
The QFA experiments summarised by Figure 3 were performed in a high-throughput manner with the systematic knock out collection in the S288C genetic background and the fitness of different query mutants was measured in slightly different types of media. It was therefore conceivable that some of the genetic interactions scored were due to: defects in the knock out collection, such as incorrect mutations being present or the presence of suppressor mutations, the S288C genetic background, subsets of the cell populations that progressed through the mass mating, sporulation and germination that occur during SGA or media differences.
To test whether genetic interactions identified by QFA with cdc13-1 or yku70Δ strains were robust observations we retested a subset of interactions in the W303 genetic background, on rich media, after construction of strains by individual tetrad dissection by manual spot test. Figure 4 and Figure S2 show the behaviour of a number of gene deletions chosen from different regions in Figure 3 to test the effects in W303. In all we measured 26 genetic interactions between 13 gene deletions and cdc13-1 or yku70Δ. Of these we estimate that 20/26 interactions were as expected, 5/26 difficult to classify, and 1/26, due to elp6Δ, opposite to that expected after QFA. In particular exo1Δ, mlp1Δ, mlp2Δ, mak31Δ and dph1Δ mutations suppress the growth defects of yku70Δ strains in W303 at 36°C, consistent with their position on the right side of Figure 3 and exo1Δ, rad24Δ, upf1Δ and upf2Δ strongly suppress cdc13-1 at 26°C consistent with their position near the top of Figure 3. upf1Δ, upf2Δ, rrd1Δ and pph3Δ mutations all reduced growth of yku70Δ strains at 36°C consistent with their position on the left of Figure 3, while elp6Δ, mak31Δ, dph2Δ, rrd1Δ and pph3Δ mutations all enhanced cdc13-1 growth defects consistent with their positions near the bottom of Figure 3. Other genes have more subtle effects, the oca1Δ and oca2Δ mutations had marginal effects on yku70Δ strains but improved growth of cdc13-1 strains (Figure S2). Interestingly the elp6Δ mutation enhanced the cdc13-1 defect at 26°C, as expected, but suppressed the yku70Δ strain growth defect at 36°C, the opposite of what was expected from Figure 3. Further experiments will be necessary to clarify the role of Elp6 and other elongator factors in cells with uncapped telomeres. However, overall, it is clear that the majority of genetic interactions identified by QFA are reproducible in smaller scale experiments in a different genetic background.
Suppressors and enhancers of the cdc13-1 and yku70Δ phenotypes were most easily identified at semi-permissive temperatures for the query mutations (Figure 2, Figure 3), however QFA at other temperatures also proved informative. For example, comparison of the fitness of yfgΔ ura3Δ strains at 37°C versus 20°C, allowed us to identify temperature sensitive mutants (Figure S1C and Table S9). Of the 57 genes which were categorized with a phenotype of “heat sensitivity: increased” in the Saccharomyces Genome Database (http://www.yeastgenome.org), as identified by low though-put experiments, which were also present in the knockout library we used, 45 (79% of total) were identified as being significantly heat sensitive by our independent QFA.
2-dimensional GIS plots, like Figure 3, also proved useful for identifying broader patterns of genetic interactions. For example, we observed a difference between the effects of deleting small and large ribosomal subunit genes on the growth of telomere capping mutants (Figure S3A, S3B). Gene deletions which affect the small ribosomal subunit are generally neutral with both cdc13-1 and yku70Δ mutations (Figure S3A, S3B red). In contrast, disruptions of large ribosomal subunit function suppress the effect of cdc13-1 on average and enhance that of yku70Δ (Figure S3A, S3B blue). Although the basis for this novel observation is unknown it may be related to the finding that the large ribosome sub-unit is subject to autophagy upon starvation, whereas the small ribosome sub-unit is not [23]. Positive and negative regulators of telomere length [24], [25], [26] also showed differing distributions in GIS comparisons – gene deletions which suppress the yku70Δ defect are more likely to result in long than short telomeres (Figure S3C). This is perhaps to be expected since yku70Δ mutants, on their own, have a short telomere phenotype. Importantly, over 90% of genes identified as suppressors of cdc13-1 in a previous study [20] showed a positive GIS with cdc13-1 (Figure S2D), demonstrating that QFA reproduces conclusions derived from qualitatively scored visual inspection. It should be noted however, that the improved sensitivity of QFA has allowed identification of significantly more enhancing mutations than were indentified in the preceding, qualitatively scored study [20].
QFA is sensitive enough to permit identification of genetic interactions even where gene deletions combined with the control ura3Δ query mutation impart a poor growth phenotype. For example, deletion of all three SPE genes resulted in low fitness that was strongly rescued by cdc13-1 (Figure S1B, blue, Figure 3 region 2). Interestingly it has recently been reported that increasing spermidine levels increases lifespan in organisms such as yeast, flies and worms [27], but no previous connection with telomeres has been made. Telomere-driven, replicative senescence is thought to be a significant component of the ageing phenotype. Our observations of interactions between SPE genes and cells with uncapped telomeres may ultimately lead to experiments to provide insight into the mechanisms by which spermidine affects lifespan.
One of the most striking results obtained from QFA experiments was the effect of deleting nonsense mediated RNA decay genes on growth of cells with telomere capping mutations. Deletion of any of the NMD genes UPF1, UPF2 or UPF3 suppresses the cdc13-1 telomere capping defect but enhances the yku70Δ defect (Figure 3, region 1). We wanted to understand the basis of this interesting interaction and decided to further analyze the NMD genes. We also investigated EBS1, a gene that has proposed roles in both the NMD pathway and telomere function [28], [29], [30] and was identified previously as interacting with CDC13 [20], [31]. EBS1 had less strong, but qualitatively similar GISs to UPF genes in our analysis (Figure 3, region 1∼2), suggesting that the position of EBS1 in Figure 3 was due a partial defect in nonsense mediated RNA decay.
One potential mechanism by which UPF genes and EBS1 affect telomere capping is if they regulate the levels of telomere capping proteins. Indeed, UPF genes have been shown to regulate transcripts of genes involved in telomere function [32], [33]. The effect of EBS1 on these transcripts has not so far been reported. Therefore we compared mRNA levels of three NMD targets with roles in telomere regulation (STN1, TEN1 and EST2) in upf2Δ, ebs1Δ and wild-type strains. Transcripts of STN1 and TEN1 were increased significantly in upf2Δ and ebs1Δ, mutants whereas EST2 transcripts were increased only in upf2Δ strains (Figure 5A). We conclude that both EBS1 and UPF2 modulate expression of STN1 and TEN1, but the effects of ebs1Δ are modest compared to those of upf2Δ. Furthermore, elevated levels of Stn1 protein were detected in both ebs1Δ and upf2Δ mutants (Figure 5B). Consistent with the measured mRNA levels of STN1, the increase in Stn1 levels was smaller in ebs1Δ strains than upf2Δ strains. Thus we concluded that the effects of UPF2 and EBS1 could be due to the effects on Stn1 and possibly Ten1 levels.
Increased Stn1 and Ten1 levels are known to suppress the cdc13-1 defect [33], [34]. To test whether elevated levels of Stn1 or Ten1 proteins could reproduce the enhancement of the yku70Δ defect observed in ebs1Δ and upf2Δ mutants, we over-expressed Stn1 and Ten1 independently of NMD by providing extra copies of STN1 and TEN1 on plasmids. Both single copy (centromeric; Figure 5C) and high copy (2 µ) Stn1-expressing plasmids [35] suppressed the temperature sensitivity of cdc13-1 strains and enhanced the temperature sensitivity of yku70Δ strains (Figure S4A), mimicking the upf2Δ and ebs1Δ phenotypes. In contrast, Ten1-expressing plasmids [35] did not affect the growth of either cdc13-1 or yku70Δ mutants (Holstein; data not shown). We therefore conclude that both UPF2 and EBS1 affect telomere capping by modulating expression of STN1. However, it is also possible that UPF2 and EBS1 affect telomere capping by modulating expression of genes other than STN1. To test this and the relative contribution of STN1 versus any other mechanisms, it would be informative to reduce STN1 expression in upf2Δ mutants. Such experiments might be difficult to perform and interpret since both centromeric single-copy and 2 micron multi-copy STN1 plasmids suppress the cdc13-1 defect to similar extents (Figure S4A), suggesting there is not a simple correlation between Stn1 levels and effects on growth of cdc13-1 and yku70Δ mutants.
Since the effect of ebs1Δ was milder than that of upf2Δ on the fitness of cdc13-1 and yku70Δ cells (Figure 3), we hypothesized that if ebs1Δ imparts a mild NMD defect, an ebs1Δ upf2Δ double mutation would result in the same phenotype as upf2Δ on its own. Figure 5D shows that both upf2Δ and ebs1Δ mutations suppress cdc13-1 temperature sensitivity and exacerbate yku70Δ temperature sensitivity in the W303 genetic background. We also confirmed that upf1Δ upf2Δ double mutants suppress cdc13-1 temperature sensitivity and exacerbate yku70Δ temperature similarly to either single mutant (Figure S4B). It is clear that the effects of ebs1Δ are less strong than upf2Δ mutations and interestingly upf2Δ ebs1Δ double mutants have slightly stronger effects on growth of both cdc13-1 and yku70Δ mutants, suggesting that ebs1Δ effects are not solely due to defects in nonsense mediated RNA decay (Figure 5D). We therefore conclude that, at least with respect to telomere capping, EBS1 and UPF2 act partially through different pathways. We do not yet understand these differences, but they may be related to the homology between Ebs1 and the telomerase protein Est1.
It is simple to hypothesize why increased Stn1 levels, caused by inactivation of nonsense mediated mRNA decay pathways, suppress the cdc13-1 defect, presumably by stabilizing the Cdc13-1/Stn1/Ten1 complex at telomeres. It is less simple to explain why increased Stn1 levels enhance the yku70Δ-induced telomere-capping defect. Our hypothesis is based on the facts that the Stn1 protein can inhibit telomerase activity [36], [37] and that Yku70 interacts with and helps recruit telomerase to telomeres [38], [39]. Thus we hypothesized that yku70Δ causes a partial defect in telomerase recruitment, one that is exacerbated by the upf2Δ mutation that causes high levels of Stn1, thus inhibiting telomerase activity. To test the simplest version of this hypothesis, that yku70Δ and upf2Δ mutations reduce the amount of telomerase binding to telomeres, we performed ChIP analyses. We examined binding of the Est2 sub-unit of telomerase in yku70Δ, upf2Δ and yku70Δ upf2Δ double mutants. Interestingly we observed a significant reduction in binding of telomerase to telomeres in yku70Δ, upf2Δ and yku70Δ upf2Δ mutants (Figure 5E). It is known that yku70Δ mutants recruit less telomerase to telomeres [39] but we are unaware of any other reports showing that upf2Δ mutants recruit less telomerase to telomeres. This observation most likely explains the short telomere phenotype of upf2Δ (as well as yku70Δ) mutants [24], [25]. It is noteworthy that although the upf2Δ mutation causes a four-fold increase in the EST2 transcript, it causes a reduction in the amount of Est2 bound to telomeres. This suggests that the increased levels of Stn1in upf2Δ cells more than counteracts any mass action effects on telomerase recruitment to telomeres caused by EST2 over-expression. However, the simple hypothesis that yku70Δ upf2Δ mutants show a more severe capping defect because of a reduction in the recruitment of telomerase appears not to be valid. Further experiments will be necessary to better understand the complex interplay between Ku, nonsense mediated decay pathways, Cdc13, Stn1 and telomere capping (Figure 6).
Systematic measurement of genetic interactions is a powerful way to help understand how cells and organisms function [40], [41]. This is because genetic approaches examine the role of individual gene products, or individual residues in genes, in the context of the whole organism and can help dissect the effects of weak biochemical interactions that are important for cells to function [42]. Systematic SGA and eMAP experiments typically examine millions of genetic interactions and use comparatively crude measures of growth (colony size) to infer genetic interactions [19], [41]. Here we have more accurately measured a smaller number of genetic interactions, focusing on interactions that affect budding yeast telomere function. The telomere is an important and interesting subject for systematic genetic analysis because it is a complex, subtle and in some senses paradoxical nucleic acid/protein structure that plays critical roles during human ageing and carcinogenesis. One paradox of telomeres is that many DNA damage response proteins, which induce DNA repair or cell cycle arrest when interacting elsewhere in the genome, induce neither response at telomeres but instead play important roles in telomere physiology.
We used Quantitative Fitness Analysis (QFA) to accurately assess the fitness of many thousands of yeast strains containing mutations that affect telomere function in combination with other deletion mutations. To assess fitness, cells were grown in parallel, in 384 spot arrays on solid agar plates. Photographs of plates were taken, images processed and analysed and growth curves for each culture generated. The growth curves are in essence very similar to those observed in liquid culture, with clear exponential and saturation phases (Figure 2 from Lawless et al. 2010) and can be summarized with as few as three logistic growth parameters. The major advantage of QFA over parallel liquid culture methods to measure yeast fitness is that many more cultures can be examined in parallel. For example we routinely follow the growth of about 18,000 parallel cultures (4,500 yeast strains, incubated at four different temperatures), whereas parallel liquid culture based methods are generally restricted to up to 200 parallel cultures [43].
QFA is similar to SGA or EMAP approaches but typically four times fewer strains per plate are cultured (384 spots versus 1536 colonies) [17], [18], [19], [41]. A further difference between QFA and SGA is that in QFA, which has a liquid growth phase, double mutants are cultured for longer before fitness is assessed. This means that that in QFA, synthetically sick double mutants often show poorer growth than is observed in SGA experiments simply because the more divisions that occur the easier it is to observe growth defects. There is a risk with QFA that during the comparatively long culturing period that suppressors or modifiers will arise. In the experiments we performed in this paper the double mutants show conditional, temperature sensitive defects and were generated in permissive conditions where there was little selection for suppressors/modifiers. The principal advantage of QFA over SGA and EMAP is that QFA provides more accurate fitness measurements that can be measured at higher culture densities. The accuracy of QFA is indicated by the tight clustering of genes affecting particular biochemical pathways/functions in Figure 3.
QFA is also lower throughput than “bar code” based assays where up to 6000 independent strains compete in a single culture [44]. One principal difference between QFA and bar code competition methods is that fitness measures are absolute, rather than comparative.
Comparison of genetic interactions observed in yeast cells containing cdc13-1 or yku70Δ mutations, affecting telomeres in different ways, has generated numerous new insights into telomere biology. For example, we have identified at least 19 groups of genes, each representing a particular protein complex or biological process, that significantly affect growth of cells with telomere capping defects in different ways and these are highlighted in Figure 3. Each of these groups of genes, as well as numerous individual genes, warrant further investigation to characterize how they influence the telomere cap. In this paper we followed up just one striking observation that deletions of NMD pathway genes suppress the cdc13-1 temperature-sensitive phenotype and enhance the yku70Δ temperature sensitive phenotype. In upf2Δ strains, levels of STN1 transcripts and levels of Stn1 protein increased. Our detailed follow-up observations are consistent with the hypothesis that the NMD pathway influences Cdc13- or Yku70-mediated telomere capping through modification of Stn1 but not Ten1 levels (Figure 6).
As well as helping generate hypotheses about the roles of individual gene products at telomeres QFA will be ideal for developing, constraining and testing dynamic, systems models of the effects of complex biological processes on telomere function. Any model describing cellular growth and division as an outcome of the complex interaction of gene products e.g. [45] could usefully be parameterized and tested by QFA. We expect QFA to be widely applicable to other quantitative phenotypic screens in budding yeast and other microbial systems.
Library strains created using SGA in this study were cultured in SD/MSG media [17] with appropriate amino-acids and antibiotics added – Canavanine (final concentration, 50 µg/ml); G418 (200 µg/ml); thialysine (50 µg/ml); clonNAT (100 µg/ml); hygromycin B (300 µg/ml). Media were made lacking arginine when using canavanine and lacking lysine when using thialysine. W303 genetic background strains were cultured in YEPD (ade).
Cell lysis and western blot analysis were performed as previously described [46]. Antibody 9E10 from Cancer Research UK was used to detect the C-Myc epitope and anti-tubulin antibodies, from Keith Gull, Oxford, UK, used as loading controls.
RNA extraction and RT-PCR were performed as previously described [47]. RNA concentrations of each sample were normalized relative to the loading control, BUD6.
Chromatin immunoprecipitation was performed as previously described with minor modifications [48]. Mouse anti-myc (9E10) or goat anti-Mouse antibodies were used for the immunoprecipitations. Immunoprecipitated DNA was quantified by RT-PCR using the SYBR Green qPCR SuperMIX-UDG w/ROX kit (Invitrogen, 11744500).
Rectangular, single chamber, SBS footprint plates (omnitrays; Nunc, Thermo Fisher Scientific) were filled with 35 ml molten agar media using a Perimatic GP peristaltic pump (Jencons (Scientific) Limited, Leighton Buzzard, UK) fitted with a foot switch. 96-well plates (Greiner Bio-One Ltd.) were filled with liquid media or distilled H2O (200 µl per well) using a Wellmate plate-filler with stacker (Matrix Technologies, Thermo Fisher Scientific).
Solid agar to solid agar pinning was performed on a Biomatrix BM3-SC robot (S&P Robotics Inc., Toronto, Canada) using either 384-pin (1 mm diameter) or 1536-pin (0.8 mm diameter) pintools. Inoculation from solid agar to liquid media was performed on the Biomatrix BM3-SC robot using a 96-pin (1 mm diameter) pintool. Resuscitation of frozen strain collections (from liquid to solid agar) was performed on the Biomatrix BM3-SC robot using a 384-pin (1 mm diameter) tool. Re-array procedures were carried out using the BM3-SC robot equipped with a 96-pin rearray pintool. Dilution and spotting of liquid cultures onto solid agar plates was performed on a Biomek FX robot (Beckman Coulter (UK) Limited, High Wycombe, UK) equipped with a pintool magnetic mount and a 96-pin (2 mm diameter) pintool (V&P Scientific, Inc., San Diego, CA, USA). Both the Biomatrix BM3-SC and the Biomek FX were equipped with bar-code readers (Microscan Systems, Inc.) and the bar-codes of plates involved in each experiment were recorded in robot log-files.
All strains, strain collections oligonucleotide primers and plasmids are described in Text S1. Single gene deletion collections (a gift from C. Boone) were stored at −80°C in 384-well plates (Greiner BioOne) in 15% glycerol and when required, were thawed and pinned onto YEPD + G418 agar. Strains were then routinely pinned onto fresh YEPD + G418 agar plates approximately every two months but were re-pinned from frozen stocks after approximately 6 months. An array containing 6 replicates of 12 telomere-related genes, 14 replicates of his3Δ and 6 replicates of 37 randomly chosen genes was created from the original deletion collection (SGAv2). This array (plate 15 in our deletion mutant collections) was designed to quickly check that gene deletions with familiar phenotypes were behaving as expected and to also provide high numbers of replicates for a small number of genes (49) allowing more robust statistical analysis. This collection was SGAv2p15. Collection SGAv3 was made by re-arraying each of the 15 plates of SGAv2p15, randomly, with the exceptions that all his3Δ strains on the plate periphery [17] were not moved and genes which were in the corner area of plates in SGAv2p15 were specifically moved to non-corner positions in SGAv3.
Liquid-to-solid agar 384-format robotic spot tests were performed as follows. Colonies were inoculated from solid agar SGA plates into 96-well plates containing 200 µl appropriately supplemented liquid SD/MSG media in each well. These were grown to saturation (usually three days), without shaking, at 20°C. Cultures were resuspended, diluted approximately 1/100 in 200 µl H2O and spotted onto appropriately supplemented solid SD/MSG media plates which were incubated at different temperatures.
SGA query strains DLY5688 (cdc13-1 flanked by LEU2 and HPHMX (HygromycinR)), DLY3541 (yku70Δ::URA3) and DLY4228 (ura3::NATMX) were crossed to SDLv2p15 and SDLv3 in quadruplicate, giving eight biological replicate crosses each. Fitness of each strain under different conditions was assayed in 384-spot growth assays. As previously [20], growth at 36°C was used as an indication of failure of the SGA process or spontaneous reversion in SGA screens where cdc13-1 was the query mutation. In this study, repeats with modeled Trimmed Area >25000 after 6 days at 36°C (provided this included no more than 3 repeats for a single gene deletion) were stripped out. In each SGA experiment, a small number of strains were missing from the starting mutant array (due to mis-pinning, strains being lost, replaced etc.). These experiment-specific missing strains; together with genes affecting selection during SGA; and experiment-specific genes situated within 20 kb of SGA markers; were removed from analysis.
Solid agar plates were photographed on an spImager (S&P Robotics Inc., Toronto, Canada). The integrated camera (Canon EOS 40D) was used in manual mode with a pre-set manual focus. Manual settings were as follows: exposure, 0.25 s; aperture, F10; white balance, 3700 K; ISO100; image size, large; image quality, fine; image type, .jpg. Using the spImager software, the plate barcode number and a time stamp (date in year, month, day and time in hour, minute, second) were incorporated as the image name (e.g. DLY00000516-2008-12-24_23-59-59.jpg).
The image analysis tool Colonyzer [21] was used to quantify cell density from captured photographs. Colonyzer corrects for lighting gradients, removing spatial bias from density estimates. It is designed to detect cultures with extremely low cell densities, allowing it to capture a wide range of culture densities after dilute spotting on agar. Colonyzer is available under GPL at http://research.ncl.ac.uk/colonyzer.
We directly compared QFA of pinned 1536- colony format versus spotted 384- culture format and found that the range of normalized 384 spot fitness is approximately 4 times that estimated from 1,536 colony growth curves (Lawless et al., in prep). We also find that 384 spot fitness estimates adequately captures the strong temperature dependent growth of cdc13-1 mutants, whereas 1536-format growth estimates do not, and that analysis of growth in 384 spot format captures a much higher dynamic range of cell densities than 1536 colony format (approx 1,000 versus 20 fold, see Fig. 2, Lawless et al, 2010). For these reasons we chose to perform QFA of telomere capping mutants arrayed as 384 spotted cultures.
Strain array positions on a 384-spot layout (plate, row, column) were defined in a comma-separated text file and tracked using bar-codes reported in robot log-files. Data was stored in a Robot Object Database (ROD) as described previously [20]. Screen data is exported from ROD in tab delimited format (Table S7) ready for modeling and statistical analysis (see below).
Culture density (G) was estimated from captured photographs using the Integrated Optical Density (IOD or Trimmed Area; Table S7) measure of cell density provided by the image-analysis tool Colonyzer (Lawless et al 2010). Observed density time series were summarised with the logistic population model, which is an ODE describing self-limiting population growth. It has an analytical solution G (t):
Modelled inoculum density (G0, AU) was fixed (at 43 AU in this case), assuming that all liquid cultures reached the same density in stationary phase before water dilution and inoculation onto agar. Logistic parameter values r (growth rate, d−1) and K (carrying capacity, AU) were inferred by least squares fit to observations, using optimization routines in the SciPy Python library (code available from http://sourceforge.net/projects/colonyzer/).
For least-squares minimisation, initial guesses for K were the maximum observed cell density for that culture. For r, we constructed initial guesses by observing that G'(t) is at a maximum when t = t*:
Linearly interpolating between cell density observations we estimated the time of greatest rate of change of density. We then estimated r as:
A quantitative measure of fitness was then constructed from the optimal parameters. The particular measure we used was the product of the maximal doubling rate (MDR, doublings.d−1), which is the inverse of the doubling time and the maximal doubling potential (MDP, doublings). These phenotypes were quantified using logistic model parameter estimates as follows.
We estimate the minimum doubling time T which the cell population takes to reach a density of 2G0 (assuming that the culture is in exponential phase immediately after inoculation):
MDP is the number of divisions the culture is observed to undergo. Considering cell growth as a geometric progression:
These two phenotypes provide different information about the nature of population fitness and both of them are important, reflecting the rate of growth (MDR) and the capacity of the mutant to divide (MDP) under given experimental conditions. Our chosen measure of fitness (F = MDR×MDP) places equal importance on these two phenotypes.
To estimate GIS, F is obtained for a particular temperature for both the QFA screen of interest and a second QFA screen using a control query mutation, ura3Δ, which is assumed to be neutral under the experimental conditions, approximating wild-type fitness. Experimental and control strain fitnesses are analysed for evidence of epistatic interactions contradicting a multiplicative model of genetic independence [49] (used due to the ratio scale of the phenotype). We denote the fitness of the query (or background) mutation Fxyz, that of a typical deletion from the yeast knockout library FyfgΔ and double mutant fitnesses as Fxyz yfgΔ. Genetic independence therefore implies:and re-arranging gives:where M = Fxyz/Fura3Δ is a constant independent of the particular knockout, yfgΔ. Thus, after normalising fitnesses () so that the means across all knockouts for both the experimental (QFA, xyz yfgΔ) and control (cQFA, ura3Δ yfgΔ) mutation strains are equal to 1, evidence that is significantly different from is evidence of genetic interaction. Thus for each knockout a model is fitted of the form: where i = 1,2, j = 1,..,ni is the jth normalised fitness for treatment i (cQFA = 1, QFA = 2), µ is the mean fitness for the knockout in the control QFA, γ1 = 0, γ2 represents genetic interaction and εij is (normal, iid) random error. Typically ni is 8 (4 replicates each of SGAv2p15 and SGAv3), but is sometimes a larger multiple of 8 for strains that are repeated in the libraries (e.g. those on plate 15). The model is fitted in R using the lmList command. For each knockout the fitted value of γ2 is recorded as an estimated measure of the strength of genetic interaction (with the sign indicating suppression or enhancement) and the corresponding p-value is used as a measure of statistical significance of the effect. The p-value is corrected using the R function p.adjust to give a FDR-corrected q-value, and it is this q-value which is thresholded to give the lists of statistically significant genetically interacting strains (see Figure 2).
The R code used for the statistical analysis of data from ROD and Colonyzer is available from the authors on request and sample logistic analysis output is presented in Table S8.
Stringent lists of genetic interactors for each query mutation and growth condition (Tables S1, S2, S3, S4, S5, S6) were compiled by imposing a 5% FDR cutoff and arbitrarily removing genes with −0.5< GIS >0.5.
Raw output data and hyperlinked supplementary tables, together with detailed legends for interpretation of data files are available from: http://research.ncl.ac.uk/colonyzer/AddinallQFA/
|
10.1371/journal.ppat.0030137 | A Virtual Look at Epstein–Barr Virus Infection: Biological Interpretations | The possibility of using computer simulation and mathematical modeling to gain insight into biological and other complex systems is receiving increased attention. However, it is as yet unclear to what extent these techniques will provide useful biological insights or even what the best approach is. Epstein–Barr virus (EBV) provides a good candidate to address these issues. It persistently infects most humans and is associated with several important diseases. In addition, a detailed biological model has been developed that provides an intricate understanding of EBV infection in the naturally infected human host and accounts for most of the virus' diverse and peculiar properties. We have developed an agent-based computer model/simulation (PathSim, Pathogen Simulation) of this biological model. The simulation is performed on a virtual grid that represents the anatomy of the tonsils of the nasopharyngeal cavity (Waldeyer ring) and the peripheral circulation—the sites of EBV infection and persistence. The simulation is presented via a user friendly visual interface and reproduces quantitative and qualitative aspects of acute and persistent EBV infection. The simulation also had predictive power in validation experiments involving certain aspects of viral infection dynamics. Moreover, it allows us to identify switch points in the infection process that direct the disease course towards the end points of persistence, clearance, or death. Lastly, we were able to identify parameter sets that reproduced aspects of EBV-associated diseases. These investigations indicate that such simulations, combined with laboratory and clinical studies and animal models, will provide a powerful approach to investigating and controlling EBV infection, including the design of targeted anti-viral therapies.
| The possibility of using computer simulation and mathematical modeling to gain insight into biological systems is receiving increased attention. However, it is as yet unclear to what extent these techniques will provide useful biological insights or even what the best approach is. Epstein–Barr virus (EBV) provides a good candidate to address these issues. It persistently infects most humans and is associated with several important diseases, including cancer. We have developed an agent-based computer model/simulation (PathSim, Pathogen Simulation) of EBV infection. The simulation is performed on a virtual grid that represents the anatomy where EBV infects and persists. The simulation is presented on a computer screen in a form that resembles a computer game. This makes it readily accessible to investigators who are not well versed in computer technology. The simulation allows us to identify switch points in the infection process that direct the disease course towards the end points of persistence, clearance, or death, and identify conditions that reproduce aspects of EBV-associated diseases. Such simulations, combined with laboratory and clinical studies and animal models, provide a powerful approach to investigating and controlling EBV infection, including the design of targeted anti-viral therapies.
| Computer simulation and mathematical modeling are receiving increased attention as alternative approaches for providing insight into biological and other complex systems [1]. An important potential area of application is microbial pathogenesis, particularly in cases of human diseases for which applicable animal models are lacking. To date, most simulations of viral pathogenesis have tended to focus on HIV [2–7], and employ mathematical models based on differential equations. None have addressed the issue of acute infection by the pathogenic human herpes virus Epstein–Barr virus (EBV) and its resolution into lifetime persistence. With the ever-increasing power of computers to simulate larger and more complex systems, the possibility arises of creating an in silico virtual environment in which to study infection. We have used EBV to investigate the utility of this approach. EBV is a human pathogen, associated with neoplastic disease, that is a paradigm for understanding persistent infection in vivo and for which a readily applicable animal model is lacking (reviewed in [8,9]). Equally important is that EBV infection occurs in the lymphoid system, which makes it relatively tractable for experimental analysis and has allowed the construction of a biological model of viral persistence that accounts for most of the unique and peculiar properties of the virus [10,11]. We are therefore in a position to map this biological model onto a computer simulation and then ask how accurately it represents EBV infection (i.e., use our knowledge of EBV to test the validity of the simulation) and whether the matching of biological observation and simulation output provides novel insights into the mechanism of EBV infection. Specifically, we can ask if it is possible to identify critical switch points in the course of the disease where small changes in behavior have dramatic effects on the outcome. Examples of this would be the switch from clinically silent to clinically apparent infection and from benign persistence to fatal infection (as occurs in fatal acute EBV infection and the disease X-linked lymphoproliferative [12], for example), or to clearance of the virus. Indeed, is clearance ever possible, or do all infections lead inevitably to either persistence or death? Such an analysis would be invaluable. Not only would it provide insight into the host–virus balance that allows persistent infection, but it would also reveal the feasibility and best approaches for developing therapeutic interventions to diminish the clinical symptoms of acute infection, prevent fatal infection, and/or clear the virus.
A diagrammatic version of the biological model is presented in Figure 1. EBV enters through the mucosal surface of the Waldeyer ring, which consists of the nasopharyngeal tonsil (adenoids), the paired tubal tonsils, the paired palatine tonsils, and the lingual tonsil arranged in a circular orientation around the walls of the throat. Here EBV infects and is amplified in epithelium. It then infects naïve B cells in the underlying lymphoid tissue. The components of the ring are all equally infected by the virus [13]. EBV uses a series of distinct latent gene transcription programs, which mimic a normal B cell response to antigen, to drive the differentiation of the newly infected B cells. During this stage, the infected cells are vulnerable to attack by cytotoxic T cells (CTLs) [14]. Eventually, the latently infected B cells enter the peripheral circulation, the site of viral persistence, as resting memory cells that express no viral proteins [15] and so are invisible to the immune response. The latently infected memory cells circulate between the periphery and the lymphoid tissue [13]. When they return to the Waldeyer ring they are occasionally triggered to terminally differentiate into plasma cells. This is the signal for the virus to begin replication [16], making the cells vulnerable to CTL attack again [14]. Newly released virions may infect new B cells or be shed into saliva to infect new hosts, but are also the target of neutralizing antibody.
Primary EBV infection in adults and adolescents is usually symptomatic and referred to as infectious mononucleosis (AIM). It is associated with an initial acute phase in which a large fraction (up to 50%) of circulating memory B cells may be latently infected [17]. This induces the broad T lymphocyte immune response characteristic of acute EBV infection. Curiously, primary infection early in life is usually asymptomatic. In immunocompetent hosts, infection resolves over a period of months into a lifelong persistent phase in which ∼1 in 105 B cells carry the virus [18]. Exactly how persistent infection is sustained is unclear. For example, once persistence is established, it is unknown if the pool of latently infected memory B cells is self-perpetuating or if a low level of new infection is necessary to maintain it. Indeed, we do not know for sure that the pool of latently infected B cells in the peripheral memory compartment is essential for lifetime persistence. It is even unclear whether the virus actually establishes a steady state during persistence or continues to decay, albeit at an ever slower rate [17].
In the current study we describe the creation and testing of a computer simulation (PathSim) that recapitulates essential features of EBV infection. The simulation has predictive power and has utility for experiment design and understanding EBV infection. One practical limitation of available simulation and modeling approaches has been their inaccessibility to the working biologist. This is often due to the use of relatively unfamiliar computer interfaces and output formats. To address these issues, we have presented the simulation via a user-friendly visual interface on a standard computer monitor. This allows the simulation to be launched and output to be accessed and analyzed in a visual way that is simple and easily comprehensible to the non-specialist.
The computer model (PathSim) is a representation of the biological model described in the Introduction. A schematic version of both is shown in Figure 1. To simulate EBV infection, we created a virtual environment consisting of a grid that describes a biologically meaningful topography, in this case the Waldeyer ring (five tonsils and adenoids) and the peripheral circulation, which are the main sites of EBV infection and persistence. The tonsils and adenoids were composed of solid hexagonal base units representing surface epithelium, lymphoid tissue, and a single germinal center/follicle (Figure 2A–2C; Video S1). Each hexagonal unit had one high endothelial venule (HEV) entry point from the peripheral blood and one exit point into the lymphatic system (Figure 2A).
Discrete agents (cells or viruses) reside at the nodes (red boxes) of the 3-D grid (white lines). There they can interact with other agents and move to neighboring nodes. Agents are assessed at regular, specified time intervals as they move and interact upon the grid. Virtual cells were allowed to leave the Waldeyer ring via draining lymphatics and return via the peripheral blood and HEVs (Figure 2A and 2B; Video S1) as in normal mucosal lymphoid tissue [19]. A brief summary list of the agents employed in our simulation, and their properties and interactions, is given in Table 1. In this report we refer to actual B cells as, for example, “B cells”, “latently infected B cells”, or “lytically infected B cells”, and their virtual representations as “virtual B cells”, “BLats”, or “BLyts”. Similarly, we refer to actual virus as virions and their virtual counterparts as virtual virus or Vir. A full description of the simulation, including a complete list of agents, rules, the default parameters that produce the output described below, and a preliminary survey of the extended parameter space is presented in M. Shapiro, K. Duca, E. Delgado-Eckert, V. Hadinoto, A. Jarrah, et al. (2007) A virtual look at Epstein-Barr virus infection: simulation mechanism (unpublished data). Here, we will first present a description of how the virtual environment was visualized and then focus on a comparison of simulation output with the known biological behavior of the virus.
Simulation runs were accessed through an information-rich virtual environment (IRVE) (Figures 2 and 3; Videos S1 and S2), which was invoked through a Web interface. This provided a visually familiar, straightforward context for immediate comprehension of the spatial behavior of the system [20]. It also allowed specification of parameters, run management, and ready access to data output and analysis.
Figure 3 demonstrates how the time course of infection may be visualized. Usually the simulation was initialized by a uniform distribution of Vir over the entire surface of the Waldeyer ring, thereby seeding infection uniformly. However, in the simulation shown in Figure 3A, virtual EBV was uniformly deposited only on the lingual tonsil. Figure 3B–3D shows the gradual spread of virtual infection (intensity of red color indicating the level of free Vir) to the adjacent tonsils. It can be seen in this case that the infection spreads uniformly to all the tonsils at once, implying that it was spreading via BLats returning from the blood compartment and reactivating to become BLyts, rather than spreading within the ring. Examples of infectious spread between and within the tonsils can also be seen in Video S2.
In this paper we present a comprehensive model of EBV infection that effectively simulates the overall dynamics of acute and persistent infection. The fact that this simulation can be tuned to produce the course of EBV infection suggests that it models the basic processes of this disease. To achieve this, we have created a readily accessible, virtual environment that appears to capture most of the salient features of the lymphoid system necessary to model EBV infection. Achieving infection dynamics that reflect an acute infection followed by recovery to long-term low-level persistent infection seems to require access of the virus to a blood compartment where it is shielded from immunosurveillance. Because we cannot perform a comprehensive parameter search (due to the very large parameter space involved), we cannot unequivocally state that the blood compartment is essential. What is clear though, is that persistence is a very robust feature in the presence of a blood compartment, and that we could not achieve an infection process that even remotely resembles typical persistent EBV infection in its absence.
The areas in which the simulation most closely follows known biology are summarized in Table 2 and include the peak time of infection, 33–38 d, compared to the incubation time for AIM of 35–50 d [21]. This predicts that patients become sick and enter the clinic at or shortly after peak infection in the peripheral blood, a prediction confirmed by our patient studies, where the numbers of infected B cells in the periphery always decline after the first visit [17].
An important feature of a simulation is its predictive power. Our analysis predicted that access to the peripheral memory compartment is essential for long-term persistence. This is consistent with recent studies on patients with hyper-IgM syndrome [31]. Although these individuals lack classical memory cells, they can be infected by EBV; however, they cannot sustain persistent infection and the virus becomes undetectable. Unfortunately, those studies did not include a sufficiently detailed time course to see if time to virus loss coincided with the simulation prediction of 1–2 mo.
Another area where the simulation demonstrated its predictive power was in the dynamics of viral replication. In the simulation it was unexpectedly observed that the level of Vir production plateaued long before BLats, predicting that the levels of virus shedding, unlike latently infected cells, will have leveled off by the time AIM patients arrive in the clinic. This prediction, which contradicted the common wisdom that virus shedding should be high and decline rapidly in AIM patients, was subsequently confirmed experimentally (V. Hadinoto, M. Shapiro, T. Greenough, J. Sullivan, K. Luzuriaga, D. Thorley-Lawson (2007) On the dynamics of acute EBV infection and the pathogenesis of infectious mononucleosis (unpublished data) and see also [22,23]). The simulation also quite accurately reproduces the relatively large variation in virus production over time, compared to the stability of B latent. This difference is likely a consequence of stochasticity (random variation) having a relatively larger impact on virus production. This is because the number of B cells replicating the virus at any given time is very small, both in reality and the simulation, compared to the number of infected B cells, but the number of virions they release when they do burst is very large. This difference may reflect on the biological requirements for persistence of the virus since a transient loss in virus production due to stochasticity can readily be overcome through recruitment from the pool of B latents. However, a transient loss of B latents would mean clearance of the virus. Hence, close regulation of B latent but not virion levels is necessary to ensure persistent infection.
Although there is now a growing consensus that EBV infects normal epithelial cells in vivo [27–29], the biological significance of this infection remains unclear. The available evidence suggests that epithelial cell infection may not be required for long-term persistence [25,26], and this is also seen in the simulation. The alternate proposal is that epithelial infection might play an important role in amplifying the virus, during ingress and/or egress, as an intermediary step between B cells and saliva. This is based on the observation that the virus can replicate aggressively in primary epithelial cells in vivo [30]. In the simulation, epithelial amplification had no significant effect on the ability of Vir to establish persistence. This predicts that epithelial amplification does not play a critical role in entry of the virus, but leaves open the possibility that it may be important for increasing the infectious dose present in saliva for more efficient infection of new hosts.
The simulation is less accurate in the precise quantitation of the dynamics. Virtual acute infection resolves significantly more slowly and persistence is at a higher level than in a real infection. In addition, virtual persistent infection demonstrates clear evidence of oscillations in the levels of infected cells that have not been detected in a real infection. The most likely explanation for these discrepancies is that we have not yet implemented T cell memory. Thus, as the levels of virtual infected cells drops, the immune response weakens, allowing Vir to rebound while a new supply of virtual CTLs is generated. Immunological memory would allow a more sustained T cell response that would produce a more rapid decline of infected cells, lower levels of sustained persistence, and tend to flatten out oscillatory behavior, thus making the simulation more quantitatively accurate. This is one of the features that will be incorporated into the next version of our simulation. It remains to be determined what additional features need to be implemented to sharpen the model and also whether and to what extent the level of representation we have chosen is necessary for faithful representation of EBV infection. Our simulation of the Waldeyer ring and the peripheral circulation was constructed with the intent of modeling EBV infection. Conversely, our analysis can be thought of as the use of EBV to validate the accuracy of our Waldeyer ring/peripheral circulation simulation and to evaluate whether it can be applied to other pathogens. Of particular interest is the mouse gamma herpesvirus MHV68 [32,33]. The applicability of MHV68 as a model for EBV is controversial. Although it also persists in memory B cells [34], it appears to lack the sophisticated and complicated latency states that EBV uses to access this compartment. However, one of the simplifications in our simulation is that the details of these different latency states and their transitions are all encompassed within a single concept, the BLat. We have also assumed a time line whereby a newly infected BLat becomes activated and CTL sensitive, migrates to the follicle, and exits into the circulation, where it is no longer seen by our virtual CTLs. In essence, we have generalized the process by which the virus proceeds from free virion to the site of persistence in such a way that it may be applicable to both EBV and MHV68. Thus, we might expect that the overall dynamics of infection may be similar even though detailed biology may vary. As a first step to test if this concept had value, we performed an analysis based on studies with MHV68 where it was observed that the levels of infected B cells at persistence were unaffected by the absolute amount of input virus at the time of infection [35]. When this parameter was varied in the simulation, we saw the same outcome. This preliminary attempt raises the possibility that the mouse virus may be useful for examining quantitative aspects of EBV infection dynamics.
The last area we wished to investigate was whether we could identify biologically meaningful “switch” points, i.e., places in time and space where relatively small changes in critical parameters dramatically affect outcome, for example, switching from persistence to clearance to death. We have observed one such switch point—reactivation of BLats upon return to the Waldeyer ring—that rapidly switches the infection process from persistence to death. How this might relate to fatal EBV infection, X-linked lymphoproliferative disease, is uncertain. However, viral production is a function both of how many B cells initiate reactivation and how efficiently they complete the process. We believe that most such cells are killed by the immune response before they release virus [16], so defects in the immune response could allow more cells to complete the viral replication process and give the same fatal outcome.
The ability to find such conditions for switch points could be very useful in the long term for identifying places in the infection process where the virus might be optimally vulnerable to drug intervention. The easiest place to target EBV is during viral replication; however, it is currently unclear whether viral replication and infection are required for persistence. It may be that simply turning off viral replication after persistence is established fails to eliminate the virus because the absence of new cells entering the pool through infection is counterbalanced by the failure of infected cells to disappear through reactivation of the virus. If, however, a drug allowed abortive reactivation, then cells would die without producing infectious virus and new infection would be prevented. This models the situation that would arise with a highly effective drug or viral mutant that blocked a critical stage in virion production (e.g., viral DNA synthesis or packaging), so that reactivation caused cell death without release of infectious virus. A similar effect could be expected with a drug or vaccine that effectively blocked all new infection. This is another case in which studies with the mouse virus, where non-replicative mutants can be produced and tested, may be informative as to whether and to what extent infection is required to sustain the pool of latently infected B cells and persistence. The simulation could then be used to predict how effective an anti-viral that blocked replication, or a vaccine that induced neutralizing antibodies, would need to be at reducing new infection in order to cause EBV to be lost from the memory pool (for a more detailed discussion of this issue see M. Shapiro, K. Duca, E. Delgado-Eckert, V. Hadinoto, A. Jarrah, et al. (2007) A virtual look at Epstein-Barr virus infection: simulation mechanism (unpublished data)).
Most modeling of virus infection to date has tended to focus on HIV and use differential equations [2–7]. One such study involved EBV infection [36], but to our knowledge none outside of our group has addressed the issue studied here of acute EBV infection and how it resolves into lifetime persistence. In preliminary studies of our own, modeling EBV infection with differential equations that incorporate features common to the HIV models, with parameters physiologically reasonable for EBV did not produce credible dynamics of infection (K. Duca, unpublished observations). Although we do not exclude the possibility that such models may be useful for simulating EBV, we took an agent-based approach because it is intuitively more attractive to biologists. Such models are increasingly being recognized as an effective alternative way to simulate biological processes [37–39] and have several advantages. The main advantage is that the “agent” paradigm complies by definition with the discrete and finite character of biological structures and entities such as organs, cells, and pathogens. This makes it more accurate, from the point of view of scientific modeling. It is also less abstract since the simulated objects, processes, and interactions usually have a straightforward biological interpretation and the spatial structure of the anatomy can be modeled meticulously. The stochasticity inherent to chemical and biological processes can be incorporated in a natural way. Lastly, it is generally much easier to incorporate qualitative or semi-quantiative information into rule sets for discrete models than it is for such data to be converted to accurate rate equations. The major drawback to agent-based models is that there is currently no mathematical theory that allows for rigorous analysis of their dynamics. Currently, one simply runs such simulations many times and performs statistical analyses to assess their likely behaviors. Developing such a mathematical theory remains an important goal in the field.
In summary, we have described a new computer simulation of EBV infection that captures many of the salient features of acute and persistent infection. We believe that this approach, combined with mouse modeling (MHV68) and EBV studies in patients and healthy carriers, will allow us to develop a more profound understanding of the mechanism of viral persistence and how such infections might be treated and ultimately cleared.
Details of the AIM patient populations tested have been published previously [17]. Adolescents (ages 17–24) presenting to the clinic at the University of Massachusetts at Amherst Student Health Service (Amherst, Massachusetts, United States) with clinical symptoms consistent with acute infectious mononucleosis were recruited for this study. Following informed consent, blood and saliva samples were collected at presentation and periodically thereafter. Diagnosis at the time of presentation to the clinic required a positive monospot test and the presence of atypical lymphocytes [21]. Confirmation of primary Epstein–Barr infection required the detection of IgM antibodies to the EBV viral capsid antigen in patient sera [40]. These studies were approved by the Human Studies Committee at the University of Massachusetts Medical School (Worcester, Massachusetts, United States) and by the Tufts New England Medical Center and Tufts University Health Sciences Institutional Review Board. All blood samples were diluted 1:1 in 1x PBS.
The technique for estimating the absolute number of latently infected B cells in the peripheral blood of patients and healthy carriers of the virus is a real-time PCR–based variation of our previously published technique [17], the details of which will be published elsewhere (V. Hadinoto, M. Shapiro, T. Greenough, J. Sullivan, K. Luzuriaga, et al. (2007) On the dynamics of acute EBV infection and the origins of infectious mononucleosis (unpublished data)). To measure the absolute levels of virus shedding in saliva, individuals were asked to rinse and gargle for a few minutes with 5 ml of water and the resultant wash processed for EBV-specific DNA PCR using the same real-time–based PCR technique. We have performed extensive studies to standardize this procedure that will be detailed elsewhere (V. Hadinoto, M. Shapiro, T. Greenough, J. Sullivan, K. Luzuriaga, et al. (2007) On the dynamics of acute EBV infection and the origins of infectious mononucleosis (unpublished data)).
In the simulation, B cells are either uninfected (BNaïve), latently infected (BLat), or replicating virtual virus (BLyt); we do not distinguish blast and memory B cells. In the biological model, newly infected B cells in the lymphoepithelium of the Waldeyer ring pass through different latency states, which are vulnerable to attack by cytotoxic T cells (CTL latent). Subsequently, they become memory B cells that enter the peripheral circulation and become invisible to the immune response by turning off viral protein expression. In the simulation, all these latency states are captured in the form of a single entity, the BLat. In addition, the blood circulation and lymphatic system are both represented as abstract entities that only allow for transport of BNaïves and BLats around the body. Virtual T cells are restricted to the Waldeyer ring. This simplification is based on the assumption that, in the biological model, EBV-infected cells entering the peripheral circulation are normal and invisible to CTLs, because the virus is inactive, and therefore the peripheral circulation simply acts as an independent pool of and a conduit for B latent. Operationally, therefore, BLats escape TLats in the simulation simply by entering the peripheral circulation. Consequently, unlike the biological model, BLats are vulnerable to TLats whenever they reenter the lymph node.
Each agent (e.g., Vir or a BNaïve) has a defined life span, instructions for movement, and functions that depend on which other agents they encounter (for example, if a Vir encounters a BNaïve, it infects it with some defined probability). The agents, rules, and parameters used are based on known biology wherever possible with simplifications (see above) where deemed appropriate. A brief description and discussion of the agents and their rules is given in Table 1. A detailed listing is provided in M. Shapiro, K. Duca, E. Delgado-Eckert, V. Hadinoto, A. Jarrah, et al. (2007) A virtual look at Epstein-Barr virus infection: simulation mechanism (unpublished data). At each time point (6 min of real time), every agent is evaluated and appropriate actions are initiated.
The simulation is invoked through a Web interface (IRVE; see movies linked to Figure 2, and [20]) that allows a straightforward visual, familiar, and scalable context for access to parameter specification, run management, data output, and analysis. This has the additional advantage that it readily allows comprehension of the spatial behavior of the system (e.g., “how does the infection spread?”). The simulation may also be invoked from the command line. Through the Web, users can process simulation data for output and analysis by a number of common applications such as Microsoft's Excel, University of Maryland's TimeSearcher [41], and MatLab.
We have developed display components that encapsulate multiple-view capabilities and improved multi-scale interface mappings. The IRVE is realized in the international standard VRML97 language. The simulation can be rerun and reanalyzed using a normal VCR-type control tool, which allows the operator, for example, to fast forward, pause, rewind, or drag to a different time point, and to play back runs or analyze simulation output dynamically. In the IRVE, any spatial object (including the global system) can be annotated with absolute population numbers (as a time plot and/or numeric table) or proportional population numbers (as a bar graph) for any or all of the agents. Spatial objects themselves can be animated by heat-map color scales. The intensity of the color associated with each agent is a measure of the absolute level of the agent; so, for example, as the level of free Vir increases, so will the level of intensity of the associated color (in this case red) both within the single units and in the entire organ.
In our simulation we manage multiple views of the dynamic population values through a higher order annotation called a PopView (population view). A PopView is an interactive annotation that provides three complementary representations of the agent population. The representations can be switched through in series by simple selection. The default view is a color-coded bar graph where users can get a quick, qualitative understanding of the agent populations in a certain location at that time step. The second is a field-value pair text panel, which provides numeric readouts of population levels at that time step. The third is a line graph where the population values for that region are plotted over time.
Because of the large amount of time points and the large number of grid locations, the IRVE manages an integrated information environment across two orders of magnitude: “Macro” and “Micro” scales. Through the standard VRML application the user has a number of options including free-navigational modes such as: fly, pan, turn, and examine. This allows users to explore the system, zooming in and out of anatomical structures as desired. In addition, the resulting visualization space is navigable by predefined viewpoints, which can be visited sequentially or randomly through menu activation. This guarantees that all content can be accessible and users can recover from any disorientation. The Visualizer manages Macro and Micro scale result visualizations using proximity-based filtering and scripting of scene logic. As users approach a given anatomical structure, the micro-scale meshes and results are loaded and synchronized to the time on the users' VCR controller. |
10.1371/journal.pcbi.1002235 | Grid Cells, Place Cells, and Geodesic Generalization for Spatial Reinforcement Learning | Reinforcement learning (RL) provides an influential characterization of the brain's mechanisms for learning to make advantageous choices. An important problem, though, is how complex tasks can be represented in a way that enables efficient learning. We consider this problem through the lens of spatial navigation, examining how two of the brain's location representations—hippocampal place cells and entorhinal grid cells—are adapted to serve as basis functions for approximating value over space for RL. Although much previous work has focused on these systems' roles in combining upstream sensory cues to track location, revisiting these representations with a focus on how they support this downstream decision function offers complementary insights into their characteristics. Rather than localization, the key problem in learning is generalization between past and present situations, which may not match perfectly. Accordingly, although neural populations collectively offer a precise representation of position, our simulations of navigational tasks verify the suggestion that RL gains efficiency from the more diffuse tuning of individual neurons, which allows learning about rewards to generalize over longer distances given fewer training experiences. However, work on generalization in RL suggests the underlying representation should respect the environment's layout. In particular, although it is often assumed that neurons track location in Euclidean coordinates (that a place cell's activity declines “as the crow flies” away from its peak), the relevant metric for value is geodesic: the distance along a path, around any obstacles. We formalize this intuition and present simulations showing how Euclidean, but not geodesic, representations can interfere with RL by generalizing inappropriately across barriers. Our proposal that place and grid responses should be modulated by geodesic distances suggests novel predictions about how obstacles should affect spatial firing fields, which provides a new viewpoint on data concerning both spatial codes.
| The central problem of learning is generalization: how to apply what was discovered in past experiences to future situations, which will inevitably be the same in some respects and different in others. Effective learning requires generalizing appropriately: to situations which are similar in relevant respects, though of course the trick is determining what is relevant. In this article, we quantify and investigate relevant generalization in the context of a particular learning problem often studied in the laboratory: learning to navigate in a spatial maze. In particular, we consider whether the brain's well-characterized systems for representing an organism's location in space generalize appropriately for this task. Our simulations of learning verify that to generalize effectively, these representations should treat nearby locations similarly (that is, neurons should fire similarly when an animal occupies nearby locations)—but, more subtly, that to enable successful learning, “nearby” must be defined in terms of paths around obstacles, rather than in absolute space “as the crow flies.” These considerations suggest new principles for understanding these spatial representations and why they appear warped and distorted in environments, such as mazes, with barriers and obstacles.
| The rodent brain contains at least two representations of spatial location. Hippocampal place cells fire when a rat passes through a confined, roughly concentric, region of space [1], whereas the grid cells of dorsomedial enthorhinal cortex (dMEC) discharge at the vertices of regular triangular lattices [2]. Behaviorally, such codes likely support decisions about spatial navigation [3]–[7], and more particularly reinforcement learning (RL [8]) or learning by trial and error where to navigate.
Here we investigate the appropriateness of the brain's spatial codes for learning value functions, guided by the influential use of RL models across many varieties of decision problems in computational neuroscience [9]–[11]. Although much work in these systems tends to focus on the “upstream” mechanisms by which place or grid fields are constructed from different sorts of inputs, we focus instead on learning downstream from these representations (e.g., where place cells synapse on striatal neurons), to ask what does this function suggest about or require from the spatial representations. This provides a complementary perspective on aspects of the neural responses, which, we argue, are well adapted to support reinforcement learning.
Importantly, this exercise views the brain's spatial codes less as a representation for location per se, and instead as basis sets for approximating other functions across space. In particular, most RL models work by learning to represent a value function over state space – a mapping of location to value. The value function measures the proximity of locations to rewards, and in this way can guide navigation towards reinforcement. Although a frequency-domain Fourier basis (often analogized to the grid representation [12], [13]) and a space-domain impulse basis (an idealized place map) are both complete representations for arbitrary functions over space, efficient RL—in the sense of rapid generalization from few experiences—depends on the features of the basis being well matched to the function being learned [14]–[17]. For instance, just as efficient visual representations are motivated by the fact that the Fourier decompositions of natural images have most of their power at low frequencies, so also value functions tend to change smoothly across space: if a given location is near reward, then so are nearby positions.
Thus, it is intuitive (and our simulations, below, verify) that low-frequency basis functions can speed up spatial RL by allowing experience about rewards to generalize over larger distances. However, we argue that considering generalization in the RL setting suggests a crucial and underappreciated refinement of this idea: in general, value functions are not maximally smooth over space “as the crow flies” (i.e. Euclidean distance). Instead, value functions exhibit discontinuities at obstacles, such as walls, which help to guide navigation around them. Building on a variety of work applying graph-theoretic distance metrics to different problems in machine learning [14], [15], [17], [18], much work in reinforcement learning [14]–[17] suggests that the demand of efficient generalization for navigation implies that basis functions—here, place or grid fields—should modulate their strength according to geodesic distance (i.e. the shortest navigable path between two points, around obstacles) rather than Euclidean.
We formalize this idea in a model of grid and place cell responses. The model and its simulations suggest novel predictions about how grid cell and place cell firing fields should behave in the presence of obstacles and other navigational constraints: in effect, these should locally warp the geometry of the representation. These predictions offer a new perspective on existing results, such as the unidirectionality of place fields on the linear track [19]–[23] and the behavior of grid cells in mazes [24].
First, we used TD(λ) learning in three simple environments (Figure 1A) to test the ability of multiscale grid cell- and place cell-like basis sets to learn value functions in spatial RL (see Materials and Methods). In order to verify the importance of generalization over long spatial scales, we compared learning with the modeled grid and place cell bases to a standard, tabular RL basis learning the same task. This is like a place cell basis using only a single, fixed scale of representation that is small with respect to the task-relevant distances. The simulated agent had to learn to navigate from a randomly chosen starting point to a goal state that contained a reward. To quantify performance, the number of steps needed to reach the reward was plotted as a function of the training trial. Although our key qualitative points are robust to changes in the free parameters (simulations not shown), to ensure a fair comparison we optimized the learning rate (a crucial free parameter) separately for each condition (i.e. basis function and gridworld) to obtain its best performance. We additionally used the TD(λ) generalization of TD with a high value (0.9) of the eligibility trace parameter λ, since this provides another mechanism for learning to generalize along trajectories and might, in principle, help to compensate for the shortcomings of the tabular or Euclidean bases.
As Figure 1B shows, the grid and place cell basis sets drastically quicken learning the value function compared to the tabular code, demonstrating the benefits of spatial generalization. Figure 2 illustrates the approximated value functions at different stages of learning and qualitatively shows the importance of generalization. In particular, the tabular basis does not take advantage of the spatial structure to generalize quickly and must learn each state's value separately from its neighbors by a slow process of TD chaining.
Figure 2 also hints at a subtler problem of overgeneralization in Euclidean space. In particular, these grid and place cell basis functions tend to smear the value function across barriers, where it should change sharply (arrows in Figures 2B and 2C, where the effects are most apparent). Because of this, value is underrepresented at states inside the walls (i.e. locations closer to the reward, as in 2B) and overrepresented on the other side of the barrier (most visible in 2C). This distortion remains at asymptote and is likely not an artifact of insufficient experience.
While this flaw does not notably degrade performance in these simple tasks, it can be detrimental when fine navigational precision is required. To demonstrate this, we tested the models in three environments that required the agent to navigate narrow halls or openings, and thus learn precise state value representations (Figure 3A). Here, the grid cell and place cell basis functions performed poorly, and were outperformed by the tabular basis (Figure 3B). Together, then, these simulations demonstrate that generalization due to spatial representations like those seen in the brain can help make reinforcement learning more efficient, but also that such generalization has drastic (and, presumably, behaviorally unrealistic) side effects, abolishing learning in tasks where paths are narrow.
In general, as can be seen directly in the recursive definition of the value function, (Equation 1 in Materials and Methods), the extent to which values are related between two states depends on how closely they are connected by the state-state transition probability function. Accordingly, work on value function approximation for reinforcement learning has proposed [14]–[17] that basis functions should be constructed to respect distance along the state transition graph. For instance, in temporal prediction tasks, value functions are smooth in time [61]. In a spatial task, the transition dynamics imply that states have similar values when they are near each other, but near as measured in geodesic (along-path) distance, rather than “as the crow flies” (Euclidean distance). Formally, geodesic distance measures the number of steps along the transition graph needed to get from one state to another. A basis over geodesic distances would treat states separated by a boundary as comparatively far apart, enabling their values to be discontinuous, whereas the Euclidean basis used above (and ubiquitously to characterize the spatial extent of place and grid fields) would inappropriately treat them as adjacent.
These considerations suggest that for efficacious representation of value functions over state space, the brain should adopt basis functions that are smooth along geodesic rather than Euclidean distances. In the open field there should be no difference between geodesic and Euclidean representations, since these metrics coincide there. However, if an environment has barriers, then Euclidean and geodesic firing fields will differ. The effect of such a difference should be to introduce geometric distortion into geodesic firing fields nearby obstacles, where geodesic and Euclidean metrics differ. Such a distortion can be characterized (and indeed implemented) by mapping the original Euclidean vector coordinates through an additional transform that accounts for geodesic distance. However, in the present work our goal is to investigate the brain's spatial representations through the lens of their downstream computations; thus, in contrast to much work on the hippocampal system [12], [35], [36], [40], [43], [44], [47], [48], [62], [63] we do not focus on the “upstream” computations by which the grid or place representations (or their hypothesized distortions) are themselves computed from inputs. That is, we take geodesic or Euclidean representations as a given and focus our analysis on hypothesized learning that relies on entorhinal and hippocampal outputs.
In particular, we modeled how basis functions would appear in environments with barriers, if they followed a geodesic metric, by evaluating Euclidean grid or place fields (characterized by spatial grids or Gaussians) over a new set of x–y coordinates, chosen such that their pairwise Euclidean distances approximated the states' geodesic distances (see Materials and Methods). When viewed in the original Euclidean space, the effect of barriers is to produce geometric distortions, such as variations in grid orientation and firing field shapes (Figure 4). As one might expect, the basis functions tend not to cross walls and instead skirt along connected paths.
We tested the geodesic bases in the environments that stressed importance of along-path generalization (Figure 3A). As can be seen, the geodesic bases alleviated the poor learning caused by the indiscriminate generalization of their Euclidean counterparts (Figure 3B). Since the geodesic grid cells and place cells generalize using the state transition graph, they learn at least as fast as the tabular TD control (Figure 3B). Figure 5A–C depicts typical value functions at different stages of training using the geodesic basis functions (25 trials for Figure 5A–B, 50 trials for Figure 5C). Also note that both the Euclidean and geodesic bases used the same multiple granularities and tiling, with the sole difference the distance metric used. To test the role of multiple tilings in learning, we performed follow-up simulations for each of the six gridworlds using three different tiles bases. While the tile bases often learned faster than the tabular basis (which one would expect), overall the geodesic bases tended to perform best (data not shown). Together, these simulations demonstrate the representation benefits conferred by geodesic generalization, in particular how generalization along paths rather than across walls solves the problem of overgeneralization interfering with learning in the presence of obstacles. That the same qualitative results hold up using both grid-cell-like and place-cell-like representations points to their generality. In simulations not shown here, we also produced similar results using an overlapping tile code at a variety of single scales [8], suggesting that the results relate to spatial generalization per se and not to the multiscale nature of the (biologically inspired) bases used here.
The foregoing simulations suggest that to support efficient navigation, the brain's spatial representations should generalize according to a geodesic rather than a Euclidean metric. Of course, these two representations coincide in the open field, where most studies have been conducted. However, we believe our model's predictions are consistent with a number of studies where researchers recorded neurophysiological activity while rats foraged in environments containing barriers. Here we compare our model to examples from three studies [24], [64], [65].
Skaggs & McNaughton [64] recorded place cells as rats moved between two separate enclosures that were connected by a narrow corridor (schematized in Figure 6, top; cf. Figure 4 in [64]). Although this was not the major experimental question of the study, the narrow corridor provides a good test for our model's prediction that place fields should track along paths rather than (as a Euclidean place field predicts) across barriers. In the examples reproduced here, for instance, place cell spikes are almost exclusively confined to either the connecting corridor's entrance (Figure 6A, left) or the pathway between the two rooms (Figure 6A, right). The spikes do not generalize across the walls separating parts of the environment, but instead appear to track along paths around them (Figure 6), even though a standard isotropic Gaussian place field over Euclidean coordinates would clearly not respect these barriers. The data are, however, similar to place field responses from the geodesic model in a similar environment (Figure 6, bottom).
In another study [65], a place field was first recorded in an open box and again after adding a barrier to the enclosure (Figure 7A; cf. to Figure 8 in [65]). Recorded hippocampal place cell responses in the open field vanished immediately when the firing field was bisected by a wall [65], Figure 6A. The geodesic model of neural spatial representation provides an elegant, intuitive account for why the place field disappears, whose graphical intuition is displayed in Figures 7A–B. In an environment without walls, one can think of the recorded place cell activity being measured over evenly spaced locations in 2D enclosure (Figure 7A, left). Once a barrier is introduced that bisects the field, the nearby locations on adjacent sides of the wall are pulled apart, which changes the spacing between neighboring points compared to its Euclidean counterpart. Locations on either side of the wall are far, in geodesic terms, from each other, and from the center of a place field centered in the wall itself. As a result, a sinkhole is created that swallows the place field in the geodesic coordinate space, thus muting its activity (Figure 7A–B).
Similar results were also seen in a recent study of how place cell firing fields changed when mazes were reconfigured [66]. In particular, this work replicated the phenomenon of place fields diminishing or disappearing near newly introduced obstacles, and verified (as in our simulations) that such changes predominate near newly introduced obstacles. The study also demonstrates a rarer, complementary phenomenon whereby the introduction of obstacles caused firing to increase or even new place fields to appear, as verified in our simulations. In our model (Figure 8), increased firing is the flip side of responses diminishing for neurons coding “holes” in geodesic space; it occurs when geometric distortion “pushes” locations into areas previously off the map.
Finally, Derdikman et al. [24] recorded from grid cells as a rat ran along a hairpin maze. Figures 1 and 2 from [24] show typical grid cell firing fields in an open field and again in a hairpin maze. The standard hexagonal pattern of responding is extremely distorted; instead, responses tend to track along the hallways but not to cross walls, and firing fields are similar between alternate arms. Grid cells simulated in the geodesic space share a number of these characteristics (Figure 9), though not (as discussed below) all of them. One limitation of the model is that it does not capture the repetitive place field firing observed by Derdikman et al. [24].
Although researchers widely assume that reinforcement learning methods such as temporal difference learning subserve learned action selection in the brain [9]–[11], it is less clear how tasks involving many structured states can be represented in a way that enables these methods to learn efficiently, due in large part to the curse of dimensionality. In computer science, stylized spatial navigation (gridworld) problems are the classic domain for studying this issue, since the state space is large but transparently visualized and manipulated [8]. Here we consider rodents' neural representations of spatial location from this perspective, treating them as basis functions for downstream reinforcement learning in high-dimensional state spaces and asking how well adapted they are to this role. Though previous modeling work has not extensively considered the constraints on the brain's location codes implied by this function, much work has more or less implicitly exploited the idea that unlike the tabular basis often assumed in simple RL, the spatial extent of place fields can help to cope with the curse of dimensionality by allowing learning to generalize between nearby locations [3], [50], [51] even over multiple scales [30].
The present study extends this idea to consider such generalization in light of work on efficient representation in machine learning [14]–[17]. These theoretical considerations, illustrated and verified by our simple simulation results, suggest that to enable efficient representation of value (or other) functions over space, grid and place fields should operate in a distorted geometry: generalizing according to geodesic (on-path) rather than Euclidean (as-the-crow-flies) distances. Although these two distance metrics coincide in the open field, they differ in the presence of boundaries. The geodesic metric predicts that grid and place fields should not spill across walls but should instead track along paths, and should also exhibit geometric distortions, such as altered grid orientation, near boundaries. We have reviewed data from a number of experiments that seem largely in accord with these predictions. It should be noted that these predictions are all at the neural level, and could be most directly tested quantitatively by simply examining whether neural firing is modulated more reliably with distances measured by either metric: e.g., regressing distance (computed according to either definition) from a place field's center on firing rate.
By contrast, since our argument is primarily one about learning efficiency (which is difficult to quantify behaviorally, since it is affected by many factors), our model does not make categorical behavioral predictions. Our simulations (Figure 3) demonstrate that simple TD models with Gaussian place fields (like that of [3]) can entirely fail to solve simple navigation problems involving narrow apertures or hallways. However, the fact that rats do not exhibit such problems of course does not by itself demonstrate that the brain adopts the same solution for this problem as the one we propose. Also, to focus on our main questions of interest, we omit many features that other models use to explain various behavioral phenomena of navigation, among them mechanisms for allocentric route-planning (important for quick goal learning [3] and for planning shortcuts [67]) and localization driven by combinations of cues and path integration [4], [68], both issues we discuss further below.
The concept of geodesic generalization provides a formal perspective on spatial representation which is different from, but complementary to, much other work in this area. Whereas much experimental and theoretical work on the hippocampal formation concerns essentially sensory-side questions—how place or grid cells combine different sorts of inputs to produce their instantaneous representations, or to learn them over time—we attempt to isolate the downstream question of how the resulting representations serve downstream learning functions. To this end, we do not address the input-side question of how the hypothesized distorted spatial representations are themselves produced from more elementary inputs. We only assume, abstractly, that the basis functions are computed on the fly from a learned map of the barriers in the environment. In sparse environments such maps could easily be learned from observation in a single trial, and may implicate the “border cells” of entorhinal cortex [69]. All this leaves open the opportunity, in future work, for studying how the input- and output-side perspectives relate: whether the mechanisms studied by previous authors might be made to produce or approximate representations of the sort we envision. For instance, in the geodesic view, place fields tend to be unidirectional on the linear track [70], [71] because the states of passing through them facing either direction are far apart in the state transition graph of a shuttling task. In input terms this more abstract relationship between states may be reflected in these situations being visually distinct [70], [71] or anchored to a different prior reference point [72].
More generally, unlike idealized RL models [3], [51], theories of how place cells arise from sensory inputs (e.g. via competitive learning [70], [71], or self-organizing maps [73]) do not necessarily imply the isotropic Gaussian firing fields we criticize, and thus may also offer (more mechanistic) explanations for phenomena such as place fields not crossing walls. It remains to be seen to what extent such local learning rules can be massaged to produce maps that accord with the globally geodesic ideal. However, such unsupervised learning models tend to envision that representations are acquired incrementally over time, which stands in contrast to our assumption (supported by data such as place field changes occurring immediately when barriers are added [65]) that the geodesic basis is computed on the fly with respect to the current barrier locations. A different mechanism that could be useful in producing geodesic firing fields is the “arc length” cell posited by Hasselmo [63], [74], a circuit for computing along-path distance using oscillatory interference mechanisms related to those thought to be involved in grid formation. This mechanism has already been used to explain several examples of context-dependent firing of hippocampal neurons similar in spirit to the phenomena we consider here.
The behavior of the entorhinal representation also raises interesting questions about the relationship between input- and output-side considerations. To start, it is often assumed that the place code is built up by linear combinations of grid cell inputs, e.g. by a sort of inverse Fourier transform [13]. In such a model, it can be shown (and simulations, not shown, verify) that place cells will inherit the geometry of their grid cell inputs. For this reason, we suggest that grid cells are likely to use a geodesic metric even if they do not directly serve as a basis for value function learning (but only indirectly, as a basis for geodesic place cells). However, this exposes some tension between the output-side imperative of generalization for RL, which we have argued calls for geodesic distortions, and the input-side implication of the system in path integration (i.e. tracking vector coordinates in a path-independent manner) [35], [37]–[40], [47], [75], which is an inherently Euclidean operation.
In this respect, the recent results of Derdikman et al. [24] showing distorted and fractionated grid fields in a hairpin maze seem difficult to reconcile with a global Euclidean path integrator (since the hairpin barriers do not change the Euclidean coordinates), and at least qualitatively more in line with the geodesic view. One possible path toward reconciling these considerations is to consider a sort of hierarchical representation that treats the environment as a collection of rooms (in the hairpin maze, hallways) whose interrelationships are represented as by a geodesic graph, but with (disjoint) Euclidean representations maintained within each of them. This has resonance with multi-level navigation models from animal behavior (e.g. [68]), with multiple map views of hippocampus [72], and, also, mechanistically, with some of the more detailed aspects of the Derdikman [24] data that are not captured by our model. Most importantly, the Derdikman data suggest that the grid phase resets and “anchors” at left or right turns, producing similar patterns in alternating arms and suggesting a possible mechanism for separating adjacent hallways' representations. Such heuristics for grid resetting and anchoring (and also stretching) [24], [34] may be able to produce a “good enough” approximation to the geodesic metric, at least in some environments, and have been examined in much more detail in more biologically detailed modeling of the task [38]. One sign of approximations is where they break down. In this respect, it is interesting that the rather extreme case of the hairpin maze results in badly fractionated downstream place fields as well [24], a phenomenon not predicted by the exact geodesic model. Finally, unlike our full model, a resetting mechanism would not in itself seem to explain phenomena related to barriers within a room, such as those we illustrate in Figure 7. A fuller understanding of these sorts of mechanisms demands additional research, both experimental and theoretical.
Our simulations also demonstrate that the grid representation itself is a suitable basis for value function learning, even without an intermediate place cell representation. On one level, these results serve to underline the generality of our points about geometry and generalization, using a rather different basis. More speculatively, they point to the possibility that the grid representation might actually serve such a role in the brain, echoing other work on the usefulness of this Fourier-like basis for representing arbitrary functions [12], particularly (as also for standard uses of Fourier representations in engineering for compressing images and sounds) smooth ones. However, although a few studies have demonstrated anatomical connections from the entorhinal cortex to striatum [55]–[57], [76], grid-like responses are less often reported in the deep layers that give rise to these subcortical projections (though see [53], [54]).
Finally, although for simplicity and concreteness we have focused on the principles of value function generalization in the context of a particular task (spatial navigation) and algorithm (TD(λ) learning), many of the same considerations apply more generally. First, across domains, in computational neuroscience, the need for (temporally) smooth basis functions been suggested to improve generalization also in learning about events separated in time rather than space [61], though there is no obvious counterpart to the geodesic distance metric in this setting.
Second, across algorithms, TD-like learning mechanisms also likely interact with additional ones in the brain, and the core considerations we elucidate about efficient generalization due to appropriate state space representations crosscut these distinctions. For instance, value functions may also be updated using replay of previously experienced trajectories (e.g., during sleep) [28], [51]. In models, this is typically envisioned to operate by the same TD learning rule operating again over the replayed experience [51], [77], and thus should imply parallel considerations of efficiency with respect to the number of replayed experiences required for convergence depending on the generalization characteristics of the basis. More distinct from these models, since the work of Tolman [67] it has been believed that spatial navigation may in part be accomplished by map-based route-planning processes that in RL terms correspond to model-based algorithms [78]–[82] rather than model-free algorithms like TD learning. These algorithms plan routes from a learned representation of the state transition matrix and rewards, typically using variants of the value iteration algorithm to compute state or action values. The core of this process is the iterative evaluation of Bellman's equation (Equation 1 in Materials and Methods), the same equation sampled with each learning step of TD. Thus, there is reason to think that efficient value iteration (here defined as fast convergence of the value function over iterations) will analogously occur when the update is over state representations that provide better generalization over states at each step. In all, then, although we exemplify them in a highly simplified model, the principles of state representation for efficient reinforcement learning are quite general.
Another issue arises when considering the present model in light of model-based RL. One of the hallmarks of model-based planning (and the behavioral phenomena that Tolman [67] used to argue for it, albeit not subsequently reliably demonstrated in the spatial domain), is the ability to plan novel routes without relearning, e.g. to make appropriate choices immediately when favored routes are blocked or new shortcuts are opened. Interestingly, rather than by explicit replanning, some such behaviors could instead be produced more implicitly by updating the basis functions to reflect the new maze, while maintaining the weights connecting them to value. This is easy to demonstrate in the successor representation [16], a model closely related to ours. To behave similarly, the present model would require additional constraints to ensure the basis functions corresponding to different mazes are interchangeable, but this would be one route toward explaining shortcut phenomena in this framework. More generally, because the present proposal uses a state transition model, implicitly, to generate a basis function that is then used with model-free learning [see also 16], [83], [84], it resembles something of a cooperative hybrid of model-free and model-based techniques somewhat different from the competitive approaches suggested elsewhere [78].
We simulate value function learning in a gridworld spatial navigation task in order to compare linear function approximation over several different spatial basis sets [8]. Our model learns to estimate the value function over states (i.e., positions in the grid), defined in the standard way as the expected future discounted reward:(1)
To simplify notation, we omit the dependence of these quantities on the action policy throughout. The model learns approximations to these values by learning a set of N linear weights w1…N for N spatial basis functions φ1…N(s) defined over the entire state space. The estimated value is thus:(2)
We use a simple temporal-difference algorithm with eligibility traces [8], [85] to learn weights. Specifically, at each run upon visiting state s receiving reward r(s) and transitioning into state s′, for each basis φi, weights wi are updated at each time step using the following algorithm:(3)
This is just the version of the familiar TD(λ) rule for linear value function approximation, with free parameters α (learning rate), λ (trace decay rate), and γ (discount factor).
We tested the model in 20-by-20 (M = 400 states) gridworlds in which the agent could move in any of the four cardinal directions, unless a wall blocked such a movement. Agents were started at a random location (i.e. state) at each trial, and had to reach the goal state, which was the only state with a reward, r(s) = 1. Individual trials ended when the agent reached the goal state, which was absorbing, or the maximum number of actions allowed, which was 500.
For simplicity, as described above the agent learns the value function over states and uses this to guide actions toward the goal, rather than directly learning the full Q-function over states and actions. This is because, in a spatial gridworld task, the state-action-state transition model is transparent, so we assume the agent evaluates the value of each action in a state as the value of the appropriate neighboring state [86]. Since the computation of Q involves a single step of what amounts to model-based lookahead, the approach is not as purely model-free as standard Q-learning or actor-critic algorithms. As with eligibility traces, we include this elaboration because it slightly improves generalization between states and actions, and might thus reduce the need for the sorts of basis-function-based generalization mechanisms we argue for.
The agent chooses actions according to a softmax policy, i.e. , where actions unavailable (due to walls) are not considered and β is the inverse temperature that balances the amount of exploration and exploitation in action selection. For these simulations, the inverse temperature was fixed to β = 80 (a factor calibrated to provide a reasonable explore/exploit balance in choice probabilities given the scale of the action values learned). To maintain such balance, because each gridworld had a different distance between the goal state and other states, for each environment the discount factor was scaled to γ = 0.9d/c so that each gridworld had the same value range. Here, d is the shortest maximum distance from any state to the goal, across all gridworlds tested, and c is the maximum interstate distance for a given gridworld (range 26 to 105 states). In order to compare fairly the different basis functions, the learning rate α was chosen for each condition and each basis set to minimize the mean number of steps to termination over a fixed number of trials, using a grid search in the range [0,1]. All simulations and analyses were performed using Matlab (Natick, MA).
We compare the model's learning using several different linear basis sets. Each basis is an M (states)×N (basis functions) matrix, with each column φi defining a function over the states. Bases were constructed as below, and lastly each row of the matrix was normalized by its L2 norm. This ensures that the learning rate parameter α in the update rule (Equation 3) has a consistent interpretation (as a fractional stepsize) between different states and basis sets.
|
10.1371/journal.pgen.1005487 | Genome-Wide Association Study with Targeted and Non-targeted NMR Metabolomics Identifies 15 Novel Loci of Urinary Human Metabolic Individuality | Genome-wide association studies with metabolic traits (mGWAS) uncovered many genetic variants that influence human metabolism. These genetically influenced metabotypes (GIMs) contribute to our metabolic individuality, our capacity to respond to environmental challenges, and our susceptibility to specific diseases. While metabolic homeostasis in blood is a well investigated topic in large mGWAS with over 150 known loci, metabolic detoxification through urinary excretion has only been addressed by few small mGWAS with only 11 associated loci so far. Here we report the largest mGWAS to date, combining targeted and non-targeted 1H NMR analysis of urine samples from 3,861 participants of the SHIP-0 cohort and 1,691 subjects of the KORA F4 cohort. We identified and replicated 22 loci with significant associations with urinary traits, 15 of which are new (HIBCH, CPS1, AGXT, XYLB, TKT, ETNPPL, SLC6A19, DMGDH, SLC36A2, GLDC, SLC6A13, ACSM3, SLC5A11, PNMT, SLC13A3). Two-thirds of the urinary loci also have a metabolite association in blood. For all but one of the 6 loci where significant associations target the same metabolite in blood and urine, the genetic effects have the same direction in both fluids. In contrast, for the SLC5A11 locus, we found increased levels of myo-inositol in urine whereas mGWAS in blood reported decreased levels for the same genetic variant. This might indicate less effective re-absorption of myo-inositol in the kidneys of carriers. In summary, our study more than doubles the number of known loci that influence urinary phenotypes. It thus allows novel insights into the relationship between blood homeostasis and its regulation through excretion. The newly discovered loci also include variants previously linked to chronic kidney disease (CPS1, SLC6A13), pulmonary hypertension (CPS1), and ischemic stroke (XYLB). By establishing connections from gene to disease via metabolic traits our results provide novel hypotheses about molecular mechanisms involved in the etiology of diseases.
| Human metabolism is influenced by genetic and environmental factors defining a person’s metabolic individuality. This individuality is linked to personal differences in the ability to react on metabolic challenges and in the susceptibility to specific diseases. By investigating how common variants in genetic regions (loci) affect individual blood metabolite levels, the substantial contribution of genetic inheritance to metabolic individuality has been demonstrated previously. Meanwhile, more than 150 loci influencing metabolic homeostasis in blood are known. Here we shift the focus to genetic variants that modulate urinary metabolite excretion, for which only 11 loci were reported so far. In the largest genetic study on urinary metabolites to date, we identified 15 additional loci. Most of the 26 loci also affect blood metabolite levels. This shows that the metabolic individuality seen in blood is also reflected in urine, which is expected when urine is regarded as “diluted blood”. Nonetheless, we also found loci that appear to primarily influence metabolite excretion. For instance, we identified genetic variants near a gene of a transporter that change the capability for renal re-absorption of the transporter’s substrate. Thus, our findings could help to elucidate molecular mechanisms influencing kidney function and the body’s detoxification capabilities.
| Genome-wide association studies with metabolic traits (mGWAS) investigate the relationship between genetic variance and metabolic phenotypes (metabotypes). In 2008, Gieger et al. presented the first mGWAS in serum of 284 individuals [1]. Since then, numerous mGWAS using different analytical platforms and ever larger study populations were published [2–8]. These studies discovered more than 150 genetic loci that associate with blood levels of more than 300 distinct metabolites. We refer to these loci as the genetically influenced metabotypes (GIMs), their ensemble defining the genetic part of human metabolic individuality. Many of the single nucleotide polymorphisms (SNPs) that associate with metabolic traits map to genetic regions coding for enzymes or metabolite transporters that are biochemically linked to the associated metabolites. Moreover, a large number of these GIMs have been previously linked to clinically relevant phenotypic traits. As intermediate traits on the pathways of many disorders, these GIMs have become valuable tools that allow unraveling disease mechanisms on the molecular level [9].
However, so far mGWAS have mostly been limited to studies of serum or plasma metabolite levels, thereby focusing on genetically influenced metabolic homeostasis in blood. Only a few studies investigated urine as a complementary body fluid enabling studies of kidney function and the detoxification capabilities of the human body. In 2011, we published the first mGWAS in urine [10] using proton nuclear magnetic resonance spectroscopy (1H NMR) to determine metabolite concentrations in urine of 862 male participants of the SHIP-0 cohort. We identified five genetic loci (SLC6A20, AGXT2, NAT2, HPD, and SLC7A9) that modulate urinary metabolite levels. While for this study metabolite concentrations were manually derived from the NMR spectra for a targeted set of metabolites, Nicholson et al. [5] directly used spectral features as abstract, non-targeted urinary metabolic traits in an mGWAS. Based on data for 211 participants of the MolTWIN and MolOBB studies, the authors identified SNPs at three loci (ALMS1/NAT8, AGXT2, and PYROXD2) that were associated with metabolic traits in urine. Two of these loci (ALMS1/NAT8 and PYROXD2) were replicated in an NMR-based mGWAS published by Montoliu et al. For that study, the authors analyzed non-targeted urinary traits from 265 subjects from the São Paolo metropolitan area [11]. Recently, Rueedi et al. [12] reported significant associations of NMR-derived non-targeted urinary traits in ten loci (ALMS1/NAT8, ACADL, AGXT2, NAT2, ABO, PYROXD2, ACADS, PSMD9, SLC7A9, and FUT2) using data from 835 participants of the CoLaus study, thus bringing the total number of reported urinary GIMs to eleven.
Here, we substantially extend our previous mGWAS with metabolic traits in urine, both in size and in scope. First, we metabolically characterize the urine samples of 3,861 male and female participants of the SHIP-0 study, thereby quadrupling the sample size when compared to previous studies. Second, we combine both targeted and non-targeted NMR-based metabolomics. In this way, we implement the approaches used in the studies by Nicholson et al., Montoliu et al., and Rueedi et al. alongside the targeted metabolomics approach used in our previous study. For an unbiased interpretation of our mGWAS results, we apply tools for evidence-based locus-to-gene mapping and automated assignment of metabolites to non-targeted NMR spectral features. Finally, besides determining the overlap of variants identified in our study with variants previously linked to clinical traits, we specifically investigate the overlap between variants influencing metabolic traits in both urine and blood.
Our study is based on one-dimensional 1H NMR spectra of urine samples from 3,861 genotyped participants in the SHIP-0 cohort (see Methods). For the targeted metabolomics analysis, we manually quantified a set of 60 metabolites in these spectra (Fig 1A). For the non-targeted analysis, we used the same spectra and applied an automated processing algorithm to align the spectra and to perform dimensionality reduction [13]. In the subsequent analysis, we screened the targeted and the non-targeted metabolic traits as well as the pairwise ratios within each trait type for associations with genotyped and imputed variants in a two-step approach (Fig 1B). We identified a total of 23 genetic loci that display significant associations with targeted and/or non-targeted metabolic traits (Fig 2, Tables 1 and 2). All but one of the discovered loci replicated in data from the KORA F4 cohort (N = 1,691). For 15 loci, our study is, to the best of our knowledge, the first to report associations with urinary traits. For 7 of these 15 loci, associations have previously been reported with blood metabolites. Thus, 8 loci are entirely new (Fig 3). Finally, 11 of the 22 replicated loci host significantly associated variants that were previously associated with phenotypes of clinical relevance (Table 3).
For the targeted metabolomics analysis, the 1H NMR spectra were manually annotated to derive absolute metabolite concentrations (out of a panel of 60 compounds) for each sample. For the non-targeted analysis, the same spectra were automatically aligned and processed using the FOCUS package [13]. This resulted in NMR signal intensities at 166 distinct spectral positions (“chemical shifts”) per sample (see Methods). In previous mGWAS, we demonstrated the potential of testing pairwise ratios of metabolite concentrations to boost genetic association signals [1, 4, 6, 8, 10, 21, 53]. Recently, we showed that this approach can also be successfully applied to NMR-based mGWAS with non-targeted features [22]. Thus, we calculated the pairwise ratios of all metabolite concentrations with at least 300 valid data points over all samples (55×54/2 = 1,485 ratios of targeted traits) and NMR signal intensities (166×165/2 = 13,695 ratios of non-targeted traits), respectively (Fig 1A). Out of all 15,401 metabolic features (targeted and non-targeted traits and the ratios thereof), a total of 15,379 features with at least 300 valid data points were screened for genetic associations using 620,456 genotyped autosomal SNPs (Fig 1B). To this end, we computed age- and sex-adjusted linear models under the assumption of additive genetic effects for each SNP-metabolic trait pair. A total of 499 genotyped variants display associations with metabolic traits with P-values below 5×10−8.
We used the 499 variants identified in the mGWAS to tag 54 distinct chromosomal regions at a window size of at least 2 Mb (centered to the tag SNPs). We then performed additional association studies using imputed variants (1000 genomes project imputation) in the tagged regions (Fig 1B). We considered associations with a P-value below the Bonferroni-adjusted significance threshold of α’ = 5×10−8/15,379 = 3.25×10−12 to be genome-wide significant. For ratio traits, we also required the P-gain to be greater than 1.52×104 for targeted traits and 1.38×105 for non-targeted traits (10 times the number of tested traits [53]). P-gain reflects the increase of association strength with the ratio trait when compared to the P-values that result from associations with the individual traits buildinging the ratio. A total of 2,882 genotyped or imputed SNPs display association signals below P < 3.25×10−12 and, in case of ratios, above the imposed P-gain threshold (Fig 2). All significantly associated SNPs within a physical distance of 1 Mb were assigned to one of 23 distinct genetic loci. Three loci display significant association signals only when imputed SNPs were used, and 8 loci show significant associations only when pairwise ratios of metabolic traits were considered. Twelve loci show significant associations in both targeted and non-targeted data sets. Three loci are only significantly associated with targeted traits (i.e., quantified metabolite concentrations or ratios thereof), whereas 8 loci are only significantly associated with non-targeted traits (spectral features or ratios thereof) (Fig 3). For each locus, we list the SNP that displays the strongest association signal (lead SNP) and its associated metabolic trait in Tables 1 and 2. In addition, we provide boxplots, regional association plots, and Q-Q plots for each locus in S2 Fig. The summary statistics for all association signals with P < 0.05 (P < 1×10−4 and P-gain ≥ 10 for associations with ratios) for each tested SNP can be downloaded from http://www.gwas.eu.
In general, the biological interpretation of association results from mGWAS requires the mapping of SNPs to candidate genes that are most likely causally linked to the observed changes in the metabotype. Furthermore, non-targeted metabolic traits that exhibit significant association signals have to be assigned to distinct metabolites. In our study, we implemented algorithmic approaches for both the locus-to-gene mapping and the assignment of non-targeted metabolic features.
As the first step in the candidate gene selection, we assigned the significantly associated SNPs to distinct loci using a physical distance threshold of 1Mb. Assigning variants within a locus to one of the covered genes based only on proximity or plausibility ignores haploblock structure and existing regulatory information for the SNPs such as expression quantitative trait loci (eQTL). To take such information into account and to achieve an unbiased selection of candidate genes, we collected evidence for each significantly associated SNP and its proxies in strong linkage disequilibrium (LD) using the SNiPA web server [51]. For each locus, we received a list of candidate genes that are linked to one or more associated variants (or a proxy in LD). Thereby, genes are linked via genomic proximity (i.e., if any of the variants is located within the candidate gene or is in close proximity), via eQTL associations (i.e., if any of the variants is associated with expression levels of the gene in a previous eQTL study), or via regulatory element association (i.e., if any of the variants is contained in a promoter/enhancer/repressor element that is associated with the gene). Moreover, missense variants or known pathogenic variants in the locus are considered to provide additional types of evidence for the linked genes. We finally assigned the locus to the gene with the strongest functional evidence (i.e., the gene showing the highest number of different types of evidences (max. 5) among the candidate genes; see Methods). In case of ambiguous assignments, the gene with the most plausible biological function was chosen. As an example, one locus contains a high number of SNPs with strong associations with non-targeted traits corresponding to N-acetylated compounds. These SNPs cover 12 different genes (see regional association plots in S2 Fig). The gene covered by the highest number of SNPs is ALMS1. However, there are 3 more genes in this locus with the same amount of functional evidence count as ALMS1 (S3 Table). One of these genes is NAT8, which encodes an N-acetyltransferase. Since there is a biologically meaningful link between the function of the NAT8 gene product and the associated metabolic traits, we annotated this locus with NAT8 as the most likely candidate gene. According to our evidence-based candidate genes assignment approach, the 23 loci map to the genes NAT8, HIBCH, CPS1, AGXT, XYLB, SLC6A20, TKT, ETNPPL, SLC6A19, AGXT2, DMGDH, SLC36A2, NAT2, ABO, GLDC, PYROXD2, SLC6A13, HPD, ACSM3, SLC5A11, PNMT, SLC7A9, and SLC13A3. S3 Table provides a complete list of candidate genes and the corresponding collected evidences.
For the identification of metabolites underlying non-targeted NMR traits, we used pseudo-spectra that display the strength of associations of a given SNP across the complete NMR spectrum [12, 22]. If the association is strong enough, these “association spectra” often exhibit a striking similarity to the reference NMR spectrum of the underlying metabolite(s). For the present study, we applied the “metabomatching” method introduced by Rueedi et al. [12] to perform an automated annotation of the association spectra for each genetic locus of interest. For 19 of the 20 loci that display significant associations with non-targeted traits, metabomatching suggests plausible metabolite candidates matching signals present in the association spectra (S1 Fig). For 10 of these 19 loci (CPS1, SLC6A20, ETNPPL, SLC6A19, AGXT2, SLC36A2, HPD, ACSM3, PNMT, and SLC13A3), the match between the association signal and the NMR spectrum of the candidate metabolite (as provided by the Urine Metabolome Database [54]) is strong and unique, which makes the assignment of a metabolite identity to a non-targeted trait unambiguous in these cases.
To replicate our findings, we used genotype data and urine samples from participants of the KORA F4 cohort (N = 1,691). From recorded 1H NMR spectra of the urine samples, we derived the targeted and non-targeted metabolic traits (metabolite concentrations, NMR spectral features, and the respective pairwise ratios) as for the discovery study. For 14 of the 15 new loci that show significant associations with targeted metabolic traits in the SHIP-0 data set, the top-ranking SNP/metabolic trait association replicates in KORA F4 (S1 Table). For the SLC7A9 locus, the association with lysine/valine does not replicate, possibly due to the difficulty in annotating lysine from the NMR spectra (> 75% missing values for lysine). However, the second-best, still genome-wide significant association of the tested SNP with valine replicates. For 15 of 20 loci that display significant association signals in the GWAS with non-targeted traits, we were able to replicate the best SNP/NMR trait association or, if this failed, the next, still significant follow-up association (S2 Table). The failure to replicate the remaining 5 loci might be due to the lower sample size in KORA, due to different fasting states of the subjects in the different cohorts, or due to a less perfect alignment of the NMR spectra, since we chose the same FOCUS parameters for aligning SHIP and KORA spectra instead of treating them separately. However, 4 of these 5 loci (ETNPPL, SLC6A19, DMGDH, PNMT) show also significant associations in the targeted SHIP-0 data set that replicate in the targeted KORA F4 data set (Table 1). Out of the 23 loci identified in the discovery study, ABO is the only locus that could not be replicated using either a targeted or a non-targeted metabolic trait in KORA F4, leaving 22 loci that display stable associations with metabolic traits in urine.
We evaluated each identified and replicated locus in the light of previously reported associations with metabolic phenotypes and clinical traits. To this end, we selected all SNPs within a locus for which we found genome-wide significant associations with any urinary metabolic trait in the SHIP-0 cohort. Furthermore, we added all bi-allelic variants from the 1000 genomes project [50] (phase 1, version 3, European ancestry) that are in strong LD to these SNPs (r2 ≥ 0.8).
For 15 of the 22 loci, no associations with urinary metabolic traits were reported so far (HIBCH, CPS1, AGXT, XYLB, TKT, ETNPPL, SLC6A19, DMGDH, SLC36A2, GLDC, SLC6A13, ACSM3, SLC5A11, PNMT, and SLC13A3) (Fig 3, Table 3). The remaining 7 loci were already identified in our previous urine mGWAS (AGXT2, HPD, SLC7A9, SLC6A20, and NAT2) [10] or in the studies by Nicholson et al. [5] and Rueedi et al. [12] (NAT8 and PYROXD2). For all 7 loci, both trait association and direction of the observed effect are consistent with the results previously published.
We further compared our association results with those of published mGWAS with metabolic traits in blood (P < 5×10−8), including all studies listed in the NHGRI GWAS catalog [23] and other studies such as the mGWAS by Shin et al., which is based on metabolomics data from the KORA F4 and TwinsUK cohorts [1, 4, 6–8, 14–22, 24]. In total, 14 loci show significant associations with metabolic traits both in blood in one of these mGWAS and in urine in our study (Fig 3, Table 3). For 3 of these 14 loci (SLC6A20, PNMT, AGXT), we consider the associated metabolic traits in both media to be unrelated. In 5 cases (NAT2, NAT8, PYROXD2, SLC7A9, TKT), the genetic association analyses identified different, but related metabolites (i.e., the associated metabolites from urine and blood are either products/substrates of the locus’ candidate gene product, or are biochemically converted within another known enzymatic reaction, or belong to the same metabolite class). In 6 cases, the associations target the same metabolites in urine and blood (CPS1, AGXT2, DMGDH, SLC6A13, HPD, SLC5A11). For 5 of these 6 loci, the direction of the observed effect is the same, whereas for SLC5A11 (associated with myo-inositol), we observe an increase in urinary metabolite concentration per copy of the effect allele, as opposed to decreased levels reported in blood (Table 3). For this locus, we additionally investigated whether the effects seen in blood and urine are directly coupled. To this end, we made use of myo-inositol levels (normalized to circulating creatinine) measured through mass spectrometry (MS) in blood serum samples of the same KORA F4 participants [6] that form the replication cohort in this study. The ratio between the urinary myo-inositol (this study) and the serum myo-inositol levels shows an increase in association strength to the lead SNP in SLC5A11 (rs17702912) by seven orders of magnitude in comparison to the association of urinary myo-inositol alone (Purine < 1.95×10−24, Pblood < 1.50×10−4, Pratio < 2.43×10−31).
For 11 loci, our mGWAS identified significantly associated variants for which either the same variant or a proxy in strong LD was previously reported to be associated with clinical phenotypes according to data from the NHGRI GWAS catalog [23] (P < 5×10−8), OMIM variation, ClinVar [55], HGMD [56], or dbGaP [57]. Amongst others, these variants have been linked to chronic kidney disease (NAT8, CPS1, SLC6A13, and SLC7A9), pulmonary hypertension (CPS1), ischemic stroke (XYLB), Iminoglycinuria (SLC6A20), heart rate variability (AGXT2), Hawkinsuria (HPD), and pharmacogenomically relevant acetylation phenotypes (NAT2) (Table 3). In addition to these 11 loci with disease-associated variants, we found previously discovered connections between clinical traits and the assigned candidate gene for another 7 loci (HIBCH, AGXT, SLC6A19, DMGDH, SLC36A2, GLDC, ACSM3).
In this study, we present the largest genome-wide association study with metabolic traits (mGWAS) in urine to date. In addition to quadrupling the sample size compared to previous mGWAS in urine, we analyzed both targeted traits (metabolite concentrations manually derived from NMR spectra) and non-targeted traits (NMR spectral features).
In total, we identified 23 genetic loci with significant associations between genetic variants and targeted or non-targeted metabolic traits in urine of SHIP-0 participants, 22 of which replicate in the independent KORA F4 cohort. To the best of our knowledge, 15 loci have not been linked to changes in the urine metabolome before. For the remaining 7 loci, our results are in line with the results from previous mGWAS in urine [5, 10, 12] regarding both the associated metabolic traits and the direction of the genetic effects (Table 3).
Though derived from the same NMR spectra, the list of GIMs identified with targeted traits and non-targeted traits partly differ. Of the 22 genetic loci reported in this study, only 12 loci were discovered in both targeted and non-targeted traits, whereas 7 loci show significant associations only with non-targeted traits, and 3 only with targeted traits (Fig 3 and Table 3).
For the data set used in the targeted analysis, the NMR spectra were manually annotated to identify and quantify the metabolites underlying the spectra. Involving human expert knowledge usually allows metabolite identification with very high confidence and yields more precise quantification, especially if signals of multiple metabolites overlap in the NMR spectrum. Furthermore, a manual annotation can to some extent compensate for different experimental and sample conditions, as alignment and pre-processing can be optimized for each spectrum individually. As an example, lysine exhibits characteristic signals in the NMR spectral region between δ = 1.68 and 1.76 ppm, which is often dominated by signals from a variety of additional metabolites, making the annotation of lysine a very difficult task. Thus, while lysine concentrations could be determined for 888 samples of the discovery cohort through manual quantification and yielded a significant genetic association at SLC7A9, the non-targeted approach did not capture any association signals for this locus.
However, a manual spectral annotation as performed in the targeted analysis is quite laborious, which limits the number of quantifiable metabolites in large studies. This leads to a bias towards a certain set of metabolites and, as a consequence, significant associations actually present in the NMR data might be missed. Also, a manual annotation in general bears some risk of annotator-induced bias [58]. As an automated method, the non-targeted analysis of spectra has the potential to overcome some of the limitations of targeted analyses. Here, the most prominent example is the PYROXD2 locus, where SNPs display exceptionally strong associations (P < 1.0×10−307) to the NMR signal intensities at δ = 2.854 ppm. We could not identify any significant associations within this locus using the targeted data. Thus, we assumed that our set of targeted traits did not cover the metabolite(s) corresponding to these signals. The challenge with genetically associated non-targeted traits lies in the lack of biochemical interpretability. To facilitate the assignment of non-targeted NMR traits to chemical compounds, we applied the metabomatching algorithm introduced by Rueedi et al. [12]. In case of PYROXD2, metabomatching suggests that the associated NMR signals correspond to trimethylamine. Thereby, the automated method replicates the findings of Nicholson et al. [5] where the authors manually annotated the associated signals based on expert knowledge. In case of our mGWAS with targeted traits, trimethylamine was not part of the metabolite panel and thus the association with PYROXD2 could only be discovered using non-targeted metabolic traits in combination with the automated metabomatching processing. Of course, automated annotation of non-targeted traits also has its limitations: the annotation through metabomatching relies on the association signals that genetic variants display over the NMR spectral range (“association spectra”) as well as on the existence of the relevant reference metabolite spectrum (see Methods and S1 Fig). In some cases, these association spectra are not meaningful enough to allow an unambiguous assignment of non-targeted features to metabolites, or they may be pointing to a metabolite not present in the reference set.
In summary, our study demonstrates that GWAS with NMR-determined metabolic traits can benefit from a combined application of both targeted and non-targeted metabolomics. Our results suggest that a targeted approach is better suited for the annotation of metabolites for which the corresponding NMR signals are in regions with a plethora of other signals as in some cases these signals cannot be resolved through non-targeted methods. Furthermore, genetic associations with targeted traits appear to be more robust, since 5 of the 12 loci that display associations with both targeted and non-targeted traits clearly display stronger association signals in the targeted data set (several orders of magnitude in case of the SLC6A20 locus; Tables 1 and 2). However, the non-targeted metabolic traits provide a less biased view on the metabolome, which in our case results in additional significantly associated genetic loci.
Fifteen of the 22 identified and replicated loci show a plausible biochemical connection between functionally annotated genes and their associated metabolic traits (Table 3). This is similar to observations from previous mGWAS. For instance, Shin et al. reported biologically meaningful links between metabolites and genetic loci for 101 of 145 GIMs [21]. In case of genes with vague functional annotations, gene-metabolite associations from mGWAS provide testable hypotheses for further gene characterization. As an example, Suhre et al. experimentally confirmed the mGWAS driven hypothesis of SLC16A9 being a carnitine transporter [6]. Vice versa, with the help of mGWAS, the chemical structure of a non-targeted metabolic trait was elucidated through the function of the associated gene [8].
Another prominent finding of previous mGWAS is the overlap between disease relevant genetic variants and variants associated with metabolic traits. In the present study, we found 11 loci hosting variants that have previously been linked to clinical phenotypes. This includes associations with the estimated glomerular filtration rate (eGFR) and chronic kidney disease (CKD). Thus, the associated metabolites might, on the one hand, serve as intermediate traits for clinical endpoints. On the other hand, the associations might provide new insights regarding the involvement of specific metabolic pathways in pathomechanisms and the mediation of genetic risk loci through metabolic changes. For all 22 GIMs, we provide information on both the match of gene and metabolite function and the link to clinical traits in Table 3. In the following, we exemplify the value of our results for the characterization of gene functions in the light of clinical phenotypes.
As a first example, we identified significant associations of variants upstream of ETNPPL with ethanolamine. Interestingly, at the time when we received the first results from our association studies this gene was named AGXT2L1 and was assumed to encode an alanine-glyoxylate-aminotransferase. Based on this gene annotation, there was no obvious relation to the associated metabolite ethanolamine. In such cases, only dedicated experiments (similar to the one for the carnitine/SLC16A9 association mentioned above [6]) could validate the connection of ethanolamine to the gene product. Meanwhile, Veiga-da-Cunha et al. experimentally investigated the locus in an independent study and found that AGXT2L1 actually encodes an ethanolaminephosphate-phospholyase [59]. As a consequence, AGXT2L1 now carries the gene symbol ETNPPL. As ethanolamine is a direct precursor of ethanolaminephosphate via ethanolamine kinase (EC 2.7.1.28), our finding indeed matches the actual gene function. Besides the functional characterization of this locus, Veiga-da-Cunha et al. suggest that the ETNPPL-mediated degradation of ethanolaminephosphate balances the concentration of that metabolite in the central nervous system. They concluded that an altered ethanolaminephosphate homeostasis might contribute to mental disorders such as schizophrenia [59]. In line with this hypothesis, the ETNPPL expression rate in brain was previously found to be associated with schizophrenia [60]. ETNPPL is primarily expressed in brain and liver and the encoded protein is, amongst other tissues, highly localized in the cerebral cortex and the kidney (S4 Table and The Human Protein Atlas [52], http://www.proteinatlas.org/ENSG00000164089/tissue). Our results suggest that an excess of ethanolamine in urine could indicate alterations in ethanolaminephosphate homeostasis linked to a genetically reduced enzymatic activity of ETNPPL.
As a second example, we identified significant associations of genetic variants with 2-hydroxyisobutyrate (2-HIBA) in a locus comprising 9 different genes. According to our evidence-based locus to gene assignment, 4-hydroxyphenylpyruvate dioxygenase (HPD) is the most probable effector gene candidate. The association between 2-HIBA and this locus represents a well replicated finding: it was already identified in our previous NMR-based mGWAS in urine [10] and it has meanwhile also been discovered in an MS-based GWAS with blood metabolites [21]. Nonetheless, to the best of our knowledge, there is no obvious, known biological link between 2-HIBA and the HPD gene or any of the remaining 8 genes covered by this locus. In the literature, 2-HIBA is often referred to as a secondary metabolite that can be found in urine of humans and rats exposed to the volatile gasoline additives methyl-tert-butylether and ethyl-tert-butylether [61–63]. However, 2-HIBA has been identified by both MS- and NMR-based methods in almost all serum and urine samples of large human cohorts (e.g. ARIC [25], CoLaus [12], KORA [6, 21], SHIP [10], TasteSensomics [12], and TwinsUK [6, 21]) in relatively high concentrations (~ 40 μM in urine in this study), which suggests sources beyond gasoline for this metabolite (e.g. microbiota [64] or medication [65]). Interestingly, Dai et al. recently showed that 2-HIBA is an intermediate for the newly discovered but common 2-hydroxyisobutyrylation of lysine residues of histones [66], thus indicating an endogenous role of 2-HIBA. In this context, it is interesting to note that SETD1B is one of the genes within the identified locus on chromosome 12. SETD1B is a component of the methyltransferase complex that specifically methylates the lysine-4 residue of histone H3 [67]. This residue is amongst the 63 sites for 2-hydroxyisobutyrylation presented by Dai et al. [66]. Thus, one could speculate that in addition to its activity as a histone methylase, SETD1B may also be involved in the newly discovered process of histone hydroxyisobutyrylation, a hypothesis that may now be tested by dedicated experiments.
As a third example, we discuss the association of variants in XYLB with increased urinary glycolate levels. One of these variants, rs17118, causes an amino acid exchange in the XYLB gene product. XYLB encodes the enzyme xylulokinase [33], which catalyzes the phosphorylation of D-xylulose to D-xylulose-5-phosphate. In humans, the vast majority of D-xylulose is metabolized via xylulokinase [33, 68, 69]. However, there is an alternative metabolic pathway in which D-xylulose is metabolized by phosphofructokinase (PFK) (S3 Fig) [70]. Therein, one of the downstream products is glycolate. Thus, the genetic variants in XYLB might reduce the enzymatic activity of xylulokinase and thereby cause a shift towards the alternative pathway. Interestingly, the minor allele of rs17118 has been implicated in increased susceptibility for ischemic stroke [32]. Furthermore, Jung et al. found a significant association between elevated glycolate levels in plasma and cerebral infarction [34]. In the alternative pathway, glycolate is a precursor of oxalate, whose toxic effect has been demonstrated repeatedly [33, 71]. Very recently, Rao et al. postulated that circulating oxalate precipitate might be a potential mechanism for stroke [72]. In this context, the association between the SNP rs17118 and glycolate (identified in our study) suggests that the carriers of this variants have a higher risk of stroke (identified in [32]) possibly via increased levels of glycolate or oxalate through favoring the alternative D-xylulose degradation. Unfortunately, oxalate or any other metabolite in the two D-xylulose degradation pathways are not detected in our metabolomics analysis to further support our hypothesis.
In total, 26 genetic loci that associate with urinary metabolic traits are known to date (22 identified or confirmed in this study plus 4 identified in previous studies [5, 11, 12], Fig 3). Of the 26 loci, only 8 loci lack corresponding SNP-metabolite associations in blood, and, based on current mGWAS, represent urine specific hits. All of these 8 loci were first reported in the present study. In case of the 14 loci with overlapping associations between blood and urine in our study (Table 3), 6 target the same metabolite in both media (CPS1, AGXT2, DMGDH, SLC6A13, HPD, SLC5A11). Interestingly, in all but one case (SLC5A11) the genetic effect has the same direction in both fluids, thus indicating that urine can be regarded as “diluted plasma” to some extent. For 5 of the 14 loci, we considered the associated metabolic traits in blood and urine to be biochemically related. Here, the metabolites are either products of the enzyme coded by the candidate gene (NAT8: N-acetylated compounds), or they are linked through an enzymatic reaction other than the reaction catalyzed by the candidate gene’s product (NAT2: 1,3-dimethylurate and 1-methylurate [73]; PYROXD2: trimethylamine and dimethylamine [EC 1.5.8.2]; SLC7A9: lysine and homocitrulline [EC 2.1.3.8]), or they belong to the same metabolite class (TKT: gluconate and erythronate are aldonates). The observed associations of related but different metabolites in blood and urine may be indicative either for biochemical conversions before excretion, or simply be a result of differences in the composition of the metabolite panels covered by the various mGWAS. In case of the remaining 3 loci, we find no direct biochemical or metabolic relationship between the metabolites in both media, since AGXT associates with an unknown compound in blood, PNMT associates with amino acids in urine and HDL cholesterol in blood, and SLC6A20 targets loosely related amino acid derivatives.
As an example for parallel effects in blood and urine, we identified an association between variants in the Carbamoyl-Phosphate Synthetase 1 (CPS1) gene and elevated urinary glycine levels. The strongest associated SNP rs715 was also identified in previous mGWAS with higher glycine concentrations in blood [14, 21, 24]. This variant has been highlighted previously as a putative regulator of CPS1 expression [74–76]. Furthermore, the second strongest glycine-associated SNP rs1047891 causes a non-synonymous mutation (Thr>Asn) in the C-terminal domain of the CPS1 polypeptide, which hosts the binding site for the allosteric activator N-acteyl-L-glutamate (NAG) [77]. Both SNPs are therefore potentially causative variants in this metabolic association. CPS1 is highly expressed in liver (S4 Table) and controls the first step in the urea cycle: ammonia is catalyzed to carbamoyl-phosphate, which in turn is the entry substrate of the urea cycle. CPS1 deficiency can lead to high ammonia levels in the body (Hyperammonemia, OMIM #237300) (S4 Fig). The association of the CPS1 variants with glycine can be explained by the conversion of excess ammonia to glycine via the glycine cleavage system [78, 79] and is thus biologically meaningful. The association between common variants in CPS1 and glycine might therefore be driven by mild forms of genetically induced Hyperammonemia. In this study, we could establish a link between genetic factors and a potential urinary marker for this condition.
SLC5A11 is the only locus where we observe an association with exactly the same metabolite in blood and urine but with reversed effects: while myo-inositol concentrations in urine increase per effect allele copy of the lead SNP, they decrease in serum (Table 3) [21]. The Solute Carrier Family 5 (Sodium/Inositol Cotransporter), Member 11 (SLC5A11) is a co-transporter of myo-inositol with sodium [80]. SLC5A11 was postulated to play a role in the regulation of serum myo-inositol concentrations [81], which was recently confirmed by an mGWAS in blood [21]. On the one hand, the influence of SLC5A11 on myo-inositol concentrations has been linked to apical transport and absorption in intestine [82]. On the other hand, SLC5A11 may be implicated in the re-absorption of myo-inositol in the proximal tubule of the kidney [83]. The opposite direction of the genetic influence in blood and urine as observed in our study suggests that SLC5A11 is actively involved in the re-absorption of myo-inositol. This assumption is further supported by the strong increase of the association strength when testing the ratio between urinary and serum myo-inositol. This could indicate that the reduced levels in blood are indeed caused by a reduced re-absorption rate in subjects that are homozygous regarding the effect allele.
As these examples demonstrate, mGWAS in urine extend our understanding of genetically influenced biochemical processes and can facilitate the knowledge transfer from blood to urine and vice versa. Currently, this transfer is limited by the comparatively low number of GIMs in urine (26) versus blood (>150). Further increasing the sample sizes of mGWAS in urine and the application of more sensitive MS-based metabolomics platforms as already used for blood mGWAS could compensate this bias.
For this study, we used data from SHIP (Study of Health in Pomerania) and, for replication, from the KORA (Kooperative Gesundheitsforschung in der Region Augsburg) study. Both studies have been described extensively in the study design papers [84–86] and in our previous publications [4, 6, 10, 21].
SHIP is a longitudinal population study conducted in West-Pomerania, located in the northeastern part of Germany. 4,308 inhabitants in that region participated in the first phase “SHIP-0”. For the GWAS presented here, metabolically characterized urine samples and genotype data were jointly available for 3,861 study participants (1,960 female and 1,901 male, aged 20 to 81 years). KORA is a population study conducted in the municipal region of Augsburg in southern Germany. The KORA F4 cohort comprises 3,080 subjects. For the study presented here, both genotype and urine samples from 1,691 participants (865 female and 826 male, age 32 to 77) were available. In both studies, all participants have given written informed consent and the local ethics committees (SHIP: ethics committee of the University of Greifswald; KORA: ethics committee of the Bavarian Chamber of Physicians, Munich) have approved the studies.
Both SHIP and KORA samples were genotyped using Affymetrix Human SNP Array 6.0 gene chips. SNPs were called using the Birdseed2 algorithm. In both data sets, the total genotyping rate was above 99%. 909,508 SNPs were genotyped in the SHIP-0 cohort, and 906,716 SNPs in the KORA F4 cohort. We excluded SNPs that violated the Hardy-Weinberg equilibrium (PHWE < 1.0×10−6, 8,623 in SHIP and 32,033 in KORA), or had a genotyping rate below 95% (57,160 in SHIP and 84,351 in KORA), or displayed minor allele frequencies (MAF) below 5% (227,967 in SHIP and 224,723 in KORA). After the exclusion, 620,456 autosomal SNPs remained in the SHIP-0 data set and 593,830 autosomal SNPs in the KORA F4 data set. Both SHIP and KORA genotypes were imputed in a two-stage process (pre-phasing followed by imputation). According to data from the 1000 genomes project (phase 1, March 2012 release [50]), we used SHAPEIT (v1.416) [87] for phasing in KORA F4 and IMPUTE (v2.2.2) [88] for imputation in KORA F4 and both phasing and imputation in SHIP-0. For the association analyses, we considered only imputed variants with a MAF ≥ 5%, PHWE ≥ 1.0×10−6, and an imputation quality score (IMPUTE info-score) ≥ 0.8.
Annotation data for genetic variants as well as linkage disequilibrium (LD) data from the 1000 genomes project (phase 1 version 3, EUR panel [50]) were retrieved from SNiPA v1 (http://www.snipa.org) [51]. Full lists of association signals from the serum-based mGWAS [21] were obtained from the Metabolomics GWAS server (http://www.gwas.eu).
All genotyped and imputed SNPs that displayed genome-wide significant association signals (according to the aforementioned P-value and P-gain criteria) were assigned to distinct genetic regions (loci), based on a physical distance threshold of 1 Mb. Each of the resulting 23 genome-wide significant loci was then projected to candidate genes using an evidence-based procedure. To this end, we used the “block annotation” feature of SNiPA on the LD-extended (r2 ≥ 0.8) list of associated variants at each locus. This feature provides a condensed view of genes that are linked to any of the significantly associated variants or their LD-proxies via genomic proximity, eQTL association, or regulatory elements. Additionally, the block annotation highlights missense and pathogenic variants. Based on these data, we defined the following criteria to identify candidate genes: 1) Genomic proximity: genes that harbor or are in close proximity (<5kb) to any of the variants in the list. 2) eQTL association: genes where altered expression levels have been discovered to associate with any of the variants in the list. 3) Regulatory elements: potentially regulated genes that are associated with a promoter/enhancer/repressor element containing a variant of the list. Further evidence for potential involvement of a gene was assumed if 4) the variant list contains a missense variant for a protein product of this gene and 5) if an intragenic variant in the list is annotated as pathogenic in one of the phenotype databases contained in SNiPA. For each gene, we counted how many of the aforementioned criteria are met. Thus, the maximum evidence count for a candidate gene is five. If evidence-based gene selection was ambiguous, the gene with the most plausible biological function was chosen (S3 Table).
Metabomatching [12] is an automated annotation method that identifies metabolites likely to underlie an observed genetic association between a SNP and one or more non-targeted metabolic traits (www.unil.ch/cbg). It does so by comparing the association signal between a SNP and all non-targeted traits (pseudo-spectrum or association spectrum) to the 1H-NMR spectra of metabolites in a reference set, and assigning a score to each pseudo-spectrum to NMR spectrum match. The metabolites most likely to underlie the genetic association are the ones with the highest scores. We applied metabomatching to each SNP showing a significant association with an NMR feature or feature-ratio, using the 180 metabolites listed in the urine metabolome database [54] with an experimental NMR spectrum as reference set. The urine metabolome database is the subset of metabolites in the Human Metabolome Database (HMDB) [90] present in urine. Where indicated, 2-compound or effect direction-specific metabomatching was applied. Final candidates were manually identified, usually among the few top-ranked matches. Most likely candidates are used in the text; potential alternatives are listed, along with the metabomatching details, in S1 Fig.
|
10.1371/journal.pntd.0000372 | Trypanosoma cruzi CYP51 Inhibitor Derived from a Mycobacterium tuberculosis Screen Hit | The two front-line drugs for chronic Trypanosoma cruzi infections are limited by adverse side-effects and declining efficacy. One potential new target for Chagas' disease chemotherapy is sterol 14α-demethylase (CYP51), a cytochrome P450 enzyme involved in biosynthesis of membrane sterols.
In a screening effort targeting Mycobacterium tuberculosis CYP51 (CYP51Mt), we previously identified the N-[4-pyridyl]-formamide moiety as a building block capable of delivering a variety of chemotypes into the CYP51 active site. In that work, the binding modes of several second generation compounds carrying this scaffold were determined by high-resolution co-crystal structures with CYP51Mt. Subsequent assays against the CYP51 orthologue in T. cruzi, CYP51Tc, demonstrated that two of the compounds tested in the earlier effort bound tightly to this enzyme. Both were tested in vitro for inhibitory effects against T. cruzi and the related protozoan parasite Trypanosoma brucei, the causative agent of African sleeping sickness. One of the compounds had potent, selective anti–T. cruzi activity in infected mouse macrophages. Cure of treated host cells was confirmed by prolonged incubation in the absence of the inhibiting compound. Discrimination between T. cruzi and T. brucei CYP51 by the inhibitor was largely based on the variability (phenylalanine versus isoleucine) of a single residue at a critical position in the active site.
CYP51Mt-based crystal structure analysis revealed that the functional groups of the two tightly bound compounds are likely to occupy different spaces in the CYP51 active site, suggesting the possibility of combining the beneficial features of both inhibitors in a third generation of compounds to achieve more potent and selective inhibition of CYP51Tc.
| Enzyme sterol 14α-demethylase (CYP51) is a well-established target for anti-fungal therapy and is a prospective target for Chagas' disease therapy. We previously identified a chemical scaffold capable of delivering a variety of chemical structures into the CYP51 active site. In this work the binding modes of several second generation compounds carrying this scaffold were determined in high-resolution co-crystal structures with CYP51 of Mycobacterium tuberculosis. Subsequent assays against CYP51 in Trypanosoma cruzi, the agent of Chagas' disease, demonstrated that two of the compounds bound tightly to the enzyme. Both were tested for inhibitory effects against T. cruzi and the related protozoan parasite Trypanosoma brucei. One of the compounds had potent, selective anti–T. cruzi activity in infected mouse macrophages. This compound is currently being evaluated in animal models of Chagas' disease. Discrimination between T. cruzi and T. brucei CYP51 by the inhibitor was largely based on the variability of a single amino acid residue at a critical position in the active site. Our work is aimed at rational design of potent and highly selective CYP51 inhibitors with potential to become therapeutic drugs. Drug selectivity to prevent host–pathogen cross-reactivity is pharmacologically important, because CYP51 is present in human host.
| The drug development pipeline targeting diseases caused by trypanosome parasites is sparse [1]. Despite significant advances in its control over the last 15 years [2], Chagas' disease, caused by the parasitic protozoan Trypanosoma cruzi [3], remains a major public health concern in Latin America, with an estimated total of 8 million people infected [4]. Nifurtimox and benznidazole, the two principal drugs for treatment of Chagas' disease, were launched in 1967 and 1972 respectively, and suffer from the twin liabilities of serious side-effects and reduced efficacy in chronic T. cruzi infections [2]. A potential new target for Chagas' disease chemotherapy is sterol 14α-demethylase (CYP51) [5], a cytochrome P450 heme thiolate-containing enzyme which is involved in biosynthesis of membrane sterols in all biological kingdoms from bacteria to animals [6]. T. cruzi sterols are similar in composition to those in fungi, with ergosterol and ergosterol-like sterols the major membrane components [7]. Clinically employed antifungal azoles [8],[9] inhibit ergosterol biosynthesis in fungi and are partially effective against Leishmania and Trypanosoma parasites [10]–[12]. Azoles block CYP51 activity, resulting in decline of the normal complement of endogenous sterols and accumulation of various 14α-methyl sterols with cytostatic or cytoxic consequences [11]. Aside from the compounds optimized for antifungal therapy, other CYP51 inhibitors with strong anti-T. cruzi activity have also been reported [13]–[15].
Mammalian CYP51 shares relatively modest overall sequence identity – below 30% – with its fungal and protozoan counterparts, but within the active site the amino acid residues are far more conserved. Based upon crystal structures of CYP51 of M. tuberculosis (CYP51Mt) [16]–[20], three of the thirteen active site residues, Y76, F83, and H259 (numbering according to CYP51Mt), are invariant throughout the cyp51 gene family. Two residues, F78 and F255, are specific to the methylation status of the C-4 atom in the sterol nucleus [18],[21], and amino acid identities of seven other positions strongly overlap across phyla [19],[20]. Of the thirteen residues, only one, R96, seems to be phylum-specific. This similarity confines design of selective CYP51 inhibitors to a species-specific cavity in the active site defined by the hydrophobic residues F78, L321, I322, I323, M433, and V434.
To discover novel inhibitors, we previously screened a library of small synthetic molecules against the CYP51Mt target [19]. The N-[4-pyridyl]-formamide moiety of the top hit, α-ethyl-N-4-pyridinyl-benzeneacetamide (EPBA), was found to bind unvaryingly in the CYP51 active site with Y76, H259, and the heme prosthetic group. The uniformity of interactions with CYP51 suggested that this scaffold could be used to target a variety of chemotypes to the active site. To verify this assumption, we determined the binding modes of second generation compounds containing the N-[4-pyridyl]-formamide moiety by determining their co-crystal structures with CYP51Mt. We also spectrally characterized binding of these compounds to CYP51 of both T. cruzi (CYP51Tc) and the related protozoan parasite T. brucei, the causative agent of African sleeping sickness, (CYP51Tb). Two compounds were selected based on their nanomolar binding affinities toward CYP51Tc and subsequently tested in vitro for inhibitory effects against both pathogens. One of the two compounds revealed potent and selective inhibitory effect against T. cruzi infection in mouse macrophage cells.
CYP51Mt double C37L/C442A and triple C37L/F78L/C442A mutants were prepared as described elsewhere [19]. The surface exposed cystein residues C37 and C442 were removed via replacement with leucine and alanine, respectively, to improve protein homogeneity and aid crystallization [18]. The functionally important F78 in the active site was replaced in the triple mutant by leucine, which invariantly occupies this position in the mammalian CYP51 isoforms.
Design of the CYP51Tc expression vector was based on an entity in the NCBI data bank (ID AY283022 [22]), which was modified by replacing the first 31 residues upstream of Pro32 with the fragment MAKKTSSKGKL from the CYP2C3 sequence [23] (CYP2C3 residues marked in bold) to improve protein solubility, and by inserting a His6-tag at the C-terminus to facilitate purification. This coding sequence (kindly provided by M. Waterman in the form of the pET vector) was subsequently sub-cloned into pCWori vector [24] between the NdeI and HindIII restriction sites and in this form used to transform Escherichia coli strain HMS174(DE3).
Transformants were grown for 5 h at 37°C and 250 rpm agitation in Terrific Broth medium supplemented with 1 mM thiamine, 50 µg/ml ampicillin, and trace elements. CYP51Tc expression was induced by the addition of isopropyl-β-D-thiogalactopyranoside (IPTG, final concentration 0.2 mM) and δ-aminolevulinic acid, a precursor of heme biosynthesis (final concentration 1 mM). Following induction, temperature was decreased to 25°C and agitation to 180 rpm. After 30 hours the cells were harvested and lysed by sonication. Insoluble material was removed from crude extract by centrifugation (30 min at 35,000 rpm). The supernatant was subjected to a series of chromatographic steps, including nickel-nitrilotriacetic acid (Ni-NTA) agarose (QIAGEN), followed by Q-Sepharose (Amersham Biosciences) in the flow-through regime, and then by S-Sepharose (Amersham Biosciences). From the S-Sepharose, protein was eluted in a 0.2 to 1.0 M NaCl gradient and observed by means of a 12% SDS-PAGE to be virtually homogeneous. Fractions containing P450 were combined, concentrated using a Centriprep concentrating device (Millipore), and stored at −80°C. Twenty mM Tris-HCl, pH 7.5, 10% glycerol, 0.5 mM EDTA, and 1 mM DTT were maintained throughout all chromatographic steps. Spectral characteristics of CYP51Tc are shown in Figure 1A.
The expression vector for CYP51Tb (ID EAN79583) was generated using T. brucei genomic DNA and upsteam GCGCGCATATGGCTCTTGAAGTTGCC and downstream CGCAAGCTTCTAGTGATGGTGATGGTGATGGTGATGAGCAGCTGCCGCCTTCC primers. The underlining denotes an NdeI restriction cloning site in the upstream primer and the HindIII restriction cloning site in the downstream primer followed by the stop codon. The bold sequence in the upstream primer highlights second codon replaced with alanine to optimize expression in E. coli cells [24]. The boldface in the downstream primer indicates the His8 tag. The original genomic DNA contained internal NdeI site at 345 base pair which was removed by introducing a silent mutation via the quick-change mutagenesis protocol (Stratagene). DNA amplification reaction was carried out as follows: 5 min at 94°C, annealing for 1 min at 55°C, and extension for 1 min at 72°C, for 30 cycles, followed by extension for 10 min at 72°C. The purified 1.5 kb PCR product was ligated into the pCR 2.1 TA cloning vector (Invitrogen). Insert was subsequently cleaved with NdeI and HindIII and ligated into pCWori vector digested with the same restriction enzymes and treated with alkaline phosphatase. The identity of the resulting vector was confirmed by DNA sequencing.
E. coli HMS174(DE3) strain was co-transformed with this vector and the pGro7 plasmid (Takara) encoding the E. coli chaperones GroES and GroEL. Double transformants were selected on agar plates containing both ampicilin and chloramphenicol. One liter of Terrific Broth medium supplemented with 1 mM thiamine, 100 µg/ml ampicillin, 40 µg/ml chloramphenicol, and trace elements was inoculated with 10 ml of overnight culture and growth continued at 37°C and 250 rpm agitation until OD600 reached 0.3. At that point expression of chaperones was induced with 0.2% arabinose. Growth continued at 27°C and 180 rpm until OD reached 0.6. Then CYP51Tb expression was induced by the addition of isopropyl-β-D-thiogalactopyranoside (IPTG, final concentration 0.3 mM), and δ-aminolevulinic acid (1 mM). Following induction, temperature was decreased to 15°C. After 48 hours the cells were harvested and lysed by sonication. Purification was conducted similarly to as described above for CYP51Tc, with the qualification that S-Sepharose was used in the flow-through regime, while the protein was bound to and eluted from the Q-Sepharose column. Spectral characteristics of CYP51Tb are shown in Figure 1B.
Five compounds (Fig. 2), purchased from ChemDiv (San Diego, California) were used for co-crystallization with the CYP51Mt C37L/C442A double mutant. Compared to the wild type, this construct has superior propensity for crystallization. Compound numbering is according to the order in which they were received in our laboratory, with number 7 being the first used in the current work. Ligands were dissolved in Me2SO at ≤100 mM stock concentration, and brought to final concentrations ranging from 1 to 5 mM in the crystallization mix, depending on ligand solubility. Protein concentration was 0.2 mM. A narrow crystallization screening grid (15–30% PEG 4000, 2–12% isopropanol, 0.1 M HEPES, pH 7.5), previously devised to obtain CYP51Mt crystals [16],[18],[19] was utilized for co-crystallization of complexes by the vapor diffusion hanging drop method. Four co-crystal forms were obtained, all diffracted to resolutions between 1.56 to 1.60 Å. Diffraction data were collected at 100–110 K at the Southeast Regional Collaborative Access Team (SER-CAT) 22ID beamline, Advanced Photon Source, Argonne National Laboratory using SER-CAT mail-in data collection program (Table 1). The images were integrated and the intensities merged with the HKL2000 software suite [25]. The structures were determined by molecular replacement using coordinates of estriol-bound CYP51Mt (Protein Data Bank ID 1X8V) as a search model. The final atomic models were obtained after a few iterations of refinement using REFMAC5 [26] and model-building using the COOT graphics modeling program [27]. The quality of the structures was assessed by the program PROCHECK [28]. One residue, A46, was found in the generously allowed region of the Ramachandran plot in all structures where, together with the adjacent G47, it enables a sharp turn between two β strands.
Spectroscopic binding assays were performed at room temperature in 1-ml quartz cuvette containing 1 µM or 2 µM CYP51 in 50 mM Tris-HCl, pH 7.5, and 10% glycerol using a Cary UV-visible scanning spectrophotometer (Varian). Concentration of CYP51 was determined at 450 nm from the difference spectra between the carbon monoxide-bound ferrous and water-bound ferric forms, with an extinction coefficient of 91,000 M−1 cm−1 [29]. In the first round, compounds dissolved in Me2SO at 10 mM concentration were added to the 2 µM protein solution in 0.5 µl aliquots, resulting in concentration increases from 5 µM to 50 µM in 5 µM increments. The same amounts of Me2SO alone were added to the protein in the reference cuvette, followed by recording the difference spectra. In the second round, compounds with high affinities were diluted to 100 µM by Me2SO and titrated into 1 µM protein solution in 1 µl aliquots to increase compound concentration from 0.1 µM to 2 µM in 0.1 µM increments. To determine the KD, we used the GraphPad PRISM software (Graphpad Software Inc.) to fit titration data to either rectangular or quadratic hyperbolas to correct for the bound ligand fraction, according to the functions ΔA = (Amax(S/KD+S) or ΔA = (Amax/2[E])((KD+[E]+[L])−((KD+[E]+[L])2−4[E][L])1/2), respectively, where E is total enzyme and L total ligand concentration, Amax the maximal absorption shift at saturation, and KD the apparent dissociation constant for the enzyme-ligand complex.
Irradiated (1000 rads) J774 mouse macrophages were plated in 12-well tissue culture plates 24 h prior to infection with 105 T. cruzi Y strain trypomastigotes for 2 h at 37°C. Cultures were maintained in RPMI-1640 medium with 5% heat-inactivated fetal calf serum and 5% CO2 with the addition of 10 µM compound 8 or 10. Untreated controls, controls treated with the inhibitor K11777 (10 µM) [30],[31], and uninfected macrophage controls were also included. All cultures were in triplicate and medium was replaced every 48 h. Treatment with CYP51 inhibitors continued for up to 27 days. Subsequently, treated cultures were maintained without inhibitor for an additional 13–15 days to confirm inhibitor effectiveness and cure of infected cells. Cultures were monitored daily by contrast phase microscopy to determine presence of T. cruzi infected cells and free infectious trypomastigotes (Table 2).
To determine IC50, mouse J774 macrophages were irradiated (1000 rads) to deter growth and plated onto 12-well tissue culture plates. Cells were infected with 105 tissue culture trypomastigotes of the Y strain of T. cruzi for 2 h at 37°C, as described above. Next, medium was replaced with the addition of compound 10 at 0, 1 nM, 10 nM, 100 nM, 500 nM, 1 µM, 5 µM, and 10 µM; these cultures were incubated for 52 h at 37°C. Controls with 10 µM K11777 and 10 µM compound 8 were also included. All treatments were performed in triplicate to ensure statistical validity. Cultures were then fixed in 4% paraformaldehyde in PBS for 2 h at room temperature and stained with DAPI (10 nM) in PBS. One hundred cells and their intracellular parasites were quantified as previously described to estimate the mean number of parasites/cell [32]. Mean P/cell data were plotted against compound concentration to estimate the IC50.
Toxicity was evaluated in bovine muscle cells (BESM), mouse J774 macrophages, and human Huh7 hepatocytes against compound 10 at 10 µM, 50 µM and 100 µM concentrations. After 48 h in culture at 37°C, cells were stained with 10% Tripan Blue and the number of live versus dead cells was quantified (Table 3).
Trypanosomes were grown in complete HMI-9 medium containing 10% FBS, 10% Serum Plus medium (Sigma Inc. St. Louis Mo. USA) and 1× penicillin/streptomycin. Trypanosomes were diluted to 1.0×105/ml in complete HMI-9 medium. Diluted trypanosomes were aliquoted in Greiner sterile 96-well flat white opaque culture plates using a WellMate cell dispenser (Matrix Tech., Hudson, NH, USA). Compounds 8 and 10 were serially diluted in Me2SO. Trypanosomes were incubated with the compounds for 48 h at 37°C with 5% CO2 before monitoring viability. Trypanosomes were then lysed in the wells by adding 50 µl of CellTiter-GloTM (Promega Inc., Madison, WI, USA). Lysed trypanosomes were placed on an orbital shaker at room temperature for 2 min. The resulting ATP-bioluminescence of the trypanosomes in the 96-well plates was measured at room temperature using an Analyst HT plate reader (Molecular Devices, Sunnyvale, CA, USA).
Co-crystals were obtained for compounds 8, 9 and 11. Compound 10 failed to generate any crystals with CYP51Mt. Compound 7 was not found in the CYP51Mt active site in the crystal, which is consistent with lack of spectrally detectable binding (Fig. 2). Compounds 8 (3-{[(4-methylphenyl)sulfonyl]amino}propylpyridin-4-ylcarbamate), 9 (cis-4-methyl-N-[(1S)-3-(methylsulfanyl)-1-(pyridin-4-ylcarbamoyl)propyl]cyclohexanecarboxamide), and 11 (N-[(1S)-2-methyl-1-(pyridin-4-ylcarbamoyl)propyl]cyclohexanecarboxamide), were observed bound in the CYP51Mt active site as predicted, through the coordination of the heme iron via a lone pair of aromatic nitrogen electrons of the N-[4-pyridyl]-formamide moiety (highlighted in gray in Fig. 2) and interactions with the invariant residues Y76 and H259 (Fig. 3). Functional groups other than the N-[4-pyridyl]-formamide moiety in compounds 9 and 11 either were accommodated in the species-specific cavity or else protruded through the opening of the active site toward bulk solvent. H259 hydrogen-bonded to the carbonyl oxygen in both compounds, while interactions with Y76 were mediated by two similarly positioned water molecules (Figs. 3A and 3B). The residual Fo-Fc electron density map suggested two alternative conformations for compounds 11 and 9, designated by pink and cyan respectively in Figures 4A and 4B. In the CYP51Mt-compound 11 complex, the cyclohexane ring protruded toward the bulk solvent (Fig. 3A), barely interacting with the protein in two alternative conformations (Fig. 4A). Together with the limited interactions of the isopropyl moiety, this lack of contact explains the low binding affinity of 11. In the CYP51Mt-compound 9 complex, the methylcyclohexane moiety protruded toward bulk solvent, while the methylsulfanyl group loosely bound in the species-specific cavity (Fig. 3B) in two alternative conformations (Fig. 4B). The side chain of M433 also adopted two alternative conformations. In both complexes, a portion of the BC-region was disordered and missing from the electron density map. Although racemic mixtures were used for co-crystallization, only one enantiomer of each compound was found in the active site.
A different binding mode was revealed for compound 8. Its flexible backbone allowed it to fold head-to-tail over the heme plane to bring the methylphenylsulfonamide group into intramolecular stacking interactions with the pyridinyl moiety and also with the heme macrocycle (Fig. 3C, 4C). Folding minimized the nonpolar surface of compound 8 by exposing the sulfonamide group to interactions with Q72, K97, and the heme propionate side chain. The hydrophobic side chain of K97 aligned along the methylphenyl moiety. A similar folding of the benzothiadiazolsulfonamide group has been observed in previous work for 2-[(2,1,3-benzothiadiazol-4-sulfonamide]-2-phenyl-N-pyridin-4-acetamide (BSPPA) [19]. Mutually stabilizing protein-ligand interactions involving the BC-loop residues including F78 result in increased binding affinity of the CYP51Mt-compound 8 complex and in unambiguous electron density both for compound 8 (Fig. 3C) and for the entire BC-region. In the CYP51Mt-compound 8 complex, H259 directly H-bound to the amide nitrogen of compound 8, whereas Y76 interacted hydrophobically with the compound's flexible backbone (Fig. 3C).
Binding affinities of all five compounds were examined against both wild type and a ‘humanized’ F78L mutant form of CYP51Mt, CYP51Tc, and CYP51Tb using spectroscopic assays (Fig. 5). These assays utilize the property of P450 enzymes to shift the ferric heme iron Soret band following replacement of a weak ligand, the water molecule, with a stronger one, the nitrogen-containing aromatic pyridinyl group (Fig. 5A). All compounds had markedly reduced or no binding affinity toward CYP51Mt, compared to the parental EPBA (Fig. 2). No binding was observed for compound 7, while the KD for compound 11 exceeded 100 µM, indicating weak binding.
However, the binding affinity of all compounds examined, including compound 7, was significantly higher to CYP51Tc than to CYP51Mt. Remarkably, the binding affinities of compounds 8 and 10 to CYP51Tc were 300- and at least 500-fold respectively higher, equaling or exceeding that of the antifungal CYP51 inhibitor fluconazole, which was used as a reference (Fig. 5B, C). A KD of at least 40 nM was estimated for compound 10 by spectral assays, with the binding curve reaching a plateau at about a 1∶1 protein to ligand ratio. This value strongly suggests that the KD must be notably higher, although further dilution of protein in an attempt to obtain a more accurate value significantly decreased the quality of the spectra and this effort was thus abandoned. The IC50 of ∼1 nM for T. cruzi intracellular growth inhibition, determined for compound 10 as described below, may better reflect true KD value. Compounds 8 and 10, which had highest binding affinity to CYP51Tc, were spectrally silent toward CYP51Tb, indicating no binding in the active site (Fig. 2). As expected, CYP51Tb had nanomolar affinity for fluconazole (Fig. 5D), but again, the plateau was reached at a 1∶1 protein to inhibitor ratio, so the binding constant could not be determined more accurately. Compound 9 bound both CYP51Tc and CYP51Tb with the same affinity.
The KD values for compounds 8 and 10 slightly decreased for the F78L CYP51Mt mutant compared to the wild type, while the KD values for the other compounds increased (Fig. 2).
With submicromolar affinities toward CYP51Tc of 160 nM and <40 nM respectively, compounds 8 and 10 were examined in vitro for inhibitory effects against both T. cruzi and T. brucei. In a mouse macrophage assay, T. cruzi completed its intracellular development in 5 days in untreated controls, resulting in death of host macrophages and abundant trypomastigotes in culture supernatant (Table 2). As anticipated, the control compound K11777 [30] cured T. cruzi infection. No parasites survived a treatment regime of 27 days with compound 10. Cure of host cells was confirmed by incubation of the cultures for an additional 15 days in the absence of inhibitor. In contrast, and similarly to untreated controls, T. cruzi completed its development in 5 days in cultures treated with compound 8.
An IC50 of ∼1 nM concentration for compound 10 (Fig. 6) was estimated for T. cruzi intracellular amastigotes. T. cruzi developed well intracellularly in untreated macrophages with a final mean number of 3.57±0.5 P/cell (0% inhibition). As determined previously, 10 µM compound 10 was deleterious for T. cruzi, with a mean of 0.25±0.01 P/cell (100% growth inhibition). Ten µM of control compound K11777 was also parasiticidal for T. cruzi with a mean of 0.25±0.01 P/cell (IC100) [30], while compound 8 was not parasiticidal at this concentration with a mean of 1.22±0.1 P/cell (data not shown).
Toxicity for mammalian cells was addressed by treating the three different cell types with increasing concentrations of compound 10 (Table 3). No toxicity was observed at 10 µM compound 10, while 50 µM was mildly toxic for muscle cells. One hundred µM compound 10 was toxic for all mammalian cells tested, especially muscle cells.
Consistent with the spectral binding assays, neither compound 8 nor 10 had any inhibitory effects against cultured T. brucei even at the highest tested concentration of 10 µM.
The atomic coordinates and structure factors determined in this study (Protein Data Bank IDs 2W09, 2W0A, and 2W0B) have been deposited in the Protein Data Bank, Research Collaboratory for Structural Bioinformatics, Rutgers University, New Brunswick, NJ (http://www.rcsb.org/).
We explored sterol 14α-demethylase (CYP51) as a potential target for trypanosomiasis chemotherapy by probing CYP51Mt, CYP51Tc, and CYP51Tc with second generation compounds that contain a universal building block, the N-[4-pyridyl]-formamide moiety, which is capable of delivering small molecule compounds to the CYP51 active site. The affinities of the N-[4-pyridyl]-formamide-derivative compounds that we tested against CYP51Mt were lower than that of EPBA (Fig. 2), from which the formamide building block was derived. Affinities of all compounds examined were much higher toward CYP51Tc than to CYP51Mt. Strikingly large increases in binding affinities – 300 and 500 fold – were observed for compounds 8 and 10. Although compound 10 did not produce crystals with CYP51Mt, based on the binding modes of compounds 9 and 11, we reason that the methylcyclohexanecarboxamide moiety of compound 10 protrudes toward the BC-loop, suggesting that the indole ring binds in the species-specific cavity, including the space occupied in CYP51Mt by the F78 aromatic ring, which is absent from CYP51Tc but present in CYP51Tb and CYP51Mt. Consistent with this hypothesis, compound 10 selectively bound CYP51Tc, inhibited T. cruzi growth with the IC50 value close to the KD estimated in the spectral binding assays, and cured mouse macrophages infected with T. cruzi Y strain at 10 µM concentration without harming them.
In contrast, compound 10 failed to bind CYP51Tb despite the identity of 12 of the 13 active site substrate binding residues, and 83% overall sequence identity between T. cruzi and T. brucei CYP51 orthologues. This result is a striking indication of the sensitivity of CYP51 to alterations of the topography of its active site at position 78. The difference in position 78 is of functional importance, because phenylalanine at this site is strictly specific to protozoa and plant species metabolizing 4α-methylated sterols [18]. Interestingly, T. cruzi is the only protozoan where the corresponding position (position 105 according to T. cruzi numbering) is occupied by isoleucine. Consistent with this observation, CYP51Tc is catalytically more closely related to its fungal and animal orthologues, preferentially converting 4α,β-dimethylated sterol substrates [21], whereas T. brucei CYP51 is strictly specific to 4α-methylated obtusifoliol and norlanosterol [33]. The proteobacterium Methylococcus capsulatus, known to synthesize sterols from squalene [34], is the only other known organism having isoleucine in the CYP51 position corresponding to F78. Not surprising, compound 10 was inactive against T. brucei in inhibitory assays in vitro.
In humans and animals metabolizing 4α,β-dimethylated 24,25-dihydrolanosterol, position 78 is always occupied by leucine. Therefore, the F78L substitution in the CYP51Mt binding site was examined and found to slightly increase binding affinities toward compounds 8 and 10, as opposed to the rest of the compounds whose binding affinities decreased (Fig. 2). Although a single amino acid substitution does not by any means convert bacterial protein into its mammalian counterpart, this finding is consistent with lack of toxicity in mammalian cells at inhibitory concentrations, and supports the possibility of rational design of highly selective anti-protozoan CYP51 inhibitors. The latter is of particular pharmacological importance as far as host-pathogen cross-reactivity is concerned, since CYP51 is present in human host.
The increased binding affinities toward CYP51Tc of all the compounds we tested may indicate more extensive involvement of the BC-loop and C helix in protein-inhibitor interactions in CYP51Tc than in CYP51Mt. Assuming that compound 8 binds CYP51Tc in a similarly compact donut-like shape that fills the space adjacent to the porphyrin ring, its 300-fold increase in binding affinity could be achieved solely by stabilization of the BC-region of CYP51Tc without engaging the species-specific cavity. This possibility opens the door to a rational design effort in which the beneficial features of both compounds 8 and 10 would be combined to yield third generation compounds that would more potently and selectively inhibit CYP51Tc. Toward this end compound 10 is currently being evaluated in animal models of Chagas' disease.
|
10.1371/journal.pntd.0005212 | Implementation of Syndromic Surveillance Systems in Two Rural Villages in Senegal | Infectious diseases still represent a major challenge for humanity. In this context, their surveillance is critical. From 2010 to 2016, two Point-Of-Care (POC) laboratories have been successfully implemented in the rural Saloum region of Senegal. In parallel, a homemade syndromic surveillance system called EPIMIC was implemented to monitor infectious diseases using data produced by the POC laboratory of the Timone hospital in Marseille, France. The aim of this study is to describe the steps necessary for implementing EPIMIC using data routinely produced by two POC laboratories (POC-L) established in rural Senegal villages. After improving EPIMIC, we started to monitor the 15 pathogens routinely diagnosed in the two POC-L using the same methodology we used in France. In 5 years, 2,577 deduplicated patients-samples couples from 775 different patients have been tested in the Dielmo and Ndiop POC-L. 739 deduplicated patients-samples couples were found to be positive to at least one of the tested pathogens. The retrospective analysis of the Dielmo and Ndiop POC data with EPIMIC allowed to generate 443 alarms. Since January 2016, 316 deduplicated patients-samples couples collected from 298 different patients were processed in the Niakhar POC laboratory. 56 deduplicated patients-samples couples were found to be positive to at least one of the tested pathogens. The retrospective analysis of the data of the Niakhar POC laboratory with EPIMIC allowed to generate 14 alarms. Although some improvements are still needed, EPIMIC has been successfully spread using data routinely produced by two rural POC-L in Senegal, West Africa.
| Infectious diseases are still a major public health problem for humanity, especially because of their unpredictability. In this context, developing surveillance systems is one of the most effective solutions to fight them, including in Sub-Saharan African countries where the situation is particularly worrying. This study describes how automated surveillance systems have been implemented in two rural areas in Senegal for the weekly surveillance of 15 pathogens responsible for fever in Africa and on the basis of laboratory data from two rural laboratories. In addition, the first results of this surveillance are presented. The two surveillance systems prospectively and retrospectively detected several abnormal events, and allowed us to observe that the distribution of the pathogens was not the same depending on the region under surveillance. This study provides evidence on the possibility of developing surveillance systems in rural areas in Africa. Such initiatives are needed, especially to improve our global knowledge on infectious diseases in this continent.
| In their 2015 report, the Global Burden of Disease study group estimated that in 2013 54.9 million people died worldwide, with 11.8 million deaths due to communicable, maternal, neonatal, and nutritional disorders [1]. Infectious diseases were directly involved in a significant part of them, with 2.7 million deaths due to lower respiratory infections, 1.3 million deaths due to HIV/AIDS, 1.3 million deaths due to tuberculosis, 1.3 million deaths due to diarrhoeal diseases, and 854,600 deaths due to malaria [1]. This situation clearly underlines that infectious diseases are still a big challenge for humanity in the 21st century, especially because i) we still discover more and more possible pathogens with new technologies, ii) pathogens are naturally evolving, leading to the emergence or re-emergence of pathogens, iii) people and resources are moving and exchanging faster and faster around the world, which directly affects the ecosystems, and iv) their appearance and disappearance in the different human populations cannot be reliably modeled [2–5].
Facing this situation, various strategies have been developed worldwide and over the time to try to monitor infectious diseases and pathogens. In this way, three main strategies of surveillance are currently extensively used for the monitoring of infectious diseases [6]: i) disease-specific surveillance allowing the surveillance of pathogens, syndromes or risk exposures defined to be public health threats in a precise population, ii) syndromic surveillance using non specific indicators not collected for surveillance purposes to be used for the early real-time identification of the impact (or non impact) of possible health threats, and iii) event-based surveillance using unstructured information from the Internet for the real-time or near real-time detection of potential or confirmed health events occuring in the world. Moreover, thirteen main sources of data are currently available for the infectious disease surveillance, including, among others, the Internet, drug sales reports, sentinel surveillance notifications, notifiable diseases reports, and microbiology orders reports.
The current and past impact of infectious diseases on humanity have also promotted the developpement and spread of new microbiology technologies and laboratories. Among the most innovative improvements made can be particularly mentioned Point-Of-Care (POC) laboratories (POC-L). They consist in on-site around the clock operating laboratories equipped to perform a wide variety of rapid diagnostic tests that are able to deliver rapid diagnosis [7,8]. Therefore, they can help to take adequate decisions for hospitalization, isolation and therapy in only few hours [7]. Moreover, these laboratories do not require higly specialized skills and equipments, making them potentially cheaper and more cost-effective than conventional laboratories, but equally easier to implement in place with low resources [8].
Sub-Saharan Africa is a 24 million km2 area that includes more than 1 billion people from 48 different countries (http://data.worldbank.org/). Although the situation is improving, with large gains in life expectancy mainly due to reductions of diarrhoea and lower respiratory infections, HIV/AIDS, malaria and tuberculosis still remained important cause of deaths in this area, especially in the southern, western and eastern sub-Saharan African countries [1]. Moreover, while the number of child deaths decreased significantly all around the world from 1990 to 2013, this number has only slightly decreased from 3,7 million (3,6–3,7) in 1990 to 3,2 million (3–3,4) in 2013 in the sub-Saharan Africa area, especially because of malaria, diarrhoeal diseases, lower respiratory infections, HIV/AIDS and measles [1].
From 2010 to 2016, two POC-L have been successfully implemented in the rural Saloum region of Senegal [9]. Using our 15-years experience of infectious diseases surveillance in the Assistance Publique- Hôpitaux de Marseille (AP-HM) institution, we recently decided to implement our homemade syndromic surveillance system EPIMIC (for EPIdemiological surveillance and alert based on MICrobiological data) [10] in Senegal to monitor infectious diseases diagnosed in the two rural POC-L using the data routinely produced by these laboratories. The aim of this study is to describe the implementation ad results of EPIMIC in a developing country using data routinely produced by two POC-L established in rural Senegal villages.
The two POC-L have been implemented in two different sites (Fig 1), one located in the research station of Dielmo and deserving two rural villages in the Sine-Saloum region (Dielmo and Ndiop), and the other in the research station of the rural village of Niakhar. The first POC laboratory has been already extensively described elsewhere [9]. The second study site is located in the Fatick region of Senegal 155km South-East of Dakar. This site theoretically involves a ~48 000 inhabitants distributed in 30 villages covering a 230km2 area.
The functioning of the POC laboratory of Dielmo/Ndiop is similar to that previously extensively described by Sokhna et al. [9], and is now equipped with a reliable Internet connexion. For the Niakhar POC laboratory, the way of functioning is different although its equipment is similar to that of Dielmo/Ndiop [9]. Thus, as the area and the population covered by this POC laboratory is far more important than that of Dielmo/Ndiop villages, this POC laboratory is working in close collaboration with 4 health structures (Niakhar, Diohine, Toucar and Ngayokhem) located in some of the biggest villages of the area (Fig 1). To sum up, 5 days on 7, the technician from the POC laboratory visits all the health strucutures in motorcyle to collect in iceboxes the samples collected by the nurses from the patients visiting the 4 different health structures mentioned above. Then, the samples are tested by the technician in the POC laboratory to look for the possible causative agent responsible for the illness of the different patients. When the tests are performed, the technician completes a preformated Microsoft Excel worksheet summarizing the main information required to properly identify each patient. The same procedure is followed in the Dielmo/Ndiop POC laboratory. The information collected in the two POC-L are structurally similar (Table 1). The POC laboratory of Niakhar is also equipped with a reliable Internet connexion.
EPIMIC has been previously extensively described elsewhere [10]. Briefly, EPIMIC is a Microsoft Excel based syndromic surveillance system allowing the weekly automated surveillance of a large panel of pathogens. Thus, it allows the real-time monitoring of the number of samples tested and positives performed every week. The percentage of positivity for each test performed is also automatically calculated. The new values entered in the software are then automatically compared to their historical mean number +/- two standard deviations (thresholds), and alarms are automatically emitted by EPIMIC when the last entered values (the number of tested samples, the number of positive samples and the percentage of positiveness) exceed the thresholds. Plots are then automatically produced by the software.
Major improvements have been performed in the current EPIMIC version since its first description. All these improvements have been made using the Visual Basic programming language. The current version of EPIMIC consists in a single soft Microsoft Excel file which can be simply configured by following a step-by-step procedure proposed by the software during its first use. After this initial configuration, EPIMIC automatically collects and analyzes data from the different POC Microsoft Excel databases without any manual contributions, avoiding any inputs errors. If duplicates are present for the same week, the software only keeps the sample from which the most important number of positive tests have been identified. Moreover, the software automatically produces a Microsoft PowerPoint summary of the epidemiologic situation in the two POC-L using plots and screen shots when alarms are emitted for pathogens monitored by the software (Fig 2). In this way, data are anonymized to garanty the confidentiality of the patients’ data. Furthermore, the software separates the data according the place they come from (per villages for the Dielmo and Ndiop database, and per health structures for the Niakhar database), facilitating the weekly analysis of the epidemiological events. Finally, the software is able to rapidly produced a precised activity report summarizing the activity of the surveillance system for each of the POC-L.
Once a week, the two Microsoft Excel databases are sent by the technicians to the coordinator of the epidemiological surveillance located in the Institut de Recherche pour le Développement (IRD) de Dakar. The databases are visually checked to avoid any inputs error. Then, EPIMIC is used to analyse the databases. The analysis is simple and rapid (less than 5 minutes per databases in routine).
Once analysed, the two Microsoft PowerPoint summaries are sent to a group of Senegalese and French specialists of the Institut Hospitalo-Universitaire Mediterranée Infections (IHU) of Marseille and of the Institut de Recherche pour le Développement (IRD) for analysis (Fig 3). If alarms are validated, indicating that the abnormal epidemiological event detected by the EPIMIC is possibly a true know or unknown epidemiological event, investigations are conducted to determine the veracity of the epidemiological events underlined by the alarms, and take countermeasures if necessary.
We performed a PubMed French-English literature search from 1990 to August 2016 in order to identify laboratory-based surveys for infectious diseases in Africa (S1 Table).
The project was initially approved by the Ministry of Health and Preventive Medicine of Senegal and the assembled village population in 1990. The project was ethically approved in 2000, 2005 and 2009 (identification number of the ethical and scientific evaluation: SEN21/09 and SEN37/09). Since that, the National Ethics Committee of Senegal has conducted three site visits, and ad-hoc research committees of the Ministry of Health and Preventive Medicine, the Dakar Pasteur Institute and the IRD provided recommendations to continue the project. Moreover, ethical approval is renewed on a yearly basis. After careful explanation of the goals of the project to the assembled village population, written informed consent was obtained after POC diagnosis from all of the adult residents in the study villages and from the guardians of children under 15 years of age.
Table 2 summarizes the data from the two POC-L databases and their EPIMIC analysis.
The full-time activity of the Dielmo and Ndiop POC laboratory started 5 years ago in February 2011, i.e 278 weeks ago [9]. Over this period, 2,577 deduplicated patients-samples couples from 775 different patients have been tested to look for the presence of at least one of the 15 causes of fever (Plasmodium falciparum, flu, dengue, Coxiella burnetii, Leptospira spp., Bartonella spp. (including B. quintana), Tropheryma whipplei, Borrelia spp., Rickettsia spp. (including R. prowazekii, R. conorii, R. africae, R. felis), Salmonella spp., Streptococcus pneumoniae, and Staphylococcus aureus) detected by the 27 POC tests routinely performed in the POC laboratory. The mean number of patients weekly tested was 10 patients (range: between 1 to 36 patients per week). 739 deduplicated patients-samples couples were found to be positive for at least one of the tested pathogens. The retrospective analysis of the Dielmo and Ndiop POC laboratory historical database with EPIMIC virtually allowed to identify 443 alarms, 373 before April 2016 and 70 since April 2016. P. falciparum was the pathogen the most cited by alarms emitted by EPIMIC (47 alarms, 2,635 tested samples, and 301 positive samples), followed by R. felis (43 alarms, 2,411 tested samples, and 21 positive samples), flu (39 alarms, 2,237 tested samples, and 135 positive samples), C. burnetii (35 alarms, 2,513 tested samples, and 14 positives), and Borrelia spp. (34 alarms, 2,577 tested samples, and 126 positives) (Table 2).
The POC laboratory of Niakhar has been set up in January 2016, i.e 26 weeks ago. Over this period, this POC laboratory tested 316 deduplicated patients-samples couples collected from 298 different patients to look for the presence of the same 15 pathogens using the same POC tests as those mentioned above. On average, 12 patients were weekly tested (range: between 3 to 26 patients per week). These tests allowed to identify 56 deduplicated patients-samples couples positive to at least one of the tested pathogens. The retrospective analysis of the historical database of the Niakhar POC laboratory with EPIMIC virtually allowed to identify 14 alarms, 8 before May 2016 and 6 since May 2016. Borrelia spp. was the pathogen the most cited by alarms emitted by EPIMIC (4 alarms, 315 tested samples, and 36 positive samples), followed by P. falciparum (3 alarms, 316 tested samples, and 5 positive samples) (Table 2).
The PubMed literature search initiated in this paper allowed us to identify 16 infectious diseases laboratory-based surveys and/or surveillance systems in Africa (S1 Table). Most of them (5, 31%) were implemented in South Africa, 3 (18.7%) in Egypt, and the 8 (50%) others in 8 different countries (1 (6.2%) in Burkina Faso, 1 in Kenya, 1 in Senegal, 1 in Sudan, 1 in Togo, 1 in Tunisia, 1 in Uganda, and 1 in Zimbabwe). Seven (43.8%) of these surveillance systems were totally devoted to the surveillance of bacteria, 3 (18.7%) to the surveillance of viruses, 2 to the surveillance of parasites (12.5%), and 1 (6.2%) to the surveillance of fungus.
EPIMIC was firstly implemented in the clinical microbiology laboratory of the Timone hospital, Marseille, France, in order to detect abnormal epidemiological events using data from both routine and POC-L [10]. Thereafter, three other laboratory-data based surveillance systems have been developed to complete this surveillance [11] (Fig 3). As EPIMIC was able to detect numerous true epidemiological events without heavy human and economic resources in Marseille [10], we decided to test if it was possible to effectively spread it in the West African country of Senegal using data routinely produced by rural POC-L (Figs 1 and 3).
The effective implementation of EPIMIC in rural Senegal was performed in less than one month, demonstrating the great scalability and versatility of the system. This is especially due to the fact that it was developed on Microsoft Excel, with low human and computer resources, using a simple algorithm for the detection of abnormal epidemiological events, and with no need for data specifically produced for surveillance as the software was specifically developed for syndromic surveillance [6,10]. However, this implementation required major improvements related to local needs. Thus, as mentioned before, the software was fully automated for the collection, anonymisation and analysis of the data, and for the production of a simple Microsoft PowerPoint activity report (Fig 2). This improvement considerably decrease the level of complexity to use the software including for people without in-depth computer skills, which is crucial to ensure its optimal long-term use [11]. Moreover, it includes a simple step-by-step easy procedure making it able to rapidly evolve with changes with no need to deeply modifiy it (for example the addition or removal of pathogens to be monitored by the system). Finally, EPIMIC is currently able to geographically localize the patients according to their villages or the health structures they visited, facilitating the on-site investigations of the validated abnormal events.
The local set-up of the software was strongly facilitated by the fact that the two POC-L have been implemented and routinely used from some time (Tables 1 and 2) with laboratory technicians trained to fill two preformatted Microsoft Excel databases, avoiding bad quality data and its consequence on syndromic surveillance [12]. Moreover, the two sites are regularly supplied with reagents necessary to perform the different POC tests, which is crucial for continuous detection and monitoring of pathogens [8,9]. Finally, they are equiped with Internet connexions, preventing disruptions in the weekly sending of microbiogical data.
By the past, numerous studies have been performed in the Dielmo/Ndiop and Niakhar areas and they allowed us to identify that some emerging pathogens like Coxiella burnetii, Bartonella spp., Tropheryma whipplei, Borrelia spp., and Rickettsia spp were responsible for unexplained fevers in these areas [13–21]. However, these studies did not have for objective to monitor infections by these pathogens over periods in time in the area. This is now possible with the two EPIMIC surveillance systems. Thus, by counting the weekly number of patients infected by these emerging pathogens in the two rural areas under surveillance and by detecting abnormal events related to these pathogens, the EPIMIC systems allowed us to collect and analyze the information that will be used to improve our knowledge on the local epidemiology of the pathogens of interest, including the emerging ones. In this way, the fact that Borrelia spp. has been identified to be the pathogen the most cited by alarms triggered by EPIMIC from January 2016 to June 2016 in Niakhar (Fig 4) is already an unexpected information. Moreover, EPIMIC also provides data on the weekly number and proportion of infections that remain unexplained. These points make us strongly believe that these two surveillance systems will be rapidly of great interest locally for the detection and management of infectious diseases, including to implement and evaluate the onsite impacts of pre-existing and future prevention plans.
Comparing our surveillance with other surveillance implemented in Africa over the two last decades (S1 Table), we identified that the main asset of EPIMIC is the number of POC tests and pathogens it currently allows to survey in near real-time (27 POC tests for 15 pathogens (bacteria, viruses and parasites)). Nevertheless, further improvements are needed to enhance our EPIMIC syndromic surveillance network. Indeed, as mentioned before, EPIMIC has been implemented using the Microsoft Excel software, and abnormal events are currently detected using statistical tools based on the historical mean of the data +/- two standard deviations. Although this strategy allows the rapid development of the system and was valuable in terms of human and economic resources [11], this is not suitable for the monitoring of big databases including data with seasonal variations in pathogen isolation like flu and P. falciparum, or in case of rare pathogens. These issues are planed to be addressed with the development of a local web-based plateform including more sophisticated statistical tools for the accurate monitoring of abnormal events [22–24]. We also plan to integrate local Health Agencies to our surveillance network (Fig 3) in order to actively participate to the Senegal national surveillance of infectious diseases like it is the case in France [11].
To conclude, EPIMIC has been successfully spread using data routinely produced by two rural POC-L in Senegal, West Africa. We are clearly convinced that such initiatives, like the implementation of rural POC-L [13,20], are needed to improve our global knowledges on infectious diseases, whatever the level of knowledge we currently have on them.
|
10.1371/journal.ppat.1003140 | Functional Characterization of HLA-G+ Regulatory T Cells in HIV-1 Infection | Regulatory T cells represent a specialized subpopulation of T lymphocytes that may modulate spontaneous HIV-1 disease progression by suppressing immune activation or inhibiting antiviral T cell immune responses. While the effects of classical CD25hi FoxP3+ Treg during HIV-1 infection have been analyzed in a series of recent investigations, very little is known about the role of non-classical regulatory T cells that can be phenotypically identified by surface expression of HLA-G or the TGF-β latency-associated peptide (LAP). Here, we show that non-classical HLA-G-expressing CD4 Treg are highly susceptible to HIV-1 infection and significantly reduced in persons with progressive HIV-1 disease courses. Moreover, the proportion of HLA-G+ CD4 and CD8 T cells was inversely correlated to markers of HIV-1 associated immune activation. Mechanistically, this corresponded to an increased ability of HLA-G+ Treg to reduce bystander immune activation, while only minimally inhibiting the functional properties of HIV-1-specific T cells. Frequencies of LAP+ CD4 Treg were not significantly reduced in HIV-1 infection, and unrelated to immune activation. These data indicate an important role of HLA-G+ Treg for balancing bystander immune activation and anti-viral immune activity in HIV-1 infection and suggest that the loss of these cells during advanced HIV-1 infection may contribute to immune dysregulation and HIV-1 disease progression.
| HIV-1 causes disease by inducing a chronic inflammatory state that leads to progressive CD4 T cell losses and clinical signs of immune deficiency. Regulatory T cells (Treg) represent a subgroup of T lymphocytes with immunosuppressive activities that can reduce HIV-1 associated immune activation, but may also worsen HIV-1 disease progression by inhibiting T cell responses directed against HIV-1 itself. Here, we describe a non-classical population of regulatory T cells that differ from conventional Treg by the expression of HLA-G, a molecule that contributes to maternal tolerance against semiallogeneic fetal tissue during pregnancy. We show that HLA-G-expressing Treg have a unique functional ability to reduce harmful bystander immune activation, while minimally inhibiting potentially beneficial T cell-mediated immune responses against HIV-1. In this way, HLA-G-expressing Treg may represent a previously unrecognized barrier against HIV-1 associated immune activation and a possible target for future immunotherapeutic interventions in HIV-1 infection.
| The hallmark of HIV-1 infection is a progressive reduction of CD4 T cells. The main function of these cells is to provide antigen-specific helper cell activity against a wide panel of microbial antigens, however, some of these cells also have regulatory immunosuppressive activities. Classical regulatory T cells (Treg) are immunophenotypically defined as being CD25hi and CD127lo, and they intracellularly express the Forkhead Box P3 protein (FoxP3) [1]. The importance of classical Treg for maintaining immune homeostasis has been highlighted by signs of autoimmune pathology that occur in the setting of deficient Treg activity [2], [3]. During progressive HIV-1 infection, the relative frequency of classical Treg is increased, while their absolute counts are reduced as a consequence of lower total CD4 T cell counts [4]. This indicates that classical Treg decline at a slower rate than conventional CD4 T cells during progressive HIV-1 infection, and suggests that these cells may play an important role in the immune pathogenesis of HIV-1 infection. Functional data from previous studies indeed demonstrated that classical Treg can potently suppress HIV-1-specific T cell responses [5], [6], and in this way may contribute to the failure of achieving T cell-mediated immune control of HIV-1 replication. However, classical Treg may also have beneficial effects on HIV-1 disease progression by reducing the deleterious consequences of HIV-1 associated immune activation [7], [8].
Recently, several alternative Treg populations have been identified that differ from classical Treg by the lack of intracellular FoxP3 expression. One group of such non-classical Tregs is defined by surface expression of HLA-G [9], an HLA class Ib molecule that is mainly expressed on placental trophoblasts. However, ectopic expression of HLA-G can also be observed on small populations of peripheral blood CD4 and CD8 T cells, which seem to be enriched at sites of inflammation [9]. These cells have the ability to suppress proliferation of T lymphocytes in a cell-contact independent manner, and their regulatory effects are reversible following neutralization with HLA-G blocking antibodies [10]. Previous reports suggested that the proportion of HLA-G-expressing CD8 T lymphocytes is increased during HIV-1 infection [11], however, such investigations were conducted in unselected populations of HIV-1 positive persons, and did not address the functional role of HLA-G+ T cells during different stages of HIV-1 disease progression.
A second group of non-classical Tregs is characterized by surface expression of the latency-associated peptide (LAP), a membrane bound form of TGF-β [12]. These LAP+ CD4 T cells lack FoxP3 expression but can inhibit proliferative activities of T lymphocytes in vitro and in vivo. Under physiologic conditions, a small proportion of LAP-expressing CD4 T cells can be detected in human peripheral blood [12]. The numeric distribution and functional role of LAP+ CD4 Treg during HIV-1 infection is not known.
In the present study, we systematically analyzed the expression and function of HLA-G-and LAP-expressing Tregs in patients with different stages of HIV-1 disease infection. Our results indicate a profound reduction of HLA-G+ CD4 Treg in individuals with progressive HIV-1 disease that may stem from a higher susceptibility of these cells to HIV-1 infection, and functionally contribute to HIV-1-associated immune overactivation.
HIV-infected patients and HIV-1 seronegative control persons were recruited according to protocols approved by the Institutional Review Board of the Massachusetts General Hospital in Boston. Samples of mononuclear cells extracted from lymph nodes and peripheral blood were obtained from HIV-1 infected study patients recruited at the University of Hamburg (Germany) according to a protocol approved by the local Ethics Committee.
All subjects gave written informed consent and the study was approved by the Institutional Review Board of Massachusetts General Hospital/Partners Healthcare.
Peripheral blood mononuclear cells (PBMC) were isolated from whole blood using Ficoll density centrifugation. Lymph node mononuclear cells (LNMC) were extracted from freshly-excised lymph node samples according to routine procedures. PBMC or LNMC were stained with LIVE/DEAD cell viability dye (Invitrogen, Carlsbad, CA) and monoclonal antibodies directed against CD4, CD25, CD127, CD45RA, CCR7 (BD Biosciences, San Jose, CA), CD57 and PD-1 (Biolegend, San Diego, CA), CD8 (Invitrogen), HLA-G (clone MEM-G/9, Abcam, Cambridge, MA), LAP (clone 27232, R&D systems, Minneapolis, MN) and, when indicated, LILRB1 (clone HP-F1, ebioscience, San Diego, CA). After incubation for 20 minutes at room temperature, cells were fixed with PBS containing 0.5% fetal calf serum and 1% formaldehyde. Anti-FoxP3 antibodies (ebioscience) were used with a dedicated staining buffer (ebioscience) per the manufacturer's instruction. Subsequently, cells were acquired on an LSR II flow cytometer (BD Biosciences, San Jose, CA) using FACSDiva software. Data were analyzed using FlowJo software (Tree Star, Ashland, OR).
Indicated total CD4 or CD8 T cell populations were isolated using a negative cell purification kit (StemCell Technologies, BC, Canada), according to the manufacturer's instructions. Cell purity was >90% in all cases. Classical LAP− HLA-G− CD25hi CD4 T cells, HLA-G+ CD4 T cells, LAP+ CD4 T cells and a control cell population of LAP− HLA-G− CD25− CD4 T cells were sorted on a FACSAria instrument (BD Biosciences) at 70 pounds per square inch. For isolation of CD8 Treg subsets, purified bulk CD8 T cells were sorted into three T cell subsets: HLA-G+ CD8 T cells, CD25hi CD28− CD8 T cells and a control cell population of HLA-G− CD25− CD8 T cells, using similar sorting conditions.
PBMC from HIV-1 infected individuals were stained with 0.25 µM carboxyfluorescein succinimidyl ester (CFSE; Invitrogen) and mixed with sorted autologous Treg populations or control T cells without regulatory activity at a ratio of 4∶1. Afterwards, cells were stimulated with a pool of overlapping peptides spanning the clade B consensus sequence of HIV-1 gag, a pool of overlapping peptides spanning the entire sequence of human CMV pp65 (concentration of 2 µg/ml per peptide), or PHA. After incubation for 6 days, cells were washed, stained with viability dye and surface antibodies, fixed and acquired on an LSR II flow cytometer. Suppression of T cell proliferation by Tregs was calculated as: (T cell proliferation (%) in the non-Treg co-culture – T cell proliferation (%) in the Treg co-culture)/T cell proliferation (%) in the non-Treg co-culture.
CFSE-stained responder T cells from HIV-1-infected patients were mixed with sorted autologous Treg populations or control CD4 T cells at a ratio of 2∶1. Cells were then stimulated with a pool of overlapping peptides spanning HIV-1 gag (concentration of 2 µg/ml per peptide) in the presence of antibodies directed against CD28 and CD49d (2 µg/ml). Cells were incubated for 6 h at 37°C, and Brefeldin A was added at 5 µg/ml after the first hour of incubation. Afterwards, cells were stained with viability dye and surface antibodies, fixed, permeabilized using a commercial kit (Caltag, Burlingame, CA), and subjected to intracellular cytokine staining with monoclonal antibodies against interferon-γ and IL-2 (BD Biosciences). Following final washes, cells were acquired on an LSR II instrument.
Responder T cells from healthy individuals were mixed with sorted autologous Treg populations or autologous control T cells without regulatory activities at a ratio of 2∶1. Following stimulation of cells with Staphylococcal Enterotoxin B (SEB, 5 µg/ml, kindly provided by Dr. Eric J. Sundberg, University of Maryland), cells were incubated at 37°C for 4 days. Afterwards, cells were stained with antibodies against CD4, CD8, CD38, HLA-DR, CD69 and Vβ13.1 and viability dye before being subjected to flow cytometric acquisition on an LSR II instrument. The surface expression of activation markers in responder T cells was analyzed after gating on T cells. Treg-dependent suppression of bystander activation was calculated as: (CD38/HLA-DR/CD69-expressing T cells (%) in the non-Treg co-culture – CD38/HLA-DR/CD69-expressing T cells (%) in the Treg co-culture)/CD38/HLA-DR/CD69-expressing T cells (%) in the non-Treg co-culture.
HLA-G+ and HLA-G− CD3 T cells were isolated by immunomagnetic enrichment and cultured in IL-2 supplemented medium for 4 days. Equal amounts of culture supernatants and cell lysates were then subjected to SDS-PAGE (8 to 16% Tris-glycine gels, Invitrogen), electroblotted and incubated with HLA-G antibodies (clone 4H84, Abcam), followed by visualization with horseradish peroxidase (HRP)-labeled secondary antibodies and enhanced chemiluminescence (ECL) detection reactions (GE Healthcare, Little Chalfont, UK) according to standard protocols [13].
CD4 T cells were activated with recombinant IL-2 (50 U/ml) and an anti-CD3/CD8 bi-specific antibody (0.5 µg/ml). On day 5, cells were infected with GFP-encoding X4- (NL4-3, MOI = 0.02) or R5- (Ba-L, MOI = 0.07) tropic viral strains [14] (kindly provided by Dr. Dan Littman, New York University) for 4 h, or with a YFP-encoding VSV-G-pseudotyped HIV-1 vector (MOI = 0.02) (kindly provided by Dr. Abraham Brass, University of Massachusetts) for 2 h at 37°C. After two washes, cells were plated at 5×105 cells per well in a 24-well plate. On day 2 (VSV-G-pseudotyped virus) or day 4 (X4-/R5-tropic viruses), cells were stained with surface antibodies and viability dye and analyzed on an LSR II instrument. For infection of quiescent cells, negatively-selected CD4 T cells with a purity of >95% were directly infected with the described HIV-1 constructs. After in vitro culture for 96 h in the absence of exogenous IL-2, cells were analyzed by flow cytometry.
Data are expressed as mean and standard deviation/standard error, or as box and whisker plots indicating the median, the 25% and 75% percentile and the minimum and maximum of all data. Differences between different cohorts or different experimental conditions were tested for statistical significance using Mann-Whitney U test, paired T test or one-way ANOVA, followed by post-hoc analysis using Tukey's multiple comparison test, as appropriate. Spearman correlation was used to assess the association between two variables. A p-value of 0.05 was considered significant. The level of significance was labeled as: *:p<0.05; **:p<0.01; ***:p<0.001.
Investigations of T cells with regulatory properties in HIV-1 infection have so far been mostly limited to classical, CD25hi and/or FoxP3 expressing Treg. To analyze the role of alternative, non-classical Treg populations in patients infected with HIV-1, we initially focused on the recently described population of Treg defined by surface expression of HLA-G [9]. These cells do not express FoxP3 or CD25 (Figure S1), and are phenotypically and functionally distinct from classical Treg [9], [10]. To analyze these cells in HIV-1 infection, we used flow cytometry to determine the relative and absolute numbers of HLA-G+ CD4 and CD8 T cells in treatment-naïve HIV-1 infected individuals with chronic progressive infection (n = 28, median viral load: 48,215 copies/ml [IQR 20,187–685,000]; median CD4 cell count: 396/µl [IQR 204–652]), spontaneous control of HIV-1 replication (n = 24, viral load <1000 copies/ml; median CD4 cell count: 924/µl [IQR 347–1879]), or patients with primary HIV-1 infection and seroconversion within 3 months prior to recruitment (n = 22, median viral load: 99,900 copies/ml [IQR 36,600–2,790,000]; median CD4 cell count: 475/µl [IQR 265–1047]). HIV-1 infected persons successfully treated with Highly Active Antiretroviral Therapy (HAART) (n = 26, viral load <50 copies/ml; median CD4 cell count: 402/µl [IQR 242–1493]), as well as a cohort of HIV-1 negative persons (n = 21), were recruited for control purposes.
Consistent with prior reports [15], we observed that relative proportions of classical CD25hi CD127lo CD4 Treg were increased in progressive HIV-1 infection, while absolute Treg numbers were decreased (Figure S2); no correlation was found between relative proportions of classical Treg and levels of immune activation (Figure S2). In contrast, we observed that the relative and absolute numbers of HLA-G-expressing CD4 T cells were lowest in HIV-1 progressors, while no significant difference was found between the numbers of HLA-G+ CD4 T cells in any of the other HIV-1 patient cohorts and HIV-1 negative persons (Figure 1 A/B). The relative frequencies of HLA-G+ CD8 T cells were lower in all HIV-1 infected patient populations compared to HIV-1 negative persons; this reduction was again most pronounced in persons with untreated progressive disease. Notably, the numbers of HLA-G-expressing CD4 and CD8 T cells were positively correlated to total CD4 T cell counts (Figure 1C), and proportions of HLA-G+ T cells were inversely associated with corresponding levels of immune activation on T cells, as determined by surface expression of HLA-DR and CD38 (Figure 1D). These data indicate a selective numerical decrease of HLA-G-expressing T cells in chronic progressive HIV-1 infection, and suggest that a reduction of HLA-G+ Treg may contribute to higher levels of immune activation during progressive HIV-1 infection.
Since HLA-G+ Treg express multiple tissue homing factors [16], a redistribution of these cells to lymphoid tissues may be responsible for the apparent reduction of HLA-G-expressing Treg in the peripheral blood during progressive HIV-1 infection. To investigate this, we analyzed the proportion of HLA-G+ T cells in lymph node and peripheral blood samples collected from patients treated with antiretroviral therapy (HIV-1 viral load<75 copies/ml, median CD4 count: 762/µl [IQR 528–1,152]) or with untreated progressive HIV-1 infection (median HIV-1 viral load: 73,500 copies/ml [IQR 1,300–252,000], median CD4 count: 430/µl [IQR 254–1,267]). Within these patients, proportions of HLA-G+ CD4 and CD8 Treg in lymph nodes and peripheral blood were not significantly different, suggesting that compartmentalization of HLA-G+ Treg to lymph nodes does not represent the major reason explaining the decreased number of circulating HLA-G+ Treg in progressive HIV-1 infection (Figure 2A). In contrast, classical CD25hi CD127lo Treg were significantly enriched in lymph nodes compared to peripheral blood in patients on and off HAART, consistent with previous results [17] (Figure 2B).
We next investigated whether the reduced frequencies of circulating HLA-G+ Treg during progressive HIV-1 infection are associated with an altered phenotypic differentiation or maturation status. We found that in all study cohorts, the T cell subset distribution of HLA-G+ CD4 T cells into naïve, central-memory, effector-memory and terminally-differentiated CD4 T cells was not substantially different from corresponding bulk CD4 T cells (Figure S3). Moreover, the expression of CD57 and PD-1, two surface markers associated with senescence and exhaustion of T cells, was not markedly different between HLA-G+ CD4 T cells and the respective bulk CD4 T cells (Figure S4). In contrast, we noted that in all study cohorts, HLA-G+ CD8 T cells tended to have a more immature naïve or central-memory phenotype when compared to reference bulk CD8 cell populations (Figure S3). There was also a trend for reduced surface expression of CD57 surface expression on HLA-G+ CD8 T cells in comparison to corresponding bulk CD8 T cells (Figure S4). Overall, these data indicate that during HIV-1 infection, HLA-G-expressing CD8, but not CD4 T cells, are skewed to a more immature differentiation status, but this difference is not correlated to the rates of spontaneous HIV-1 disease progression.
T cells expressing LAP, a membrane-bound form of TGF-β, have recently been characterized as an alternative, FoxP3-negative population of lymphocytes with immunosuppressive properties [12] [18]. To determine whether this non-classical population of regulatory cells is involved in HIV-1 disease pathogenesis, we analyzed the frequency of LAP+ T cells in our study cohorts. We did not observe significant differences in the proportions of LAP+ T cells between our study groups (Figure S5). Absolute numbers of LAP+ CD4 Treg were positively associated with total CD4 T cell counts (Figure S5), and were lowest in progressors, likely reflecting the decline of total CD4 T cells in this patient population (Figure S5). Proportions of neither LAP+ CD4 nor LAP+ CD8 T cells were significantly associated with corresponding levels of immune activation (Figure S5). LAP+ T cells did not substantially differ from bulk T cells in terms of T cell subset distribution, although LAP+ CD8 T cells appeared to be slightly overrepresented in central-memory cells during HIV-1 infection (Figure S6). No difference was found between the surface expression of PD-1 and CD57 on LAP+ T cells and bulk T cells (Figure S4). Taken together, these results do not suggest that LAP+ T cells play a major role in HIV-1 immune protection or restriction of HIV-1 associated immune activation.
A functional hallmark of classical Treg is their ability to inhibit antigen-specific T cell responses [19]. Prior work has shown that non-classical Tregs can also inhibit proliferative properties of T cells, but their functional effects on HIV-1-specific T cells remain unclear [9]. To investigate this, CFSE-labeled PBMC from HIV-1 controllers were stimulated with viral peptides or PHA and individually mixed with sorted autologous HLA-G+ CD4 Treg, HLA-G+ CD8 Treg or classical CD25hi CD4 Treg; HLA-G− CD25− CD4 or CD8 T cells were added as negative controls. Subsequently, proliferation of HIV- and CMV-specific T cells was monitored after six days of culture. These experiments demonstrated suppressive effects of classical CD25hi Treg on the proliferative activities of HIV-1- and CMV-specific CD4 and CD8 T cells, consistent with prior reports showing potent Treg-mediated inhibition of T cell proliferation [5]. In contrast, HLA-G+ Treg did not effectively suppress the proliferative activity of autologous virus-specific CD4 (Figure 3A/C) or CD8 (Figure 3B/D) T cells in these study patients. LAP-expressing CD4 Treg had a moderate suppressive effect on proliferative activities of HIV-1-specific T cells (Figure S7). None of the tested classical or non-classical Treg populations had a measurable impact on interferon-γ or IL-2 secretion in HIV-1-specific CD8 T cells (Figure S8). Taken together, these data show that HLA-G+ Treg have minimal effects on the functional activities of virus-specific T cell responses in controllers.
To further explore the role of non-classical Tregs in HIV-1 disease pathogenesis, we focused on how these cells influence T cell activation. Activation of T lymphocytes can either occur through direct antigenic triggering of the TCR, or by mechanisms involving a TCR-independent mode of T cell stimulation, commonly referred to as “bystander activation” [20], [21]. Both of these pathways seem to contribute to the pathological immune activation observed during progressive HIV-1 infection [22], [23], and may be influenced by the non-classical Treg populations described in this manuscript. As a functional assay to investigate and quantify the effects of non-classical Tregs on TCR-dependent and bystander immune activation, we stimulated T cells with Staphylococcal Enterotoxin B (SEB), an antigen that elicits T cell responses by a broad panel of different TCR clonotypes, but cannot be recognized by T cells using TCR Vβ13.1 [24], [25]. Immune activation in Vβ13.1-expressing T cells following exposure to SEB can therefore only be attributed to bystander activation, while immune activation in Vβ13.1-negative T cells after SEB exposure reflects classical TCR-dependent activation.
To analyze the effects of non-classical Tregs on immune activation, SEB-stimulated responder T cells were individually co-cultured with autologous populations of sorted LAP+ CD4 Treg, HLA-G+ CD4 Treg, or classical LAP− HLA-G− CD25hi CD4 Treg; LAP− HLA-G− CD25− CD4 T cells were added for control purposes. Alternatively, HLA-G+ CD8 T cells, CD25hi CD28− CD8 T cells or HLA-G− CD25− CD8 control cells were added to autologous SEB-stimulated responder T cells. On day 4 of culture, immune activation was measured by flow cytometric analysis of CD38, HLA-DR and CD69 surface expression in Vβ13.1-expressing and Vβ13.1-negative T cells. As demonstrated in Figure 4, we observed that classical CD25hi Treg potently suppressed CD38/HLA-DR expression in Vβ13.1-negative T cells, consistent with prior reports about the immunosuppressive properties these cells [5]. In contrast, HLA-G-expressing Treg led to a significantly reduced surface expression of CD38 on Vβ13.1-expressing T cells, but had limited effects on immune activation of Vβ13.1-negative cells. This selective inhibitory effect on bystander activation was seen both for HLA-G+ CD4 (Figure 4A/C) and CD8 (Figure 4B/D) T cells and substantially exceeded regulatory effects on bystander activation of classical CD25hi Treg or LAP+ Treg. None of the tested Treg populations significantly affected CD69 expression on responder cells over the 4-day incubation period, likely because in comparison to CD38, CD69 is only transiently upregulated for a short period after immune activation [26], and therefore could not be properly evaluated in our 4-day co-culture experiment.
To explore reasons for the differential susceptibility of Vβ13.1-positive and Vβ13.1-negative responder T cells to classical and non-classical Tregs, we analyzed the dynamics of LILRB1 surface expression on responder T cells over a 4-day incubation period. LILRB1 can effectively inhibit functional properties of T cells [27] and represents one of the highest-affinity receptors for HLA-G [28], which is secreted by HLA-G+ Treg (Figure S9) and responsible for the immunomodulatory effects of HLA-G+ Treg [9], [10]. Interestingly, we observed that following TCR-dependent T cell activation, LILRB1 surface expression on responder T cells declined, while stable or slightly increased LILRB1 surface expression was observed on Vβ13.1-negative T cell after “bystander activation” (Figure 5). Overall, these data indicate that HLA-G+ Treg differ from alternative Treg populations by their ability to reduce bystander activation of T cells, and suggest that TCR-dependent and TCR-independent mechanisms of immune activation are associated with altered susceptibilities to inhibitory effects of classical and non-classical Tregs.
Conventional CD25hi CD4 Treg express HIV-1 co-receptors and are targets for HIV-1 infection [29], [30]. Direct HIV-1 infection of HLA-G+ CD4 Treg may contribute to the reduction of these cells in progressive HIV-1 infection. To investigate this, we analyzed the susceptibility of HLA-G+ CD4 Treg to X4- or R5-tropic HIV-1 viruses, or to a VSV-G-pseudotyped HIV-1 construct causing single-round HIV-1 infection. We observed that HLA-G+ CD4 Treg were significantly more susceptible to HIV-1 infection than autologous HLA-G− CD4 T cells; this was true both for in vitro activated cells and for cells directly infected ex-vivo (Figure 6A–C, E). This enhanced susceptibility was in line with higher expression of the HIV-1 co-receptors, CXCR4 and CCR5, on HLA-G+ CD4 Treg, in comparison to HLA-G− CD4 T cells (Figure 6D/F). These data suggest that reduction of circulating HLA-G+ CD4 Treg in progressive HIV-1 infection may, at least in part, be due to their enhanced susceptibility to HIV-1 infection.
Regulatory T lymphocytes can influence immune homeostasis by suppressing innate and adaptive effector cell activity, and in this way may importantly modulate immune defense mechanisms against HIV-1 [31]. The majority of currently available data indicate that classical CD25hi CD127lo Treg are expanded during chronic progressive HIV-1 infection [32], [33], [34], [35], [36], [37] and may worsen spontaneous HIV-1 disease progression by potently suppressing functional activities of HIV-1-specific T cell responses [5], [17], [38]. Here, we demonstrate several numerical and functional aspects of non-classical HLA-G-expressing Treg in HIV-1 infection that clearly distinguish them from these recognized characteristics of classical Treg. We found that absolute numbers and relative proportions of HLA-G-expressing Treg are diminished in progressive HIV-1 infection, that they are inversely correlated to phenotypic markers of immune activation, and that they may have a functional role for reducing bystander immune activation, while only minimally suppressing proliferative activities of HIV-1-specific T cells. In contrast, an alternative population of non-classical Treg expressing the TGF-β latency-associated antigen (LAP) was not correlated to immune activation during HIV-1 infection and weakly affected immune activation in functional assays. Overall, these data suggest that HLA-G-expressing Treg may contribute to balancing and fine-tuning anti-viral immune activity and bystander immune activation during HIV-1 infection.
HLA-G+ Treg represent a relatively recently discovered group of suppressive T cells that can inhibit the activation and proliferation of T cells after TCR triggering with CD3/CD28 antibodies. However, how HLA-G+ Treg functionally compare to classical Treg in terms of their ability to suppress virus-specific T cells or TCR-independent bystander activation of lymphocytes remained unclear. Our data show that HLA-G+ Treg do not effectively inhibit proliferation of HIV-1- and CMV-specific T cells, compared to the effects of classical Treg in HIV controllers. In contrast, we observed a seemingly stronger ability of HLA-G+ Treg to reduce TCR-independent bystander activation of T cells, using an assay that excludes TCR cross-reactivity as a possible source of activation in heterologous T cells. Yet, due to the numeric reduction of HLA-G+ Treg in progressive HIV-1 infection, all functional effects of these cells could not be evaluated using cells from this particular patient population. Whether functional properties of HLA-G+ Treg from HIV-1 progressors or HAART-treated patients resemble those of HIV-1 negative persons, or exhibit an altered or dysfunctional profile, remains to be investigated. Nevertheless, our results suggest that HLA-G+ Treg differ from alternative Treg populations by a unique profile of suppressive functions that may allow for reducing bystander immune activation while simultaneously minimizing inhibitory effects on virus-specific T cell immune responses. The preservation of this HLA-G-expressing Treg population in HIV-1 controllers may represent an additional immunological feature of this specific patient population.
This work demonstrates that in contrast to classical Treg, HLA-G-expressing Treg progressively decline during advanced HIV-1 infection. This selective loss of HLA-G+ Treg during advanced HIV-1 infection may, in conjunction with other mechanisms, contribute to immune overactivation during progressive HIV-1 infection. The reduction of HLA-G+ CD4 Treg during progressive HIV-1 infection may be related to their increased susceptibility to HIV-1 infection, which is likely due to enhanced expression of the viral co-receptors CCR5 and CXCR4 demonstrated in this study. An upregulation of these chemokine receptors may also lead to elevated sequestration of HLA-G+ Treg into inflamed tissues, where these cells were indeed preferentially observed in previous investigations [9], [39]. However, in our study, we did not find any positive evidence for a selective enrichment of HLA-G+ CD4 and CD8 Treg in lymphoid tissues, either in HAART-treated or in untreated HIV-1 patients; but this observation in a limited number of patients does not exclude the possibility of tissue compartmentalization of HLA-G+ Treg in HIV-1 infection. In addition, the specific reason for the loss of HLA-G+ CD8 Treg in untreated progressive HIV-1 infection remains unclear and warrants further investigation.
Over the recent years, HIV-1 infection has increasingly been recognized as a chronic inflammatory condition characterized by elevated T cell immune activation [40]. The mechanisms leading to this abnormal immune activation are most likely multifactorial and include direct stimulation of T cells by HIV-1 antigens, as well as direct TCR-mediated activation of T cells by alternative viral and bacterial antigens that challenge the host during conditions of HIV-1 associated immune deficiency. TCR-independent bystander immune activation does not seem to play a significant role under physiologic conditions, however, increasing data suggest that bystander activation represents a major driving factor for pathological immune activation during progressive HIV-1 infection. For instance, bystander activation occurs mainly through cytokines, including interferon-α/β, IL-2 and IL-15 [41], which are all increased in HIV-1 infection and represent independent and accurate predictors of disease progression [42]. Moreover, the majority of activated T cells in HIV-1 infected patients typically do not exhibit phenotypic markers of recent TCR stimulation [43], suggesting that their activation occurred by TCR-independent processes. In addition, activation of T cells specific for Influenza virus has been documented during HIV-1 infection in the absence of serological evidence of Influenza co-infection, or detectable TCR cross-reactivity between HIV-1 and Influenza antigens [44]. Interestingly, our data suggest that T cells activated by bystander mechanisms may have a higher susceptibility to inhibitory effects of HLA-G+ Treg, likely because they do not downregulate the HLA-G receptor LILRB1 in a similar way as T cells activated by TCR triggering. These observations indicate that TCR-dependent and TCR-independent mechanisms of immune activation are associated with altered susceptibilities to classical and non-classical Tregs, and shed new light on target cell characteristics that influence inhibitory effects of Tregs. By selectively reducing the deleterious effects of TCR-independent bystander activation, HLA-G+ Treg may provide a previously unrecognized form of immune protection against HIV-1 associated disease manifestations.
|
10.1371/journal.ppat.1007252 | Control of primary mouse cytomegalovirus infection in lung nodular inflammatory foci by cooperation of interferon-gamma expressing CD4 and CD8 T cells | Human cytomegalovirus (CMV) and mouse cytomegalovirus (MCMV) infection share many characteristics. Therefore infection of mice with MCMV is an important tool to understand immune responses and to design vaccines and therapies for patients at the risk of severe CMV disease. In this study, we investigated the immune response in the lungs following acute infection with MCMV. We used multi-color fluorescence microscopy to visualize single infected and immune cells in nodular inflammatory foci (NIFs) that formed around infected cells in the lungs. These NIFs consisted mainly of myeloid cells, T cells, and some NK cells. We found that the formation of NIFs was essential to reduce the number of infected cells in the lung tissue, showing that NIFs were sites of infection as well as sites of immune response. Comparing mice deficient for several leukocyte subsets, we identified T cells to be of prime importance for restricting MCMV infection in the lung. Moreover, T cells had to be present in NIFs in high numbers, and CD4 as well as CD8 T cells supported each other to efficiently control virus spread. Additionally, we investigated the effects of perforin and interferon-gamma (IFNγ) on the virus infection and found important roles for both mechanisms. NK cells and T cells were the major source for IFNγ in the lung and in in vitro assays we found that IFNγ had the potential to reduce plaque growth on primary lung stromal cells. Notably, the T cell-mediated control was shown to be perforin-independent but IFNγ-dependent. In total, this study systematically identifies crucial antiviral factors present in lung NIFs for early containment of a local MCMV infection at the single cell level.
| Cytomegalovirus (CMV) is worldwide a highly prevalent β-herpesvirus. While the primary infection in healthy individuals does not cause disease, infection of immunocompromised patients can lead to multiple organ disease and can sometimes be lethal. CMV becomes latent and uses a series of mechanisms to circumvent its elimination by the immune system. Therefore, the establishment of therapies or vaccination strategies is a difficult endeavor. Murine CMV (MCMV) is a well-established model to study CMV infection in mice, since it targets many immune defense mechanisms in a similar manner as human CMV. Thus, studying immune responses of mice to MCMV can help to develop new therapeutic strategies for patients. In this study we focused on MCMV infection in adult mouse lungs and found that, in the immunocompetent host, immune cells infiltrate the lung tissue as nodular inflammatory foci that help to control the acute infection in one week. We identified specific lymphocyte subsets that are pivotal for efficient containment of the infection and showed that this process requires cooperation between CD4 and CD8 T cells. Furthermore, these cells need to secrete the multipotent cytokine IFNγ to successfully clear the lungs from infectious viruses.
| The immune response against CMV infection in humans and murine cytomegalovirus (MCMV) in mice has been studied for decades [1]. Understanding immune-cell mediated control of CMV infection is of essential interest since immunocompromised patients are particularly susceptible to CMV-disease [2]. From earlier human and mouse studies, it is known that virus-specific T cell responses contribute to virus control in both species [3,4]. It is thought that both CD8 T cell-mediated killing of virus-infected cells and secretion of cytokines by CD8 and CD4 T cells contribute to protective immunity [4,5]. Since different lines of evidence indicate that T cells can control CMV infection, the adoptive transfer of HCMV-specific T cells has been used in clinical trials [6]. In patients that suffer from CMV reactivation or primary infection during periods of severe immunosuppression, the infusion of donor HCMV-specific T cells was found to improve the control of infection [7,8]. However, it remains unclear how CD8 T cells are able to control CMV infection, given that both HCMV and MCMV encode for proteins that efficiently down-modulate surface major histocompatibility complex class I (MHCI) on infected cells [9]. This immune evasion mechanism can prevent virus-specific T cells from recognizing their target cells and thus interferes with CD8 T cell-mediated killing [10]. Other immunoevasins interfere with the presentation of antigen via major histocompatibility complex class II (MHCII) molecules to avoid recognition by CD4 T cells [11,12] but they also downregulate ligands of natural killer (NK) cell receptors to avoid cytotoxic killing [13].
The immune response against CMV infection seems to be affected by the site of infection, since various cell populations and effector mechanisms differentially contribute to virus control in different organs [14–16]: While IFNγ secreting NK cells have a more pronounced role than CD8 T cells in controlling the infection in the liver [14,15], an early, perforin-dependent NK cell mediated, and a late CD8 T cell based mechanism control virus load in the spleen [14]. In salivary glands, IFNγ-secreting CD4 T cells are the major cell population crucial for viral reduction [16,17]. Although CMV pneumonitis is a common clinical problem, it is not entirely understood how the infection is controlled in lungs [18,19]. Moreover, it is still unclear which immune cells are locally present at the site of infection and how many cells of each type are needed to successfully control the infection in the affected tissue.
To simultaneously quantify the amount of infected cells and the ensuing cellular immune response at the primary site of infection, we have previously generated MCMV reporter mutants expressing the red fluorescent protein mCherry to directly observe single virus-infected cells by microscopy [20,21]. Following infection with these reporter viruses, histological sections of the infected organ reveal the number of virus-infected cells and simultaneously allow tracking of the immune response at the site of infection. Further, the MCMV reporter virus used in this study is deficient for m157. In C57BL/6 mice–but not in other mice, including wild mice–specific binding of the m157 protein to the NK-cell expressed Ly49H receptor leads to NK cell activation and fast elimination of infected cells [22]. Besides Ly49H, other specific NK cell receptors have been identified to be involved in the immune response during MCMV infection [23]. However, mice that do not express the activating NK cell receptor Ly49H are more susceptible to MCMV infection [24,25]. To investigate immunity against MCMV infection in C57BL/6 mice without an excessive NK cell response we here applied a m157-deficient reporter virus. Following infection of lungs in neonatal mice, we previously showed that MCMV infection can be precisely characterized, revealing the formation of dynamic nodular inflammatory foci (NIF) with accumulation of virus-infected and immune cells [26,27]. This inflammatory response is also found in humans after CMV infection and has been termed nodular inflammation due to the abundancy of myeloid cells and corresponding granuloma-like appearance [28]. Moreover, we observed how MCMV-specific CD8 T cells interact with infected cells to mediate antiviral immunity in vivo [10].
In the current study, we identified the types of immune cells that control acute MCMV infection in lung NIFs of adult mice. To systematically address the impact of different immune cells, we used several knock-out mice, as well as adoptive cell transfers, and cell depletion approaches. To clarify how leukocytes mediate antiviral immunity, we studied the main lymphocyte populations important for virus control in the lungs and the underlying mechanisms. For the lungs, we found that local inflammation with recruitment of immune cells was necessary to control infection. The formation of NIFs occurred independently of the presence of lymphoid cells. We confirmed an important role of T cell-secreted IFNγ, while perforin secretion by T cells was not essential for controlling MCMV infection in lung NIFs. However, in the absence of T cells we observed a NK cell-mediated antiviral effect that was independent from m157 recognition and conveyed by IFNγ and perforin. Furthermore, in adoptive cell transfer experiments we found that only the combination of CD8 with CD4 T cells was sufficient to control MCMV infection whereas either cell population on its own had only minor antiviral effects.
To define the composition of immune cells in lung NIFs of adult immunocompetent mice, we intranasally infected 6–14 weeks old C57BL/6 mice with 106 PFU of recombinant MCMV-3D [20]. At 5 days post infection (dpi), we found accumulations of CD45+ immune cells at multiple sites of the lungs (Fig 1A). In contrast, no inflammation could be observed in uninfected control animals (Fig 1A). The cellular infiltrates in lungs of MCMV-infected adult mice at 5 dpi localized in close proximity to MCMV-infected cells and consisted mainly of myeloid cells positive for CD11b and/or CD11c, as well as CD3+ T cells and NK1.1+ NK cells (Fig 1B). These findings demonstrate that adult mice develop NIFs around MCMV-infected cells in the lungs in a similar manner as previously described for infected neonatal mice [26]. Next, we analyzed the number of infected cells and NIF dynamics in the first two weeks following infection. We found an increasing number of MCMV-infected cells in NIFs within the first 3 dpi, whereas at 8 dpi only few mCherry+ cells remained (Fig 1C and 1D). In parallel, the average size of NIFs increased within the first 5 days and contracted until 8 dpi (Fig 1E). At day 8, NIFs were dissolving and only some mCherry+ remnants of infected cells were detected. Thus, in immunocompetent mice acute MCMV infection in the lungs was controlled during the first week of infection, most likely by cells that infiltrated the lung tissue and resided within NIFs at the site of infection.
To investigate whether hematopoietic cells present within NIFs mediate antiviral activity, we compared MCMV infection in gamma-irradiated versus non-irradiated mice. Following gamma-irradiation, mice showed reduced numbers of CD45+ cells in the peripheral blood, confirming interference with hematopoiesis. Five and eight days after MCMV infection of irradiated mice, we found no structures in the lungs that could be classified as NIFs (Fig 2A). Instead, we observed numerous clusters of morphologically intact MCMV-infected cells at both time points (Fig 2A). Only few apparently irradiation-resistant hematopoietic cells were scattered through the lung tissue. Because of the absence of NIFs in irradiated mice, we quantified by histology the number of infected cells per lung slice and measured in a separate experiment virus titers and Gaussia luciferase activities (Fig 2B–2E). At 5 dpi, we found the number of intact infected cells, Gaussia luciferase activity, and virus titers to be comparable in irradiated and control animals. In contrast, at 8 dpi, control mice showed a reduction in infected cell numbers, luciferase activities, and virus titers, whereas in irradiated mice virus loads in the lungs had slightly increased compared to 5 dpi (Fig 2B–2D). In general, we found a correlation between viral titers and luciferase activity in lungs as well as salivary glands (Fig 2E, S1C Fig). Analysis of salivary glands showed a comparable infection level in irradiated and control mice at 5 dpi (S1A and S1B Fig). However, virus loads were increased in both conditions at 8 dpi indicating that the virus spreads slowly towards salivary glands (S1A and S1B Fig), confirming previous reports on infection dynamics in salivary glands [26,29]. Furthermore, virus loads in irradiated mice were even higher at 8 dpi than control mice, suggesting a role for immune cells to reduce virus dissemination from the lungs into salivary glands.
Taken together, the disappearance of MCMV-infected cells in lungs of immunocompetent mice was most likely due to antiviral effects executed by NIF-resident immune cells.
Next, we asked whether NIF formation occurs in the absence of defined subpopulations of immune cells and whether virus spread can be controlled under such conditions. To screen for the contribution of different types of lymphoid cells to NIF formation and antiviral control, we used mice deficient for B cells (Ighm-/-), T cells (Cd3e-/-), B and T cells (Rag2-/-), and mice lacking the entire lymphoid compartment (Rag2-/-Il2rg-/-). Following intranasal infection, we found typical NIFs in all these immunodeficient mouse strains (Fig 3A). At 8 dpi, NIFs in Ighm-/- mice were already dissolving and negative for intact mCherry+ cells, thus being comparable to observations made in immunocompetent mice (Fig 3A and 3B). In contrast, in T cell-deficient Cd3e-/- mice as well as T and B cell-deficient Rag2-/- mice NIFs were still present and contained intact mCherry+ cells (Fig 3A and 3B and S2 Fig). These animals showed a small but consistently higher number of infected cells per NIF than observed in wild-type mice (Fig 3B), while a robust elevation of the total number of infected cells could be observed (Fig 3C). Mice deficient for T, B, and NK cells (Rag2-/-Il2rg-/- mice) or Rag2-/- mice depleted for NK cells showed the highest average number of virus-infected cells per NIF, suggesting that the different types of lymphocytes cooperate to limit virus replication inside NIFs (Fig 3B). Notably, we found no difference in the average number of infected cells per lung slice between Rag2-/-Il2rg-/- and Rag2-/- mice, while NK cell-depleted Rag2-/- mice showed elevated numbers of infected cells (Fig 3C). To further address the role of NK cells in the presence of T and B cells we next depleted NK cells in WT mice. For these experiments we chose to analyze the lungs at 5 dpi, since the NK cell-mediated response is faster than T cells and at 8 dpi the infection is already controlled in WT mice. In NK cell-depleted mice we could detect slightly increased numbers of infected cells per NIF but not per lung slice (S3 Fig). Together, these data suggest that NIF formation occurs independently of the presence of lymphocytes and control of viral infection in NIFs primarily relied on T cells. Further, independent of m157, NK cells can reduce the infection in the absence of T cells and may have the potential to counteract the infection in the presence of T cells.
The MCMV-encoded chemokine 2 (MCK2) was reported to increase virulence of the virus [27,30–32] and may interfere with T cell responses [33,34]. To address whether the presence of MCK2 affects T cell-mediated control of lung MCMV infection we infected WT and Rag2-/- mice with MCMV-3DR. In contrast to the MCMV-3D recombinant the insertion of a missing base pair in the Mck2 ORF of the MCMV-3DR recombinant allows for accurate translation of the whole MCK2 protein [30]. Similarly to the MCMV-3D virus, the number of infected cells per NIF in Rag2-/- mice was increased compared to WT mice infected with MCMV-3DR at 8 dpi (S4A Fig). However, MCMV-3DR infected alveolar macrophages leading to a higher total number of infected cells per lung as compared to animals infected with MCMV-3D –which shows no tropism to alveolar macrophages (S4B Fig) [27]. Thus, the T cell-mediated control of MCMV infection in NIFs seems to be independent of the presence of the full MCK2 protein.
We next studied the dynamics of the number of lymphocytes present in NIFs of MCMV-infected wild-type mice. Histological analysis at 1 dpi revealed hardly any lymphocytes residing in close proximity to infected cells, whereas at 3 dpi and 5 dpi a substantial number of T cells is present in NIFs. The number of T cells per NIF is even further increased at 8 dpi, when most of the infected cells already disappeared. In contrast, the number of NK cells present in NIFs peaked at 3 dpi and declined until 5 dpi (Fig 4A and 4B and S5 Fig). These data indicate that NK cells populate NIFs early after infection, while T cells might mediate their antiviral effects between 3 and 8 dpi.
To determine whether the T cell-mediated antiviral response is dose-dependent, we varied the number of T cells available in the lungs of infected mice. To this end, we adoptively transferred increasing numbers of purified polyclonal T cells isolated from untreated wild-type mice into MCMV-infected Rag2-/- recipients (Fig 4C). At day 8 post infection and transfer of 106 or less T cells, many NIFs still contained virus-infected cells (Fig 4D). In contrast, the transfer of 107 T cells resulted in a reduced number of infected cells per NIF (Fig 4D). We observed a similar effect when analyzing the whole lung slice for the average number of infected cells (Fig 4E). Simultaneously, we analyzed the number of T cells present in NIFs. Only few T cells were detected in mice receiving 104 or 105 T cells and a robust increase in the number of T cells was observed after transfer of 106 T cells (Fig 4F). However, following the adoptive transfer of 107 T cells into Rag2-/- recipients, we found T cell densities comparable to those in NIFs of wild-type mice (approximately 100 T cells per NIF; Fig 4G). Interestingly, FACS analysis of infected lungs revealed that a substantial amount of CD8 T cells is specific for MCMV (S6 Fig). These data indicate that a certain T cell density is needed to sufficiently contain the MCMV infection. In summary, we found that the number of T cells per NIF increased over time in wild-type mice and that adoptive transfer of 107 naïve T cells into Rag2-/- mice was necessary to accumulate enough T cells in NIFs to control the infection.
To screen whether perforin and/or IFNγ might be essential for antiviral immunity in MCMV-infected lungs, we analyzed mice deficient for perforin (Prf1-/-), IFNγ(Ifng-/-), or IFNγ-receptor-1 (Ifngr1-/-). At 8 dpi, we quantified the number of infected cells per NIF but did not find any significant differences between WT and the three knock-out strains tested (Fig 5A). Additionally, we analyzed the number of intact infected cells per whole lung slice and found an increase of intact infected cells in Prf1-/-, Ifng-/-, and Ifngr1-/- mice (Fig 5B) as well as an increase in the number of NIFs per lung section (Fig 5C). Compared to mice lacking T and B cells (Fig 3C), the differences of infected cell numbers observed here were less pronounced. Thus, these data indicate that antiviral effector mechanisms in NIFs of perforin- or IFNγ deficient mice may substitute for each other.
Next, we investigated the impact of IFNγ and perforin in the absence of T and B cells. To this end, we analyzed Rag2-/-Prf1-/- and Rag2-/-Ifng-/- mice 8 dpi. We found a slightly increased number of infected cells per NIF in Rag2-/-Prf1-/- mice when compared to Prf1-/- or Rag2-/- mice (Fig 5A and 5D), while the number of infected cells per lung slice significantly increased in these mice (Fig 5B–5E). Furthermore, in Rag2-/-Ifng-/- mice, an even higher number of intact infected cells was observed per lung slice, when compared to Rag2-/- or Ifng-/- single knock-out mice (Fig 5B and 5E). These findings support the idea that perforin as well as IFNγ can both play a role in controlling acute MCMV infection in the lungs. In absence of T cells, the secretion of these two molecules, most probably by NK cells, gains more importance in controlling the virus. However, it is likely that in Rag2-/- and Rag2-/-Il2rg-/- mice IFNγ is also produced by cells that do not express classical NK cell biomarkers [35].
To further characterize the cells producing IFNγ, we compared cytokine production of leukocytes isolated from MCMV- or mock-infected lungs following in vitro re-stimulation (Fig 5F). In total we observed more IFNγ-producing cells in infected animals than in control mice (Fig 5G). At 5 dpi, NK cells were found to be the dominant cell population to produce IFNγ followed by CD8 and CD4 T cells (Fig 5H). Interestingly, by histology less NK cells were found to be present in NIFs at 5 dpi, while flow cytometric analysis showed NK cells to be present in higher numbers in the lung (Fig 5H and S5 Fig). This indicates that most of the NK cells are present in the lung tissue but do not remain or proliferate in NIFs as efficiently as T cells. Taken together, NK and T cells were the major IFNγ-expressing cell types in the lungs during the peak of anti-MCMV immune response in an immunocompetent host leading to control of infection.
To test how IFNγ could change the dynamics of MCMV spread in NIFs, we next modeled the situation in the lungs applying an in vitro plaque reduction assay using primary lung stromal cells. First, we purified adherent cells from the lungs and found approximately 40% of these cells to be gp38+PDGFRA+PDGFRB+CD31- stromal cells (Fig 6A). The remaining cells were CD45+ hematopoietic cells, mainly of the myeloid lineage (CD11b+, partly CD11c+ and F4/80+; Fig 6A). Next, we infected this cell mixture in vitro with MCMV and followed the formation of plaques after addition of a single dose of recombinant IFNγ. At 4 and 8 dpi, we found that administration of IFNγ reduced plaque size in a dose-dependent manner (Fig 6B). Since about 60% of the cells in the culture were myeloid cells, the slower plaque growth might also be due to the activating effect of IFNγ on these myeloid cells. To investigate a direct effect of IFNγ on virus replication, we purified stromal cells by depletion of CD45+ cells. The resulting cell suspension lacked hematopoietic cells (Fig 6C). In this setup, we found IFNγ to have minor effects on plaque size at 4 dpi (Fig 6D). However, after 8 days, smaller plaques were observed in IFNγ-treated cell cultures in comparison to control cultures suggesting a direct interference of IFNγ with virus replication (Fig 6D). As control, we used stromal cells isolated from Ifngr1-/- mice and found no effect for any of the IFNγ concentrations tested (S7 Fig). Together, these data indicate that IFNγ can slow down cell-to-cell spread in lung stromal cells. The difference observed in stromal cell cultures in the presence or absence of hematopoietic cells might be caused by indirect mechanisms mediated by IFNγ treated myeloid cells [36].
It is generally believed that IFNγ and perforin can inhibit local viral spread in the lungs. We also showed that NK and T cells were the major sources of IFNγ in MCMV-infected lungs. Therefore, we next asked whether T cells deficient for IFNγ or perforin could still control MCMV infection. Following adoptive transfer of naïve T cells isolated from wild-type, Ifng-/-, Ifngr1-/-, or Pfr1-/- donor mice into Rag2-/- recipients, we quantified the number of infected cells per lung (Fig 7A). At 8 dpi, the transfer of wild-type T cells reduced the number of infected cells per lung slice (Fig 7B) as observed before (Fig 4E). Importantly, following the transfer of 2x107 IFNγ-deficient T cells, the control of MCMV infection was limited with many intact infected cells still detectable at 8 dpi (Fig 7B). In contrast, T cells deficient for the IFNγ-receptor showed a comparable antiviral effect to wild-type T cells, indicating that IFNγ signaling on T cells is not essential for T cells to suppress virus replication. Surprisingly, perforin-deficient T cells also reduced the number of infected cells (Fig 7B). Thus T cells need to secrete IFNγ to efficiently control MCMV infection in the lungs.
Finally, we asked whether helper or cytotoxic T cells are responsible for the T cell-mediated control of MCMV infection in the lungs. To address this question, we transferred CD4 and CD8 T cells isolated from naïve wild-type donors, separately or together, into MCMV-infected Rag2-/- recipients (Fig 7A). We found that CD4 T cells in the absence of CD8 T cells were unable to suppress virus replication (Fig 7B). In contrast, CD8 T cells in absence of CD4 T cells had a detectable antiviral effect (Fig 7B). However, only the combination of CD4 and CD8 T cells lead to effective control of the infection indicating a cooperative effect of these two cell populations (Fig 7B). Taken together, these data suggest that i) for T cells secretion of IFNγ is essential to interfere with local MCMV replication in NIFs, ii) IFNγ-receptor signaling on T cells is not necessary to mediate their antiviral effect and iii) CD4 and CD8 T cells work together to fight virus infections in MCMV-infected lungs.
The immune response initiated by CMV infection can protect the host from severe disease. However, several immunosuppressive conditions can lead to recurrence of an unnoticed primary infection and during pregnancy primary infection or reactivation can lead to vertical transmission of the virus to the neonate [18]. Therefore, it is important to understand why the immune system can control, but not eradicate CMV infection in different organs. There are still many uncertainties regarding the antiviral factors needed to control CMV infection. In the mouse model, a classical read-out to assess the viral load of an organ is to determine viral titers via plaque assays or quantification of reporter virus-expressed luciferase activity within an organ of interest. Although these approaches can give a general overview of the viral load in individual organs, the localization of infected cells, the infected cell type, and the micro-anatomical context remain enigmatic.
In the present study, we combined a MCMV reporter mutant with various genetically modified mice to study the immune response to lung infection in vivo. The use of a reporter virus encoding the red fluorescent protein mCherry allows for simultaneous quantification of the infection as well as visualization of the immune cell-mediated antiviral response. Previously, we applied this method to assess the killing capacity of specific effector CD8 T cells [10] and to study differences between MCK2-proficient and -deficient virus mutants [27] as well as the immune responses in lungs of neonatal mice [26].
For the lungs, it has been reported that CMV infection causes pneumonia in immunocompromised humans [18] as well as in mice [29]. Focal infiltrations of immune cells following CMV infection have been observed in mice [29,37] and humans [28] and characterized as NIFs [26]. NIFs have been described as sites of infection but also as the site of antiviral immune response and we proposed recently that NIFs act as sites for T cell priming, at least in neonatal mice [26]. Thus, it is still unclear how NIFs are formed and how these structures contribute in detail to MCMV control in the lungs. Here, we showed that NIF formation is independent of the presence of lymphocytes and identified T cells as essential contributors to the antiviral effects observed within NIFs. Presumably, the recruitment of myeloid cells is triggered by the infected cells directly or by resident cells in close proximity to infected cells. In parallel to this first wave of cellular influx lymphocytes locate in forming NIFs to mediate their antiviral effect directly at the site of infection. Once the infection is kept under control within NIFs the pro-inflammatory signals vanish and the structures dissolve to allow tissue healing. In the liver, CCL3 was reported to be essential for the infiltration of immune cells after MCMV infection [38], while CCL5 was proposed as a possible recruitment factor for T cells into the lung interstitium [39,40]. However, the factors inducing and sustaining NIFs in the lungs and chemokines involved in recruiting immune cells to these structures need further investigation.
Primary MCMV infection of the adult lungs can be controlled within approximately one week in immunocompetent mice [26]. In contrast to wild-type mice, we find that various strains of immunocompromised mice are impaired in reducing the number of infected cells in the lungs. NIFs with intact infected cells are still detectable 8 dpi in immunocompromised but not in wild-type mice. Although we used MCMV reporter viruses deficient for m157 that encodes a ligand for the NK activating receptor Ly49H [22], our results indicate that NK cells contribute to virus control in NIFs even in the absence of this strongly activating ligand especially in a T cell deficient host. Therefore, the NK cell-mediated contribution to virus control must be induced by other mechanisms that have not been addressed here. In fact, we found NK cells to be the major population of lymphocytes able to secrete IFNγ at 5 dpi. Notably, an IFNγ-producing cell population lacking T and NK cell surface markers has been identified in Rag1-/- and Rag2-/-Il2rg-/- mice [35]. Furthermore, the secretion of perforin might be crucial for the NK cell mediated control, since we found increased virus loads in perforin-deficient mice while this effect was apparently not T cell-mediated.
Using various experimental approaches, we show in lungs that T cells have a major effect on NIF dynamics and virus control in NIFs. Effector CD8 T cells are known to essentially contribute to the control of MCMV infection in mice [41–44] and humans [4] and adoptive transfers of specific T cells are applied as clinical therapy for several years now [7]. However, so far the efficacy of T cells controlling the primary infection in the lungs of immunocompetent mice has not been addressed in detail. We found that T cells migrate into NIFs as early as these structures can be detected (3 dpi); with an increasing number of T cells being present until 8 dpi. The observed increase of T cell counts in NIFs can be either due to enhanced recruitment to the site of infection or to local proliferation following priming in NIFs—as shown before in neonatal mice [26]–or both. Earlier studies revealed that the adoptive transfer of 105−107 T cells (either activated lymphocytes from lymphoid organs, or pulmonary T cells, or pulmonary CD8 T cells from previously infected donors) were sufficient to reduce the viral titer or the amount of infected cells in several organs of infected immunocompromised recipients [42–44]. In contrast, our data indicate that only the adoptive transfer of higher numbers of naïve T cells (107) led to T cell densities in NIFs that contributed to sufficient local virus control to a similar degree as observed in WT mice 8 dpi. This high number of polyclonal naïve T cells is essential to provide sufficient numbers of precursors for MCMV-specific T cells that can be expanded and subsequently control the infection.
The role of CD4 T cells in controlling cytomegalovirus infection has been discussed controversially. Several adoptive transfer experiments showed that CD4 T cells on their own cannot control MCMV infection [42,45,46]. Nevertheless, CD4 T cells have been shown to be necessary to control MCMV in the salivary glands [16,17] and other organs [47] in mice and clinical case reports revealed that the presence of CD4 T cells might be crucial for efficient viral control in patients [7,48–53].
Applying adoptive transfers into Rag2-/- mice, we showed that CD4 T cells support CD8 T cells in controlling virus infection in the lungs. This effect might be mediated by a type of T helper-1 function supporting CD8 T cell differentiation as suggested elsewhere [54] but it seems also likely that CD4 T cells possess cytotoxic effector functions directly contributing to the control of infection [7,55]. Intracellular staining for IFNγ revealed that around 20% of the CD4 T cells secrete IFNγ indicating that a considerable proportion of the CD4 T cells show a T helper-1 phenotype.
Furthermore, we showed that the secretion of IFNγ by T cells was essential to reduce the number of infected cells while the absence of perforin did not affect the T cell mediated control. IFNγ has been reported to have multiple effects on different cell types [56–58]. For example, IFNγ drives differentiation of monocytes into macrophages that mediate anti-microbial effects [59]. Likewise the function of the myeloid cells abundantly present within NIFs might also be affected by the locally secreted IFNγ. In an in vitro assay we found that IFNγ affected plaque size of lung stromal cells after MCMV infection, indicating that IFNγ stimulates intrinsic cell mechanisms to interfere with virus replication [60,61]. Interestingly, not only pretreatment but also treatment of cells with IFNγ after an initial MCMV infection can result in an inhibiting effect on the viral spread [36]. It is known that IFNγ can counteract to some degree MCMV-mediated down-modulation of MHC I that in turn might directly facilitate CD8 T cell-mediated killing [62]. Previously, we reported that the presentation of viral antigen via MHC I by MCMV-infected target cells and the following direct recognition of the infected cell by antigen-specific effector CD8 T cells is essential for the cytotoxic killing [10]. In addition to IFNγ and perforin other factors, such as TNFα and Fas/FasL, play an important role in the T cell effector function and must be addressed in similar models in the future.
Based on the data presented in this manuscript we envisage the following model how T cells contribute to the local control of MCMV in lung NIFs: T cells are primed in lymphoid organs and within NIFs to secrete IFNγ directly at the site of infection. This secretion of IFNγ not only slows down the viral cell-to-cell spread but might also interfere with MHC I molecule downregulation on the surface of infected cells allowing for better recognition of viral antigen by MCMV-specific CD8 T cells. Furthermore, the IFNγ secretion by NK, as well as both CD4 and CD8 T cells, can stimulate macrophages to efficiently phagocytose infected cell remnants and clear the tissue from infectious virus particles. In parallel, NK cells contribute to MCMV control in a perforin-dependent manner. Thus, a combination of virus-specific CD8 and CD4 T cells, with the addition of NK cells, might be an optimal recipe for anti-CMV adoptive immunotherapy.
All animal experiments were performed according to the recommendations and guidelines of the Federation of European Laboratory Animal Science Associations and Society of Laboratory Animals and approved by the institutional review board and the Niedersächsische Landesamt für Verbraucherschutz und Lebensmittelsicherheit (33.12-42502-04-10/0225, 33.12-42502-04-12/0921, 33.12-42502-04-13/1255, and 13.12-42502-04-17/2737). All experiments were performed according to the Tierschutzgesetz and the Tierschutz-Versuchstier-Verordnung. Anesthesia of mice was performed with Ketamin and Xylazin injection. Mice were sacrificed following CO2 anesthesia by cervical dislocation.
Mice were bred at the central animal facility of Hannover Medical School under specific pathogen free conditions. The different mouse strains (Ighmtm1Cgn [63], Cd3etm1Mal [64], Rag2tm1 [65], Rag2tm1Il2rgtm1[66,67], Ifngtm1Ts [68], Ifngr1tm1Agt [69], Prf1tm1Sdz [70], Ncr1gfp/wt [71]) were maintained on a C57BL/6 (B6) background. Rag2-/-Prf1-/- and Rag2-/-Ifng-/- mice were generated by intercrossing Rag2-/- and Prf1-/- or Rag2-/- and Ifng-/- mice, respectively. 6–14 weeks old mice were used for the experiments. All mouse experiments were performed in accordance with local animal welfare regulations.
The MCMV strain named MCMV-3D has been described previously [20]. Virus stocks were produced and titrated on mouse embryonic fibroblasts. MCMV-3D encodes Gaussia luciferase and mCherry and carries an additional sequence within the m164 ORF encoding the SIINFEKL peptide. It lacks the m157 ORF that encodes a ligand for the activating receptor Ly49H present on a subset of NK cells in C57BL/6 mice and expresses a truncated version of the viral MCK2 protein [27,30]. In some experiments, the reporter virus MCMV-3DR with a repaired m129 ORF resulting in an intact MCK2 protein was used for intranasal infection [27]. All animals were infected intranasally with 106 PFU MCMV-3D or MCMV-3DR. In some experiments C57BL/6 mice were irradiated with a total dose of 6 Gray in one irradiation procedure the day before intranasal infection.
In experiments Rag2-/- mice were used as recipients of adoptive T cell transfers. T cells were isolated from lymph nodes and spleens of untreated wild-type B6, Ifng-/-, Ifngr1-/-, or Prf1-/- mice using MACS pan T cell, CD4 T cell, or CD8 T cell negative selection kits (Miltenyi Biotech) and adoptively transferred intravenously in parallel to MCMV infection. For in vivo depletion of NK cells in B6 and Rag2-/- mice 300 μg anti-NK1.1 (antibody clone PK136) was applied intraperitoneally four hours before and every second day after MCMV infection.
Organs were perfused in situ with cold PBS, explanted and stored in 500μL DMEM containing 10% FCS and 1% P/S. Next, the organs were mechanically disrupted by metal beads and shaking at 25/s over 4 min (TissueLyser II, Quiagen). Then, 350 μL of the homogenate were separated and stored at -80°C for viral plaque titration, while the rest was centrifuged at maximum speed for 15 min. 20 μL of the substrate (1 μg Coelenterazine / ml in PBS) were added freshly to 180μL of a 1:10 dilution of the organ homogenate and measured for 10s directly with Lumat LB 9507 (Berthold Technologies) in duplicates. The viral plaque titration was performed on MEF monolayers in triplicates.
Mice were sacrificed and lungs were perfused with cold PBS via the right heart ventricle and then filled in situ with PBS solution containing 2% paraformaldehyde and 30% sucrose. Fixed lungs were embedded in OCT compound (TissueTek, Sakura) before freezing at -20°C.
7 μm thick cryosections were blocked with rat serum in TBST and stained with antibodies and nuclei were stained with DAPI. The following antibodies (clones) were used after additional blocking of Fc receptors with rat anti-CD16 and anti-CD32 (2.4G2): CD11b-bio (MAC1), CD3-Cy5 (CD317A2), CD11c-APC (N418), CD45.2-FITC (104), CD45-APC (30-F11), and CD8a-PE (53–6.7). For staining with anti-mouse NK1.1-APC (PK136) and GFP-binding protein (GFP-boost, Chromotek), rehydration was performed in 0.5% Triton-X following blocking with mouse serum and Fc receptor blocking, while slides were stained in 1%Triton-X. Images were taken with an AxioCam MRm camera (Carl Zeiss) attached to an Axiovert 200M fluorescence microscope (Carl Zeiss) with PlanApochromat objectives 10x/0,45 and 20x/0,75 (magnification/numerical aperture) and with AxioScan.Z1 (Carl Zeiss) and processed with AxioVision 4.8.2 software and with ZEN blue 2.3 software.
Lung sections were analyzed at defined anatomical positions (through the center; slices included right and left lobes as well as main bronchi). Sections were stained with DAPI. For each animal, 4 lung sections were analyzed by an observer blinded for mouse identity. Manual counting of virus-infected cells and NIFs was performed directly at the microscope. Intact cells were distinguished from remnants of infected cells based on their mCherry expression, morphology, and DAPI staining while NIFs were quantified based on localized accumulation of nuclei as well as presence of mCherry signals. The number of independent experiments performed is indicated in each figure. In general, we only show data from at least 4 mice and at least 2 independent experiments. For the analysis of infected cells per NIFs, microscopy pictures were taken with the 20x/0.75 objective. T cells were quantified based on CD3 staining, CD8 T cells were quantified based on CD8a staining, and NK cells were quantified based on GFP expression in Ncr1gfp/wt mice and additional staining with a GFP-binding protein in the area of infiltration that was semi-automatically measured based on CD45 staining using ImageJ. Between 10 and 15 NIFs were analyzed per animal.
Leukocytes from lungs were isolated as described before [26]. Briefly, right heart ventricle was perfused with PBS until blood cells were removed from the lungs. Fragmented tissue was digested with Collagenase D (Roche, 0.5 mg/ml) and DNAse I (Roche, 0.025 mg/ml) for 45 min at 37°C, meshed through 40 mm Falcon Cell Strainer and leukocytes isolated with Lympholyte-M (Cedarlane) gradient centrifugation technique. For tetramer staining, lung leukocytes were stained with a mix of APC-tetramers specific for MHCI complexes presenting M45, M38, or m139 for 15 min at 4°C, following a surface staining for 15 min on ice. Tetramers were provided by Ramon Arens (LUMC, Netherlands). For intracellular staining, lung leukocytes were cultured in vitro in 200 μL RPMI/FCS/HEPES. For re-stimulation 50 ng/mL PMA and 500 ng/mL Ionomycin were added and cells were incubated for 2 hours at 37°C with 5% CO2. To allow detection of intracellular proteins 10 μg/mL Brefeldin A were added and incubated for another 2 hours at 37°C and 5% CO2. Fc-receptors were blocked in PBS with 3% FCS (PBS/FCS), 5% serum, and rat antibodies to mouse CD16 and CD32 (2.4G2). Surface staining was performed for 15 min on ice following the fixation and intracellular staining procedure with the Intracellular Fixation & Permeabilization Buffer Set (eBiosciences) according to the manufacturer protocol. Cells were re-suspended in PBS/FCS for flow cytometry analysis using a LSRII (BD Biosciences). Flow cytometry data was analyzed using FlowJo 7.5 and 10. Gates for IFNγ+ cells were set based on the border of 99% of the isotype control. Cells from in vitro culture experiments were trypsinized at room temperature (for 3 min using PBS containing 0.5 mg/mL Trypsin and 3 mM EDTA), blocked for Fc-receptors and stained with antibodies before FACS analysis. Following antibodies (clones) were used for flow cytometry staining: CD8a-PE (53–6.7), CD8-APC (RmCD8-2), Thy1-Alexa488 (RMT1), CD103-FITC (M290), CD31-FITC (MEC13.3), Ly6G-PE (1A8), CD11b-eF450 (M1/70), CD45.2-APC (104), EpCAM-eF450 (G8.8), F4/80-APC-eF780 (BM8), gp38-bio (eBio8.1.1), Ly6C-PerCP-Cy5.5 (HK1.4), CD11c-PeCy7 (N418), CD19-BV510 (6D5), CD4-PerCP (RM4-5), CD45.2-PerCP-Cy5.5 (104), IFNγ-PE (XMG1.2), NK1.1-PeCy7 (PK136), Pdgfra-PE (APA5), and Pdgfrb-APC (APB5).
Animals were sacrificed and the right heart ventricle was perfused with cold PBS to remove blood cells from the lungs. Lungs were explanted, chopped into pieces and digested in RPMI with 0.125 mg DNAse I and 0.269 mg collagenase D at 37°C for 2h. Cell suspensions were treated with erythrocyte lysis buffer and stromal cells either purified by depletion of CD45+ cells with the use of a MACS CD45 positive selection kit (Miltenyi) or cultured as a mixed population. Cells were cultured in RPMI with FCS, P/S and L-Glutamine, and medium was exchanged every day to remove dead and non-adherent cells. After 6 days of culture, cells were trypsinized, seeded on 48 well plates and cultured for additional 7 days to reach a confluent layer. Cells were then infected with 20 PFU/well for 30 mins. Afterwards, virus solution was discarded and cells were covered with carboxy-methyl-cellulose in addition to different concentrations of mouse IFNγ protein (0, 0.1, 1, 10, 100 and 1000 ng per well). After 4 and 8 days, mosaic images were taken with Zeiss Axiovert M microscope, and plaque sizes were measured with ImageJ.
Statistical analysis was performed using GraphPad Prism 4. A non-parametric two-tailed t-test (Mann-Whitney test) was applied to compare data from experimental groups to that of wild-type mice (Fig 2, Fig 3, Fig 5A–5C, S1 Fig, S2 Fig, S3 Fig, and S4 Fig) or of Rag2-/- mice (Figs 4D, 4E, 5D, 5E and 7). In Fig 4F, the Wilcoxon signed-rank test was applied comparing experimental groups to a fixed value of 0, since Rag2-/- do not harbor any T cells. Significant differences are marked as follows: P>0.05 (ns), P<0.05 (*), P<0.01 (**), P<0.001 (***).
|
10.1371/journal.pntd.0003127 | Viral Aetiology of Central Nervous System Infections in Adults Admitted to a Tertiary Referral Hospital in Southern Vietnam over 12 Years | Central nervous system (CNS) infections are important diseases in both children and adults worldwide. The spectrum of infections is broad, encompassing bacterial/aseptic meningitis and encephalitis. Viruses are regarded as the most common causes of encephalitis and aseptic meningitis. Better understanding of the viral causes of the diseases is of public health importance, in order to better inform immunization policy, and may influence clinical management.
Study was conducted at the Hospital for Tropical Diseases in Ho Chi Minh City, a primary, secondary, and tertiary referral hospital for all southern provinces of Vietnam. Between December 1996 and May 2008, patients with CNS infections of presumed viral origin were enrolled. Laboratory diagnostics consisted of molecular and serological tests targeted at 14 meningitis/encephalitis-associated viruses.
Of 291 enrolled patients, fatal outcome and neurological sequelae were recorded in 10% (28/291) and 27% (78/291), respectively. Mortality was especially high (9/19, 47%) amongst those with confirmed herpes simplex encephalitis which is attributed to the limited availability of intravenous acyclovir/valacyclovir. Japanese encephalitis virus, dengue virus, herpes simplex virus, and enteroviruses were the most common viruses detected, responsible for 36 (12%), 19 (6.5%), 19 (6.5%) and 8 (2.7%) respectively, followed by rubella virus (6, 2%), varicella zoster virus (5, 1.7%), mumps virus (2, 0.7%), cytomegalovirus (1, 0.3%), and rabies virus (1, 0.3%).
Viral infections of the CNS in adults in Vietnam are associated with high morbidity and mortality. Despite extensive laboratory testing, 68% of the patients remain undiagnosed. Together with our previous reports, the data confirm that Japanese encephalitis virus, dengue virus, herpes simplex virus, and enteroviruses are the leading identified causes of CNS viral infections in Vietnam, suggest that the majority of morbidity/mortality amongst patients with a confirmed/probable diagnosis is preventable by adequate vaccination/treatment, and are therefore of public health significance.
| Central nervous system (CNS) infections are important diseases worldwide. The spectrum of infections is broad, encompassing bacterial/aseptic meningitis and encephalitis. Viruses are regarded as the most common causes of encephalitis and aseptic meningitis. Better understanding of the causes of the diseases is of public health importance, in order to better inform immunization policy, and influence clinical management. We describe the clinical features and infectious causes of 291 adults with clinically suspected CNS infections of presumed viral origin. We show that CNS viral infections in Vietnam are associated with high morbidity and mortality. Mortality was especially high (47%) amongst those with herpes simplex encephalitis which is attributed to the limited availability specific antiviral drugs in our setting. Japanese encephalitis virus, dengue viruses, herpes simplex virus and enteroviruses were the most common viruses detected, followed by rubella virus, varicella zoster virus, mumps virus, cytomegalovirus, and rabies virus. Our study represents the broadest yet investigation of the possible viral causes of the CNS infections in adults in Vietnam, with a diagnostic yield of 32%. The results show that the majority of morbidity/mortality amongst patients with a confirmed/probable diagnosis could be prevented by adequate vaccination or treatment, and are therefore of public health significance.
| Central nervous system (CNS) infections are important diseases worldwide. The spectrum of infections is broad, encompassing bacterial meningitis, aseptic meningitis and encephalitis. The estimated incidence of encephalitis worldwide is between 3.5 and 7.4 cases per 100,000 person years [1]. While the clinical course of viral meningitis and encephalitis may overlap, viral meningitis is usually self-limiting [2], whereas the mortality from viral encephalitis ranges from 4.6% to 29% [3]–[8] and nearly 50% of survivors have persistent neurological sequelae after 6 months follow up [7]. Viruses are regarded as the most common causes of encephalitis and aseptic meningitis, and the specific viral aetiology of the diseases is diverse and dependent on geographical, temporal, host-immunity and age factors [3], [7]–[10]. The aetiology is in part driven by the introduction of immunization programs and/or the (re)emergence of (new) pathogens.
In Southeast Asia, in 1998 Nipah virus emerged in Malaysia and Singapore [11], [12] and spread to Bangladesh where it causes annual outbreaks of fatal encephalitis [13]. Over the last 16 years, enterovirus 71 has emerged and caused large outbreaks of hand foot and mouth disease, sometimes associated with fatal encephalitis in young children. Likewise, Japanese encephalitis virus (JEV) is a leading cause of encephalitis in children and occasionally in young adults in many countries in Asia including Vietnam [9], [14]–[16]. In Vietnam, JEV vaccine was first introduced in 1997, and had been administered to all children 1–5 years of age in 437 (65%) of 676 districts by 2007 [14]. By 2008, 91% of the target population in Vietnam had received JEV vaccination [17].
Studies in Western countries have revealed that herpes simplex virus (HSV) varicella-zoster virus (VZV) and enteroviruses (EVs) are the leading causes of encephalitis/aseptic meningitis in adults [4], [7], [18], whereas in two recent prospective descriptive studies in Vietnam JEV, dengue virus (DENV), HSV, EVs and VZV were frequently detected in adults CNS infections of presumed viral origin [9], [19]. Of note, in these two studies virus diagnostic tests were limited to these 5 types, possibly underestimating the aetiological diversity of the infections. Better understanding of the causes of the diseases is of public health importance, in order to better inform immunization policy, and may influence clinical management. Herein we report the results of a 12 year study investigating the clinical and laboratory features of 291 HIV uninfected adults with presumed viral CNS infections admitted to a tertiary referral hospital in southern Vietnam.
The study was conducted at an infectious disease ward of the Hospital for Tropical Diseases (HTD) in Ho Chi Minh City, a primary, secondary, and tertiary referral hospital for infectious diseases in both children and adults for all southern provinces of Vietnam. The hospital has 550 beds, 35,000 admissions annually, and serves a population of 42 million people. Any adult with severe CNS infections in southern Vietnam is referred to HTD. The patients who present to this hospital are therefore representative of the whole of southern Vietnam. Patient enrolment started in December 1996 and is on-going. The present study reports findings of the 291 consecutive patients enrolled between December 1996 and May 2008.
All adult patients (age ≥15 years) presenting with clinically suspected CNS infections of presumed viral origin, based on the clinical judgment of admitting physicians, negative HIV serology, and with no evidence of purulent bacterial, eosinophilic, cryptococcal and tuberculous meningitis by CSF cell count, culture and/or microscopy, were eligible to enter the study.
Detailed demographic and clinical data, including routine blood and CSF haematology and biochemistry were collected on case record forms at enrollment and during hospitalization. Clinical outcomes were assessed at discharge using the Glasgow Outcome Scale and was defined as death, full recovery (no abnormality), and severe (greatly affecting function, i.e. dependence), moderate (deficit affecting function but not dependence), or minor (abnormality detectable but not affecting function) neurological sequelae, based on neurological examination, degree of independent functioning and controllability of seizures [20]. For microbiological investigations, acute CSF and serum were collected at enrolment.
As part of routine care, CSF specimens of the enrolled patients were cultured and/or examined by microscopy for detection of bacterial/C. neoformans/M. tuberculosis infection with the use of standard methods if clinically indicated. Eosinophilic meningitis was diagnosed by CSF examination, and defined by the presence of more than 10 eosinophils per ml of CSF or >10% of total CSF white cells. All patients were tested for antibodies to HIV.
Between 1997 and 2004, oral acyclovir 4 g per day was given if herpes simplex encephalitis (HSE) was suspected because intravenous (IV) acyclovir was not available in our hospital at that time. From 2005, patients with suspected HSE were either given oral valacyclovir 3 g per day or oral acyclovir 4 g per day. IV acyclovir 1.5 g per day was only given to patients who could afford to pay for their medications pending HSV PCR results. Irrespective of the aetiology, supportive therapy for patients with encephalitis of presumed viral infection was an important cornerstone of management. Seizures were controlled with IV benzodiazepines, phenytoin or phenobarbital, and where necessary sedation and mechanical ventilation. Medical management of raised intracranial pressure included elevating the head of bed, IV mannitol, and intubation with mechanical hyperventilation. Careful attention was paid to the maintenance of respiration, cardiac rhythm, fluid and electrolyte balance, prevention of deep vein thrombosis, aspiration pneumonia, and secondary bacterial infections.
Patients were categorised as having confirmed, probable or possible diagnoses, and no aetiological agent found (see Table 1). A confirmed diagnosis was established if viral nucleic acid or viral specific IgM was detected in CSF for JEV, DENV, HSV, EVs, mumps and VZV. In some instances, the detection of the virus in CSF or in other body fluids alone is insufficient and requires further supporting evidence for interpretation of the results. Details are presented in Table 1.
Chi-square test, Fisher's exact test, independent samples t test and the Wilcoxon rank-sum test were used to compare data between groups of patients when appropriate by using either SPSS for Windows version 14 (SPSS Inc, Chicago IL, USA) or statistical software R version 2.9.0 (http://www.r-project.org).
For the period between 1996 and October 2007, the samples analysed were anonymized, residual CSF and blood specimens that were collected as part of routine care. The use of these clinical specimens for the purpose of improving knowledge about the causative agents of CNS infections in Vietnam was approved by the Scientific and Ethical Committee of the Hospital for Tropical Diseases, Vietnam.
From November 2007 onward, the prospective study of patients was approved by the Scientific and Ethical Committee of the Hospital for Tropical Diseases, Vietnam and the Oxford University Tropical Research Ethics Committee, UK, OxTREC number: 004-07. Physicians entering patients into the study were responsible for obtaining their written informed consent from the patient. If the patient was unconscious, the written informed consent was obtained from a relative or a family member.
Between December 1996 and May 2008, 1601 patients with CNS infections were admitted to the infectious ward of the HTD, of whom 291 fulfilling the entry criteria of CNS infections of presumed viral origin were enrolled. 190 (65%) were male, and 235 (89%) were referred from other hospitals. The median duration of hospital stay was 10 days [interquartile range (IQR): 7–18]. A total of 71 (25%) patients received acyclovir or valacyclovir: oral acyclovir (n = 39), IV acyclovir (n = 2) and oral valacyclovir (n = 30) (Table 2).
A low Glasgow coma score (≤9) was recorded in 82 (29%) of the patients on admission. A fatal outcome was recorded in 28 (10%) patients; and 78 (27%) patients suffered neurological sequelae at discharge, which was severe in 49 (17%), moderate in 17 (6%) and mild in 12 (4%). There were no differences in outcome detected between patients who had confirmed/probable causes identified versus those with possible/no aetiological agent identified (Table 2).
The patient's characteristics, clinical outcome and distribution of the enrolled patients over the study period are presented in Table 2.
Virus and bacterial specific PCR investigations were performed on CSF specimens of all 291 enrolled patients, while serologic tests for IgM antibodies against DENV and JEV were done in 278 patients due to insufficient volumes of CSF from the remaining patients (Table 3). Rabies PCR was performed on saliva of 1 patient with a clinical syndrome suggestive of rabies. Serologic testing for IgM antibodies to rubella virus was performed on sera of six patients with a non-confluent maculopapular rash, suggestive of rubella on admission. Further testing for the presence of antibodies to EBV (VCA IgM, VCA IgG and EBNA IgG) and CMV (IgM and IgG) were done on available acute blood samples from 22/24 and 4/5 patients that had viral DNA detected in CSF by EBV and CMV PCRs, respectively (Table 3).
We found confirmed/probable viral infection in 93 (32%) patients (see Table 4). JEV was the most common virus detected (n = 36, 12%) followed by DENV (n = 19, 6.5% - CSF serology, n = 17 and CSF PCR, n = 2), HSV (n = 19, 6.5%) and EVs (n = 8, 2.7%) (Table 4).
Characteristics and clinical outcome of patients with or without a confirmed/probable viral aetiology, and patients with JEV, DENV, EVs and HSV infections, are presented in Tables 2, 3 and Table 5 (with additional data for DENV patients). Overall, the characteristics of patients with or without a confirmed/probable viral aetiology were similar, although patients with a confirmed/probable viral aetiology were significantly younger, had higher frequency of history of fever, vomiting and headache on admission, and higher CSF white blood cell counts, but they required a shorter hospital stay (p<0.05 for all).
The incidence of death (47%) and neurological sequelae (26%) by discharge in the 19 patients with HSV encephalitis (HSE) was significantly higher than that of patients with other infections (Table 3). Nine (47%) patients with HSE received acyclovir/valacyclovir; in one of whom the discharge outcome was not fully assessed because of being transferred to another hospital. There was a trend towards better outcome amongst those who received drug earlier. The median days of illness before acyclovir/valacycolovir administration was 3 days (range 3–5) in survivors compared to 7 days (range 5–13) (P = 0.05) in those who died.
Patients infected with JEV were mostly young adults and were significantly younger than the other patients: median age 18 years (IQR 16–22) compared to 28 years (IQR 17–38; P = 0.02) for those with DENV and 32 years, (IQR 21–41; P<0.001) for those with HSE.
A possible viral cause was detected in 28 (10%) patients. EBV DNA was detected by PCR in CSF of 24 patients (Table 4). Among these, evidence of past EBV infection was seen in 22/22 (100%) tested patients (VCA IgG and EBNA IgG positive) but none had evidence of acute infection (negative VCA IgM) (Table 4). Similarly, CMV DNA was detected in the CSF of 5 patients of whom 4 had acute sera available for serologic tests. All had detectable CMV specific IgG, but three had undetectable IgM (Table 4).
Of 291 patients fulfilling the inclusion criteria of CNS viral infections, evidence of bacterial infection by PCR analysis of CSF was found in 8 (2.7%), including 5 with S. pneumoniae, 2 with S. suis serotype 2, and one with PCR positive for both S. suis and N. meningitidis. Three of the 8 bacterial PCR positive patients were treated with antibiotics prior to admission. Clinical outcomes and CSF laboratory data of these patients are detailed in Supplementary Table S1.
Twenty-three patients (8%) had tests that were positive for more than one agent (Table 4). In the majority (16/23, 70%) of these, the CSF was PCR positive for EBV or CMV plus another agent. In those with positive CSF EBV PCR, 6 were positive for JEV, 2 with DENV, and 1 each with HSV, VZV and S. pneumoniae, respectively. In those with positive CSF CMV PCR, 2 were also positive for JEV (n = 2), and 1 each with mumps and EVs, respectively. Among EBV patients there was no statistical difference in viral load (as suggested by the obtained Ct values) between patients with or without other positive results (median Ct value 37, range 34–40 vs. 35, range 32–41, respectively, P = 0.19).
Twenty-eight (10%) patients died during hospitalization. 60% died within the first two weeks of illness, and 93% died within the first two weeks of hospital admission.
Two hundred and ninety one patients were enrolled over a 12-year period (Figure 1A). Encephalitis cases distributed throughout the year with a slight peak in October and November. Similarly, there is no clear seasonal trend for specific viruses except for JEV cases, which peaked in June (Figure 1B), when the southern part of Vietnam enters the rainy season.
We describe the clinical features and infectious aetiology of 291 adults with suspected CNS infections of presumed viral origin admitted to a tertiary referral hospital in southern Vietnam over 12 years. The incidence of death in this study was high at 10% (28/291). And this was particularly marked in patients with confirmed HSE (9/19, 47%). The mortality from viral encephalitis reported in other studies ranged from 4.6% to 20% [3]–[7]. Studies assessing outcome from HSE have reported mortality ranging from 5% in France [4], to 11% in the UK [7] and 18% in the USA [3]. High-dose IV acyclovir is recommended for the treatment of HSE [24], but is often unavailable or unaffordable in resource-poor countries, such as Vietnam. The limited availability of acyclovir/valacyclovir and the lack of its empirical prescription may explain the high mortality from HSE in our study. According to a recent study conducted in our hospital, oral valacyclovir, which is considerably cheaper, may be an acceptable early treatment for suspected HSE in resource limited settings [25].
Neurological sequelae were also frequent in our patients, affecting 27% (78/291). More subtle impairments of cognitive function may have been missed in our assessment; a recent study performed in the UK found 45% of patients to have persistent neurological sequelae 6 months after diagnosis [7]. The lack of specialist rehabilitation services in less well-resourced countries makes the problems of neurological sequelae particularly serious for patients and their families.
The diagnostic yield of 32% of the present study is in accordance with previous findings (16–52%) [3]–[7], [18], and further illustrates the big challenge to establish a confirmed infectious aetiology in patients with acute CNS infections. Together with our previous reports, the data confirm that JEV, DENV, HSV and EVs are the leading causes of CNS viral infections in Vietnam [8], [9], [19], and highlights much-needed efforts for national vaccination campaigns against vaccine preventable diseases due to viruses as JEV, rubella and rabies in Vietnam.
As of 2008, which is at the end of this report, it was recorded that JEV vaccine was administered to 91% of the target population in Vietnam [17]. Because, data on vaccination status of the patients was not available, it remains unknown whether the JEV patients in the present study had received (sufficient doses of) JE vaccine. Of note, in our previous study in children with viral encephalitis in southern Vietnam in 2004, of 191 enrolled children, only 19.5% had received at least one dose of JE vaccine. [8]. Follow-up study is therefore needed to assess the effect of this immunization campaign on the overall incidence of JEV in Vietnam.
Currently, Vietnam is amongst the countries that have yet to include rubella vaccination in their routine immunization programmes [26], and has experienced notable outbreaks between 2005–2007 with ≥3000 cases/year, and ∼800 cases in 2008 [27]. We may have underestimated the number of cases of encephalitis due to rubella in our study, since only patients with rashes where the diagnosis was suspected were tested. Likewise, rabies is responsible 100 cases per year annually in Vietnam. Rabies control can be achieved through vaccination of humans and animals, dogs in particular [28]–[30]. Rabies control in Vietnam is challenging because of limited public awareness (particularly among pet owners), large numbers of stray dogs, and the lack of a national vaccination program [30]–[32].
Neurological manifestations of DENV infection have been recorded in about 0.5–20% and 4–47% of patients admitted with classical dengue and encephalitis-like illness, respectively in endemic areas [33]. Similarly, the 19 DENV patients were all admitted with clinical sing/symptom of acute CNS infection without a typical picture of classical dengue infection. Although currently there are no standardised case definitions or diagnostic criteria available for this clinical entity [33], [34], according to criteria recently proposed by Carod-Artal et al. [33], 15/19 (79%) dengue patients in our study can be classified as having dengue encephalitis.
While the biological significance of the detection of EBV/CMV DNA in CSF remains unknown, the high frequency of EBV/CMV DNA detection in CSF together with other potential CNS pathogens observed in this study confirms the findings of previous reports [35]–[38]. Past infection was also documented in 22/22 and 3/4 EBV and CMV patients, respectively. The detection of EBV/CMV DNA in CSF of these patients may be a result of the inflammatory processes and white blood cell recruitment leading to CSF entry of EBV/CMV infected cells. However, it cannot be ruled out that under certain circumstances (e.g. co-infection with another CNS pathogen) the virus may reactivate and cause or aggravate CNS infection [39]–[41].
Evidence of bacterial infection was detected in 2% of patients with negative CSF Gram stains and culture and with clinical and laboratory data compatible with CNS viral infections, suggesting that PCR should always be considered to exclude CNS infections in patients with treatable bacterial meningitis, particularly in countries as Vietnam where antibiotics are frequently prescribed in community and hospital settings prior to presentation at a facility where a definitive microbiological diagnosis could be made [42].
Our study has some limitations. First, this is a hospital-based descriptive study and patient admission to the research ward is biased by the availability of beds at the time the patients are admitted to the hospital. This in part, explains the fluctuation in patient numbers enrolled over the study period (Figure 1A). Therefore the data may not closely represent for the wider community. Second, we did not look for all potential infectious causes of CNS infections in Vietnam (including measles virus [43]), or for non-infectious (immune or endocrine) causes. Third, diagnostics by IgM testing of acute samples might be suboptimal (e.g. in case of JEV and DENV) both in terms of sensitivity and specificity. Fourth, samples from other body compartments such as rectal and throat swabs were not collected for aetiological investigation in this study (e.g. in case of enterovirus infection), which could have increased the total diagnostic yield [8]. Fifth, undiagnosed cases could be caused by novel pathogens which would have gone undetected. However, this study represents the broadest investigation yet of the possible viral causes of the CNS infections in adults in Vietnam, with a diagnostic yield of 32%. The results suggest that the majority of morbidity/mortality amongst patients with a confirmed/probable viral CNS infection is preventable by adequate vaccination and/or treatment, and are therefore of public health significance.
|
10.1371/journal.pgen.1007031 | SNPs near the cysteine proteinase cathepsin O gene (CTSO) determine tamoxifen sensitivity in ERα-positive breast cancer through regulation of BRCA1 | Tamoxifen is one of the most commonly employed endocrine therapies for patients with estrogen receptor α (ERα)-positive breast cancer. Unfortunately the clinical benefit is limited due to intrinsic and acquired drug resistance. We previously reported a genome-wide association study that identified common SNPs near the CTSO gene and in ZNF423 associated with development of breast cancer during tamoxifen therapy in the NSABP P-1 and P-2 breast cancer prevention trials. Here, we have investigated their roles in ERα-positive breast cancer growth and tamoxifen response, focusing on the mechanism of CTSO. We performed in vitro studies including luciferase assays, cell proliferation, and mass spectrometry-based assays using ERα-positive breast cancer cells and a panel of genomic data-rich lymphoblastoid cell lines. We report that CTSO reduces the protein levels of BRCA1 and ZNF423 through cysteine proteinase-mediated degradation. We also have identified a series of transcription factors of BRCA1 that are regulated by CTSO at the protein level. Importantly, the variant CTSO SNP genotypes are associated with increased CTSO and decreased BRCA1 protein levels that confer resistance to tamoxifen. Characterization of the effect of both CTSO SNPs and ZNF423 SNPs on tamoxifen response revealed that cells with different combinations of CTSO and ZNF423 genotypes respond differently to Tamoxifen, PARP inhibitors or the combination of the two drugs due to SNP dependent differential regulation of BRCA1 levels. Therefore, these genotypes might be biomarkers for selection of individual drug to achieve the best efficacy.
| Many studies have demonstrated that germline genetic variation can contribute to both breast cancer disease risk and treatment response. However, the underlying mechanisms associated with these biomarkers often remains understudied. As part of functional genomic studies following up a case-control genome-wide association study (GWAS) performed with the large and influential National Surgical Adjuvant Breast and Bowel Project P-1 and P-2 SERM breast cancer prevention trials, we investigated the top GWAS SNPs in CTSO gene on chromosome 4 and mechanisms of CTSO involvement in the regulation of BRCA1 and response to therapy. We showed that, based on individual’s genotype, CTSO contributes differentially to tamoxifen response in ERα-positive (ER+) breast cancer cells by regulating ZNF423 and BRCA1levels and that PARP inhibitors can effectively restore tamoxifen sensitivity in subjects with unfavorable genotypes of CTSO and ZNF423 associated with tamoxifen resistance. Our work highlights the potential value of a new biomarker signature involving CTSO and ZNF423-related SNPs for selection of tamoxifen or PARP inhibitors.
| Approximately 80% of breast tumors express estrogen receptor α (ER) [1–3], a receptor that binds and mediates many of the effects of estrogens. Estrogen signaling is known to modulate several processes relevant to breast cancer cell proliferation, predominately as a result of the activity of ER as a transcription factor [4]. Therefore, selective estrogen receptor modulators (SERMs) such as tamoxifen have been widely used clinically in endocrine therapies for patients with ERα-positive (ERα+) breast cancer [5–7]. Tamoxifen is not only effective in the treatment of ERα+ breast cancer, but it is also effective in the chemoprevention of breast cancer [8, 9]. However, resistance to tamoxifen therapy also occurs in that 22.7% of patients treated in the adjuvant setting had recurrence of breast cancer by 10 years in a meta-analysis, and in the prevention setting [10] tamoxifen reduces risk by 49%, but the number needed to treat to prevent one case of breast cancer is in excess of 50 [8]. Several mechanisms have been associated with resistance to tamoxifen [11, 12]. Of particular importance are the effects of estrogen/ER on BRCA1. The BRCA1 protein directly interacts with ERα and inhibits ERα transactivation and downstream signaling [13]. Decreased BRCA1 expression has been shown to be present in 30–40% of sporadic breast cancers [14]. BRCA1 deficiency is known to play a role in breast cancer development. Furthermore, decreased BRCA1 expression results in tamoxifen resistance by altering ERα co-regulator association in breast cancer cells [15]. These findings suggest that BRCA1 may regulate the response of ERα to its canonical ligand E2 and to tamoxifen, a compound known to exert either agonistic or antagonistic activity toward ERα in different cellular and tissue contexts [16]. In addition, BRCA1 is also known to play a major role in the DNA double-strand break (DSB) repair during the S and G2 phases by mediating homologous recombination (HR) to maintain replication fidelity and genome integrity [17]. Studies have demonstrated that BRCA1 dysfunction results in the lack of HR and markedly sensitizes cells to the inhibition of PARP enzymatic activity, which seemed to be attributable to the persistence of DNA lesions that are normally repaired by homologous recombination [18, 19]. Therefore, genetic factors that might contribute to BRCA1 regulation could significantly affect response to drugs like SERMs and PARP inhibitors.
Our previous case-control genome-wide association study (GWAS) performed with samples from the NSABP P-1 and P-2 breast cancer SERM chemoprevention trials identified two SNP signals that were associated with breast cancer risk, including one in which the variant SNP genotype near the CTSO gene was associated with increased risk for the development of breast cancer and a second signal for which the variant SNP genotype in the ZNF423 gene was associated with decreased risk for the development of breast cancer in women treated with tamoxifen or raloxifene [20]. ZNF423 appeared to be a transcription factor that regulated BRCA1 expression in an estrogen-dependent fashion, while CTSO also showed weak estrogen-dependent induction of BRCA1 mRNA expression in a CTSO SNP-dependent fashion [20]. In a separate study, it was also shown that the variant GG genotype for the CTSO rs10030044 SNP was an independent factor indicating a poor prognosis in ER+ breast cancer patients receiving adjuvant tamoxifen therapy [21], which suggested the involvement of this genetic locus in tamoxifen response. CTSO, cathepsin O, is a member of the cysteine protease family that is involved in cellular protein degradation and turnover. Another member, cathepsin D, has been associated with poor prognosis for breast cancer as a result of stimulation of breast cancer cell proliferation, fibroblast outgrowth, angiogenesis, breast tumor growth and metastasis formation [22]. Even though in our previous study, we have observed a correlation between CTSO and BRCA1 in an estrogen and SNP dependent fashion, how CTSO regulates BRCA1 remains unclear. In the current study, based on our prior findings [20], we investigated the possible role of CTSO in drug response and breast cancer risk as a result of the regulation of ZNF423 and BRCA1. Finally, we also explored the role of both ZNF423 and CTSO SNP genotypes to help selection of tamoxifen and PARP inhibitors.
Our previous GWAS involved 592 cases and 1171 matched controls selected from the 33,000 participants enrolled in the NSABP P-1 and P-2 breast cancer prevention trials identified two SNPs on chromosome 4 (rs10030044 and rs4256192) that were associated with breast cancer risk, with odds ratios of 1.42 and 1.44 respectively [20]. To gain a comprehensive understanding of the contribution of genetic variants in that region, together with the two top genotyped SNPs, based on our previous imputation results [20], we chose additional six imputed SNPs associated with increased risk for the development of breast cancer (OR 1.42–1.45) with adjusted p-values < 5.00E-6 (rs6835859, rs4550865, rs10030044, rs62328155, rs11737651, rs6810983, rs4256192, rs11724342). All eight SNPs were located at 5′ of the CTSO gene. These variant SNP genotypes are common with MAFs ranging from 0.39 to 0.45. We then performed linkage disequilibrium (LD) analysis and the analysis showed that all 8 SNPs were in significant linkage with each other. The top two genotyped SNPs, rs10030044 and rs4256192 were in strong LD (r2 = 0.78). The SNP rs10030044 was also in strong LD with the three imputed SNPs: rs6835859 (r2 = 1), rs4550865 (r2 = 1), and rs6810983 (r2 = 1), while the rs4256192 SNP was in strong LD with the other three imputed SNPs: rs11724342 (r2 = 1), rs62328155 (r2 = 1) and rs11737651 (r2 = 1). Because of the importance of understanding breast cancer risk and because P-1 and P-2 are the largest breast cancer chemoprevention trials ever performed, we pursued the possible functional implications of these SNP signals.
We began by analyzing the top 8 SNPs for their associations with expression levels of all genes including CTSO within 1 Mb up- and downstream of the SNPs of interest using the Genotype-Tissue Expression (GTEx) database. Although, we did not find eQTL relationships between these SNPs and CTSO in normal breast tissue in GTEx, significant eQTL associations between the SNPs and CTSO were present in stomach, skin, pancreas, and testis. The variate SNP was associated with higher CTSO expression (p = 0.0077–4.3E-7). We did not observe eQTL relationships between these SNPs and CTSO at baseline in our panel of LCLs for which we had genome-wide genotype data and mRNA expression data [23]. Because 94.2% of the participants on P-1 and P-2 were Caucasian, our GWAS was restricted to only Caucasian subjects [20]. Therefore, we randomly selected LCLs from Caucasians that were either homozygous wild type (WT) or variant for the SNPs 5’ of CTSO to validate the eQTL relationships in a setting mimicking the estrogenic environment in patients. These LCLs were grown in medium containing charcoal-treated serum to deplete the levels of endogenous steroids and supplemented with physiological concentrations of E2. CTSO mRNA and protein were higher in LCLs homozygous for the variant genotype as compared with LCLs homozygous for the WT genotype (p<0.05; Fig 1A). However, the induction of CTSO mRNA was more significant in the WT than variant cells, consistent with our previous finding [20], even though the variant cells had higher baseline level of CTSO (S1 Fig).
We next determined which of the SNPs 5’ of CTSO might influence expression. Our previous study suggested that the expression of CTSO was estrogen-dependent, and only the rs6810983 SNP disrupted an estrogen response element (ERE) for the variant SNP genotype [20]. We decided to directly determine the possible role of these eight SNPs in transcription regulation using luciferase reporter gene assays performed in ZR75-1 breast cancer cells. Specifically, we cloned a 200 bp DNA sequence that included either WT or variant sequence for each of the eight SNPs, together with the CTSO promoter, into the pGL3 basic reporter plasmid. We then transfected these constructs into the ER+ cell line, ZR75-1 cells in a normal medium with 10% FBS. Cells transfected with constructs with variant genotypes for rs10030044 and rs6810983 SNPs displayed 2–3 fold greater luciferase activity than did those transfected with constructs with WT SNP sequences, indicating increased transcriptional activity (Fig 1B)—compatible with the results in LCLs.
We then determined the possible functional effect of CTSO on BRCA1 based on our previous finding [20]. We genotyped the ZNF423 SNP and CTSO SNP in a panel of breast cancer cell lines and chose T47D, CAMA-1, and ZR75-1 cell lines carrying homozygous genotypes for ZNF423 and CTSO SNPs (S1 Table) for further functional study. When CTSO was overexpressed significantly in T47D, CAMA-1, and ZR75-1 cells, there was a striking decrease of BRCA1 protein levels as well as protein levels for the BRCA1 transcription factor, ZNF423, in all the cell lines tested (Fig 2A, left panel). To determine how generalizable this phenomenon might be, we also measured the level of BRCA1 protein in triple negative MDA-MB-231 breast cancer cells. In agreement with ER+ breast cancer cell line data, BRCA1 protein was significantly decreased after overexpressing CTSO in triple negative breast cancer cells (Fig 2A, left panel). Quantitative RT-PCR revealed excellent transfection efficiency of CTSO in all of the cell lines, with modest but statistically significant decreases in BRCA1 transcript levels (Fig 2A, right panel), while ZNF423 mRNA remained unchanged after CTSO overexpression (Fig 2A, right panel). Next, we asked whether CTSO might influence BRCA1 and ZNF423 protein stability through its cysteine proteases activity. Overexpression of CTSO decreased ZNF423 and BRCA1 protein levels in CAMA-1 and ZR75-1 cells, while treatment with the cathepsin inhibitor E-64 resulted in increased levels of BRCA1 and ZNF423 protein (Fig 2B). Previous work has largely focused on CTSO SNP-dependent estrogen induction of CTSO and BRCA1 mRNA in LCLs. Consistent with our previous finding [20], both CTSO and BRCA1 mRNA was moderately induced by E2 in LCLs with WT CTSO SNP genotype (S1 Fig). However, in this study, we further demonstrated that, more importantly, CTSO can also directly regulate BRCA1 protein turnover in breast cancer cells.
Since CTSO is able to stimulate BRCA1 and ZNF423 protein degradation, we determined the possible interaction between CTSO and BRCA1or ZNF423. Immunoprecipitation using CTSO antibody showed endogenous interaction of CTSO with BRCA1 and ZNF423 (Fig 2C). These results indicated that CTSO regulates BRCA1 and ZNF423 protein stability through a cysteine protease- mediated degradation pathway—at least in part.
We next examined possible mechanisms by which CTSO might influence BRCA1 transcription. We first confirmed that knockdown of CTSO resulted in increased BRCA1 expression, both at the mRNA and protein levels in both CAMA-1 and ZR75-1 cells (Fig 3A). Our previous GWAS study had reported that ZNF423 binds to the 5′-flanking region of BRCA1 and regulates BRCA1 transcription [20]. We also showed in the present study that CTSO interacts with ZNF423, leading to ZNF423 degradation (Fig 2), suggesting that CTSO may regulate BRCA1 transcription partially through its effect on ZNF423. In order to identify additional factors involved in the CTSO-dependent regulation of BRCA1 transcription, we performed mass spectrometry screening of a pool of proteins that co-precipitated with CTSO. During this process, we identified 130 proteins that interacted with CTSO (S2 Table). We then interrogated the Cancer Genome Atlas (TCGA) breast cancer data [24] for possible relationships between the expression of BRCA1 and these 130 genes, and identified 20 genes that were associated with BRCA1 with p< 1E-05 (S3 Table). We then knocked down these 20 genes to determine the effect on BRCA1 levels (S2 Fig), and found that knockdown of 4 out of the 20 genes, MTDH, PABPC4L, LMNA, and EEF1A1, resulted in striking decreases of BRCA1 mRNA expression level (Fig 3B), consistent with the TCGA data that showed positive correlations between these 4 genes and BRCA1. Furthermore, in CAMA-1 and ZR75-1 cells, overexpression of CTSO decreased expression of all four genes (Fig 3C), which could explain the down-regulation of BRCA1 mRNA level when overexpressing CTSO (Fig 2A). In summary, these results indicate that the up-regulation of CTSO could reduce BRCA1 levels by promoting the cysteine protease—mediated degradation of MTDH, PABPC4L, LMNA, and EEF1A1 protein levels in addition to the effect on ZNF423 that we had already identified, all of which regulate BRCA1 transcription. Thus, it appears that tumor expression of CTSO may play a role in the regulation of BRCA1 transcription in addition to having an effect on BRCA1 protein degradation.
We hypothesized that, because CTSO regulates BRCA1 stability, it may play a role in endocrine resistance. Previous studies demonstrated that BRCA1over-expression can inhibit cell proliferation by activating p21WAF1/CIP1 [25, 26]. We had demonstrated that CTSO regulates the stability of BRCA1 (Fig 2). Therefore, we next determined whether the down-regulation of CTSO inhibited cell proliferation in breast cancer cells due to the up-regulation of BRCA1. BRCA1 protein increased after CTSO knockdown in CAMA-1 and ZR75-1 cells (Fig 4A, lower panel). Depletion of CTSO inhibited cell growth compared with negative siRNA transfected control cells (Fig 4A, upper panel). To further confirm that the CTSO effect on cell proliferation was mediated through the regulation of BRCA1, we knocked down BRCA1 in cells with down-regulation of CTSO. Knockdown of BRCA1 in CTSO-depleted cells resulted in the abrogation of decreased proliferation due to CTSO depletion in both cell lines (Fig 4A, upper panel). We next tested the effect of CTSO on tamoxifen treatment based on the observations from our previous study [20] and others. In the presence of 100 nM 4OH-tamoxifen (4OH-TAM), CTSO-deficient cells exhibited increased sensitivity to 4OH-TAM compared with negative siRNA-transfected control cells (Fig 4B), and BRCA1 might be responsible for the increased sensitivity since BRCA1 depletion in siCTSO cells significantly decreased 4OH-TAM sensitivity (Fig 4B). These results demonstrated that depletion or inhibition of CTSO can increase BRCA1 levels with potential therapeutic effects, resulting in growth arrest.
Since our previous study had identified ZNF423 and CTSO SNPs that were associated with breast cancer risk [20], both of which appeared to regulate BRCA1, we examined their joint effect on cell proliferation in the presence of tamoxifen or E2 treatment. We utilized a model system consisting of 300 individual human LCLs (100 European-American, 100 African-American and 100 Han Chinese-American subjects). The “Human Variation Panel” that had been SNP genotyped previously and has repeatedly demonstrated its value as a platform to study genetic variants [20, 27, 28]. Specifically, we selected 4 groups of LCLs to perform 4OH-TAM treatment:
Notably, in the presence of 4OH-TAM, the growth of CTSO WT/ZNF423 WT and CTSO V/ZNF423 V cells decreased significantly (Fig 5A and 5B, and Table 1) suggesting that the therapeutic effects of tamoxifen are seen mainly in the CTSO WT/ZNF423 WT and CTSO V/ZNF423 V groups, not the CTSO WT/ZNF423 V and CTSO V/ZNF423 WT groups (Fig 5C and 5D, and Table 1). We also measured BRCA1, CTSO and ZNF423 protein levels in cells with different ZNF423 SNP and CTSO SNP combinations (Fig 6). The estradiol-, 4OH-TAM -dependent and SNP-dependent regulation of BRCA1 protein level was more pronounced against the background of homozygous variant for the CTSO SNP. BRCA1 protein level in the CTSO V / ZNF423 WT group was significantly upregulated in the presence of E2 and then decreased upon addition of 4OH-TAM treatment (Fig 6). The opposite effects on BRCA1 protein level upon treatment of E2 or E2 plus 4OH-TAM were observed in CTSO V / ZNF423 V group compared with CTSO V / ZNF423 WT group (Fig 6B). The higher BRCA1 level in CTSO V / ZNF423 V group compared to the CTSO V / ZNF423 WT group in the presence of 4OH-TAM could explain the tamoxifen response seen in CTSO V / ZNF423 V group, but not in CTSO V / ZNF423 WT group (Figs 4 and 5B and 5D). In the presence of TAM, cells with CTSO W / ZNF423 W genotype were also showed relatively higher BRCA1 levels, even though with this genetic background the baseline BRCA1 was higher compared with other genotype groups (Fig 6B). Therefore, cells with CTSO W / ZNF423 W also benefit from TAM treatment (Fig 5A). We also measured ER level in these four groups of LCLs upon different treatment to account for its potential impact, and did not observe difference in ER level among the four genotype combination groups, furthermore, E2 and TAM treatment did not change the level of ER compared to vehicle treatment for each genotype combination (Fig 6). Therefore, the ZNF423 and CTSO SNPs-dependent effects on TAM response were not due to ER expression level.
When compared the cell proliferation in the presence of different treatments among different genotypes, cells with CTSO V/ZNF423 W showed the fastest growth rate, regardless of whether they received no treatment, estradiol (E2) alone, 4OH-TAM alone, or the combination of E2 plus 4OH-TAM (S3 Fig), while cells with CTSO WT/ZNF423 V grew slowest among all genotype combination groups in all treatment groups (S3 Fig). This was consistent with our previous finding that the odds ratios for CTSO V/ZNF423 W (OR = 5.71) was the highest, and that for CTSO WT/ZNF423 V (OR = 1.00) was the lowest for breast cancer risk in the P-1, P-2 trials [20].
Loss of BRCA1 function leads to defects in the HR DNA repair pathway, which renders cells more sensitive to PARP inhibitors [29–32]. In BRCA1/2 mutated cells, the DSBs at the replication fork caused by PARP inhibitor treatment cannot be repaired, resulting in synthetic lethality and cell death. We have shown that the LCL CTSO WT/ZNF423 WT (Fig 5A) and CTSO V/ZNF423 V (Fig 5B) groups respond to 4OH-TAM treatment but not the CTSO WT/ZNF423 V (Fig 5C) and CTSO V/ZNF423 WT (Fig 5D) groups (Table 1). In addition, comparing the two 4OH-TAM-resistant groups, CTSO WT/ZNF423 V cells showed higher BRCA1 level upon 4OH-TAM treatment than CTSO V/ZNF423 WT cells (Fig 6B). As a result, we hypothesized that the combination of a PARP inhibitor and 4OH-TAM might achieve better therapeutic outcomes in the CTSO V/ZNF423 WT group that displayed lower levels of BRCA1. To determine the effect of a PARP inhibitor in this setting, we treated 4OH-TAM-responsive CTSO WT/ZNF423 WT and CTSO V/ZNF423 V LCLs as well as 4OH-TAM-resistant CTSO WT/ZNF423 V and CTSO V/ZNF423 WT LCLs with either 4OH-TAM alone or 4OH-TAM plus the PARP inhibitor, olaparib. Olaparib did not increase 4OH-TAM sensitivity in the two 4OH-TAM-responsive CTSO WT/ZNF423 WT and CTSO V/ZNF423 V groups (Fig 7A, upper panel, and Table 1). However, olaparib significantly sensitized the 4OH-TAM-resistant CTSO V/ZNF423 WT cells to tamoxifen treatment, but not the CTSO WT/ZNF423 V cells (Fig 7A, lower panel, and Table 1). The differential effects of olaparib in the two 4OH-TAM-resistant groups can be explained, at least partially, by the differences in BRCA1 levels (Fig 6B). Upon 4OH-TAM treatment, the 4OH-TAM-resistant CTSO V/ZNF423 WT cells had lower BRCA1 levels compared with the CTSO WT/ZNF423 V cells, resulting in sensitization by combining olaparib with 4OH-TAM. The 4OH-TAM-resistant CTSO WT/ZNF423 V cells had high level of BRCA1, consistent with olaparib having little effect.
We also confirmed the therapeutic effect of the combination of olaparib and 4OH-TAM in ER+ breast cancer cells, CAMA-1 and ZR75-1 that had WT BRCA1 and were resistant to olaparib (Fig 7B). Knock down of CTSO resulted in striking increases of BRCA1 protein level (Fig 3A), therefore, the addition of olaparib did not increase 4OH-TAM sensitivity (Fig 7B). However, olaparib significantly increased 4OH-TAM sensitivity in cells transfected with negative control siRNA due to lower baseline BRCA1 level comparing with CTSO knockdown cells (Fig 7B, p<0.05). 4OH-TAM showed the 50% inhibitory concentration (IC50) of 11.22 μM for CAMA-1, and 10.17 μM for ZR75-1 cells transfected with negative control siRNA respectively. The IC50 of 4OH-TAM decreased significantly when co-treated with olaparib in negative control siRNA transfected CAMA-1 and ZR75-1 cells (CAMA-1: IC50 = 5.10±0.26μM; ZR75-1: IC50 = 4.70±0.18 μM) (Fig 7B, p<0.05). In summary, these results indicated that the down-regulation of CTSO could increase BRCA1 levels, resulting in decreased cell growth and potential therapeutic effects.
Understanding intrinsic or acquired resistance to endocrine therapy in the treatment or prevention of breast cancer is of great importance [11]. Tamoxifen is still widely used to treat ER+ breast cancer and, along with the SERM raloxifene, are the only FDA-approved drugs for prevention of breast cancer in high-risk women. Our previous GWAS study identified SNPs on chromosome 4, near the CTSO gene that were associated with increased risk for the development of breast cancer during five years of breast cancer prevention therapy with tamoxifen or raloxifene in the NSABP P-1 and P-2 breast cancer prevention trials [20]. Recently, Hato et al reported a correlation between the variant (GG) genotype for CTSO rs10030044 and shorter disease-free survival, and shorter overall survival in hormone receptor-positive breast cancer patients receiving adjuvant tamoxifen therapy [21]. Multivariate Cox regression analysis revealed that this genotype was an independent factor indicating a poor prognosis in hormone receptor-positive breast cancer patients receiving adjuvant tamoxifen therapy [21]. Our previous work has largely focused on CTSO SNP-dependent estrogen induction of CTSO and BRCA1 mRNA in LCLs [20]. However, the exact mechanism by which CTSO regulates BRCA1 is not clear. Adding E2 can have multiple effects on both CTSO and BRCA1. Additionally, the different combinations of ZNF423 and CTSO genotypes also add additional complexity of regulation of downstream proteins like BRCA1. Therefore, in this study, we focused on the possible mechanisms of CTSO gene involvement in the regulation of BRCA1 and response to therapy in different genotype background. The data presented here demonstrated a possible role for CTSO in resistance to tamoxifen, since the down-regulation of CTSO led to the inhibition of cell growth and increased BRCA1 protein level through both regulation of BRCA1 transcription factors and BRCA1 protein degradation in ER+ breast cancer cells. In addition, we obtained evidence that the addition of PARP inhibitor to tamoxifen could reverse resistance to tamoxifen in breast cancer cells with higher levels of CTSO gene expression. Genotypes for ZNF423 and CTSO could regulate gene expression in an estrogen or tamoxifen-dependent fashion, in turn, influencing downstream BRCA1 levels. Therefore, based on individual genotypes, we could potentially select different treatments to achieve the best outcomes, i.e. precision breast cancer prevention or therapy.
CTSO is a cysteine protease. This class of proteases mediates catabolism of intracellular proteins and selectively activates extracellular protein degradation, macrophage function, and bone resorption [33]. Cysteine proteases have been shown to function extracellularly as well as intracellularly [34, 35], and have been suggested as potential targets for anti-cancer therapy [35, 36]. Cathepsins B, D, H, L, or L2 are thought to play a role in several cancers [37–39]. The role of cathepsins in resistance to cancer therapy is an area of emerging interest [40, 41]. Our current studies demonstrate the mechanisms underlying CTSO-mediated tamoxifen resistance in ER+ breast cancer.
Specifically, our functional genomic studies demonstrated that, among the top 8 SNPs near the CTSO gene from our previous GWAS, the rs10030044 and rs6810983 SNPs could regulate CTSO gene expression, and these SNPs were associated with higher CTSO gene expression levels (Fig 1). We next examined the possible relationship between CTSO expression and that of BRCA1, a gene known to be induced by estrogen exposure through a mechanism that has remained unclear [42, 43].
We found a negative correlation between CTSO and BRCA1 protein levels (Fig 2). Based on our observations of the effect of CTSO on both BRCA1 protein and mRNA levels, we first hypothesized that CTSO might regulate BRCA1 through a cysteine protease -mediated pathway, which we experimentally confirmed by treatment with a cysteine protease inhibitor (Fig 2B). Furthermore, regulation of the transcription of BRCA1 by CTSO was found to be through the regulation of ZNF423 [20], MTDH [44, 45], PABPC4L [46], LMNA [47], and EEF1A1 [48] transcription factors (Figs 2 and 3). MTDH (AEG-1) regulates c-MYC through PLZF, and c-MYC induces BRCA1 gene expression [44, 45]. PABPC4L (Poly A Binding Protein Cytoplasmic 4 Like) is a member of PABP family. PABP recognizes the 3′ mRNA poly (A) tail and plays critical roles in eukaryotic translation initiation and mRNA stabilization/degradation [46, 49]. LMNA (A-type lamin) has been shown to control transcription of BRCA1 [47]. EEF1A1 (translation elongation factor 1-alpha 1) affects gene expression through regulating mRNA stability [48], and could also regulate BRCA1 through E2F1 [50, 51]. Therefore, the ultimate BRCA1 protein level is regulated by CTSO at both transcription as well as protein levels. Decreased BRCA1 has been shown to abolish tamoxifen suppression of cell proliferation [15]. We showed that down-regulation of CTSO increased BRCA1 protein level and inhibited proliferation of ER+ cells with or without tamoxifen treatment (Figs 3 and 4). Inhibition could be restored by co-silencing BRCA1 and CTSO gene expression (Fig 4), suggesting that CTSO may regulate cell proliferation and tamoxifen response through BRCA1.
Our initial GWAS had identified SNPs associated with decreased (ZNF423) and increased (CTSO) risk for breast cancer occurrence [20], both of which appeared to regulate BRCA1. The joint odds ratios for the development of breast cancer while on SERM therapy for five years for these two sets of SNPs ranged from 1.00 for women homozygous for both sets of favorable, low-risk alleles, to 5.71 for women homozygous for unfavorable, high- risk alleles for both ZNF423 and CTSO. In the present study, we also evaluated their joint effect on cell proliferation in the presence of tamoxifen in LCLs carrying different combinations of ZNF423 and CTSO genotypes. We found that the cells homozygous for the favorable alleles of both CTSO and ZNF423 (CTSO W/ZNF423 V) proliferated slowest, while cells homozygous for the unfavorable alleles of both CTSO and ZNF423 (CTSO V/ZNF423 W) proliferated fastest at baseline without treatment (S3 Fig). With tamoxifen treatment, these two genotype groups remained the slowest-growing (favorable) and fastest-growing (unfavorable) groups among the four different genotype groups (S3 Fig), suggesting that tamoxifen had no further effect on the proliferation of cells with these two genotype groups (Fig 5C and 5D). At the mechanistic level, tamoxifen benefit is partially determined by the induction of BRCA1 level. Cells homozygous for one favorable allele and the other unfavorable allele (CTSO W/ZNF423 W, and CTSO V/ZNF423 V groups) responded to tamoxifen treatment (Fig 5A and 5B, Table 1), both of which showed high induction of BRCA1 levels in the presence of TAM (Fig 6B), indicating that patients with these two genotype groups might benefit the most from tamoxifen treatment. For the two tamoxifen-nonresponsive cell groups, in cells carrying CTSO V/ZNF423 W genotypes, PARP inhibitor treatment restored tamoxifen sensitivity (Fig 7A). However, a PARP inhibitor did not sensitize tamoxifen in the other tamoxifen- non responsive cells with the CTSO W/ZNF423 V genotypes, which might be due to the higher level of BRCA1 level in those cells (Figs 6 and 7A, Table 1). Consistent with a previous study [52], we found that cells with lower BRCA1 level due to higher CTSO were very sensitive to PARP inhibition (Fig 7B). The combination of genotyping for CTSO SNPs and ZNF423 SNPs offers the potential for the stratification of ER+ breast cancer patients into different drug response subgroups. Specifically, the use of PARP inhibitors in combination with tamoxifen in patients carrying the CTSO V/ZNF423 W SNP genotypes offers an opportunity for improving tamoxifen sensitivity and prognosis in these patients. The findings of no efficacy for tamoxifen alone or in combination with a PARP inhibitor in patients with the favorable SNP genotype profile of CTSO W/ZNF423 V raises the possibility that alternative approaches to prevention in low-risk patients should be studied in such patients.
In conclusion, we present evidence in the present study that CTSO is a new factor of importance for tamoxifen efficacy as a chemopreventive agent in women at high risk of developing breast cancer as well as evidence for a potential mechanism by which this effect involves BRCA1. The underlying mechanisms identified require validation and further refinement but they also provide pharmacogenomic insights into tamoxifen as a preventative agent. We have demonstrated that a PARP inhibitor, which can effectively restore tamoxifen sensitivity in tamoxifen—resistant ER+ breast cancer cells, might be a potentially promising addition to tamoxifen as a combination regimen for patients carrying the CTSO V/ZNF423 W SNP genotype. As a result, our study has revealed a new potential biomarker signature involving CTSO and ZNF423-related SNPs for the therapeutic stratification of patients at high risk for the development of breast cancer.
Dulbecco's minimum essential medium (DMEM), glutamine and penicillin/streptomycin/glutamine stock mix were purchased from Life Technologies, Inc. (Carlsbad, CA, USA). Fetal bovine serum (FBS) and charcoal-stripped FBS were from Invitrogen (Carlsbad, CA, USA). L-trans-Epoxysuccinyl-leucylamido (4-guanidino) butane (E-64) was from Sigma-Aldrich (St. Louis, MO USA). CTSO, MTDH, PABPC4L, LMNA, EEFiA1and control small interfering RNAs (siRNA) were purchased from Dharmacon (Thermo Scientific Dharmacon, Inc.). CTSO plasmid was purchased from OriGene (Rockville, MD, USA). Affinity purified rabbit and mouse antibodies against human BRCA1 and CTSO were from Santa Cruz Biotechnologies (Santa Cruz, CA, USA). ZNF423 antibody was purchased from Abcam (Cambridge, MA, USA). Actin, MTDH, PABPC4L, LMNA, and EEFiA1 antibodies were from cell signaling (Danvers, MA, USA). For standard PCR, HotStart Taq Plus DNA Polymerase was used (Qiagen, Germantown, MD, USA). Reagents and primers for real time PCR were purchased from Qiagen (Valencia, CA, USA). The protease inhibitor cocktail kit was obtained from Pierce Biotechnology (Rockford, IL, USA). 17β-estradiol (E2) and 4-hydroxytamoxifen (OH-TAM) were purchased from Sigma Aldrich (Saint Louis, MO USA). Olaparib was from Selleckchem (Houston, TX, USA).
Lymphoblastoid cell lines (LCLs) with known genotypes for the chromosome (chr) 4 CTSO SNPs were cultured in RPMI 1640 media containing 15% (vol/vol) FBS (Invitrogen, San Diego, CA). T47D, ZR75-1, CAMA-1, MDA-MB-231 cell lines were obtained from American Type Culture Collection (ATCC) (Manassus, VA). T47D and ZR75-1 were cultured in RPMI-1640 (Grand Island, NY) containing 10% fetal bovine serum (FBS). CAMA-1 cells were cultured in Eagle's Minimum Essential Medium containing FBS to a final concentration of 10%. MDA-MB-231 cells were cultured in Leibovitz's L-15 Medium containing 10% FBS at 37°C without CO2.
Luciferase reporter gene constructs containing various SNP genotypes were generated by PCR based mutagenesis. Specifically, a 1924 bp segment of the CTSO promoter containing ERE was PCR amplified with the primers: 5’- TAAGCAGATATCACTGACATCATGCCACACCT’ and 5- ACGATGCTGAGATTGACCCTAAGCTTTAAGCA -3’ and was cloned into the EcoRV and HindIII sites of pGL3 basic plasmid to make the pGL3-CTSO construct. A 150–250 bp DNA segment that included the rs10030044, rs6810983, rs6835859, rs4550865, rs62328155, rs11737651, and rs4256192 SNPs respectively was also PCR amplified using primers as described in S1 File.
These fragments were cloned into the KpnI and NheI sites upstream of the CTSO promoter sequence to make the plasmids pGL3-WT-CTSO or pGL3-V-CTSO. The WT SNP sequence was amplified with LCL genomic DNA as a template that was homozygous for this WT SNP genotype. This variant SNP sequence was amplified using LCL genomic DNA shown to be homozygous for the variant genotype as template. These 150 -250bp amplicons contained the rs10030044, rs6810983, rs6835859, rs4550865, rs62328155, rs11737651, and rs4256192 SNPs respectively.
T47D and ZR75-1 cells were then seeded in triplicate in 12-well cell culture plates at a concentration of 105 cells / well. After 24 h, the cells were transfected using Lipofectamine 2000 (Invitrogen) with 4 μg of the pGL3-WT-CTSO or pGL-V-CTSO constructs and 2 μg pRL-CMV encoding a CMV-driven renilla luciferase vector (Promega), together with the carrier DNA (pGL3 basic). Luciferase assays were performed 48 h after transfection using a luciferase reporter assay system (Promega). The renilla luciferase activity was used to correct for the transfection efficiency.
The human variation panel model system consists of LCLs from 300 healthy subjects (100 European-Americans, 100 African-Americans, and 100 Han Chinese-Americans). This panel was generated by the Coriell Institute (Camden, New Jersey). We genotyped all 300 cell lines for genome-wide SNPs using Illumina 550K and 510S SNP BeadChips (Illumina), and the Coriell Institute obtained Affymetrix SNP array 6.0 (Affymetrix) data for the same cell lines. These combined SNP genotype data (~1.3 million genotyped SNPs) were used to impute a total of approximately 7 million SNPs per cell line. This LCL model system has been used repeatedly to generate and/or test pharmacogenomic hypotheses arising from clinical GWAS [3, 12, 17–19, 53]. The application of these cell lines made it possible to evaluate the function of CTSO and ZNF423 SNP genotypes. To study the effect of the SNP on CTSO expression, LCLs were cultured in base media containing 5% charcoal-stripped FBS for 24 hours and were subsequently cultured in FBS-free base media containing 0.1 nM E2 for another 48 hours. Cell lysates were used to perform Western blot analysis, and total RNA was isolated for qRT-PCR.
Breast cancer cells were cultured in specific base media, as described above, supplemented with 10% FBS. 5000 cells were seeded in triplicate in 96-well plates, and were cultured in base media containing 5% (vol/vol) charcoal-stripped FBS for 24 hours and were subsequently cultured in FBS-free base media for another 24 hours. Cells were then transfected with either control siRNA or siRNA targeting CTSO. Twenty-four hours after transfection the media was replaced with fresh FBS-free base media and the cells were treated with 0.1 nM E2 for 24 hours, and then treated with 100 nM 4-OH- tamoxifen. Cell growth was measured at different time points (0, 24, 48, and 72 hours) post tamoxifen treatment using the BrdU Cell Proliferation Assay kit (Cell Signaling, Danvers, MA) at intervals of 24 h following the manufacturer's instructions. The plates were measured in a Safire2 microplate reader (Tecan AG, Switzerland).
LCLs selected based on ZNF423 and CTSO genotypes were cultured in RPMI 1640 media (Cellgro) supplemented with 15% FBS. Cells were cultured in RPMI 1640 media containing 5% (vol/vol) charcoal-stripped FBS for 24 hours and were subsequently seeded in triplicate in 96-well plates and cultured in FBS-free RPMI 1640 media for another 24 hours before treatment. Cells were treated with 0.1 nM E2, 50nM tamoxifen, or the combination of both 0.1 nM E2 and 50nM tamoxifen. Cell growth was measured at different time points (0, 24, 48, 72, and 96 hours) post treatment using the CYQUANT Direct Cell Proliferation Assay (#C35012, Invitrogen) following the manufacturer’s instructions at intervals of 24 h. The plates were measured in a Safire2 microplate reader (Tecan AG, Switzerland).
Cells were plated at 70% confluence in culture medium supplemented with 10% FBS, and were transfected with empty vector or CTSO plasmid (OriGene) using lipofectamine 2000 (Invitrogen, Carlsbad, CA) according to the vendor's protocol. Cells were collected for protein analysis 48 hours after transfection. In some experiments, 24 hours after transfection, cells were treated with 10 μM E-64, a cysteine proteases inhibitor, for additional 24 hours. Cells were then collected for protein analysis.
Specific siGENOME siRNA SMARTpool reagents against a given gene as well as a negative control, siGENOME Non-Targeting siRNA, were purchased from Dharmacon Inc. (Lafayette, CO, USA). Cells were transfected with control siRNA, and specific siRNAs (10nM) in 96-well plates or 12-well plates using lipofectamine RNAiMAX (Invitrogen, Carlsbad, CA) according to the vendor's protocol. For the purpose of cell growth assay, cells were plated in base medium supplemented with 5% charcoal stripped FBS for 24 hours, and then cultured in FBS-free RPMI 1640 media for another 24 hours before transfection. Different treatments were started 24 hours after transfection. For the purpose of testing gene expression level, cells were transfected with control siRNA and specific siRNAs (10nM) in 12-well plates using lipofectamine RNAiMAX for 48 hours.
Breast cancer cells were harvested by trypsinization, lysed in SDS buffer. Cell lysates were heated to 95°C for five minutes. Protein samples (10 to 20 μg) were resolved by electrophoresis on 10% sodium dodecyl sulfate polyacrylamide gel electrophoresis (SDS-PAGE) gels and electrophoretically transferred to PVDF membranes (Millipore Corporation, Bedford, MA, USA). The blots were probed with the appropriate primary antibody and the appropriate horseradish peroxidase conjugated secondary antibody. The protein bands detected with the Pierce enhanced chemiluminescence Western blotting substrate (Thermo Scientific, Rockford, IL, USA) and were visualized using Geldoc (Bio-Rad Laboratories).
LCLs selected based on ZNF423 and CTSO genotypes were cultured in RPMI 1640 media containing 5% (vol/vol) charcoal-stripped FBS for 24 hours and were subsequently seeded in 6-well plates and cultured in FBS-free RPMI 1640 media for another 24 hours before treatment. Cells were treated with 0.1 nM E2, 50nM tamoxifen, or combination of both 0.1 nM E2 and 50nM tamoxifen for 48 hours and lysed in RIPA buffer supplemented with protease and phosphatase inhibitors. Cell lysates were used to perform Western blot analysis. Quantification of the blots was analyzed using Image J.
Cells were lysed in NETN buffer (20 mM Tris-HCl, pH 8.0, 100 mM NaCl, 1 mM EDTA, 0.5% Nonidet P-40) supplemented with protease and phosphatase inhibitors. Lysates were clarified by centrifugation (13,000 r.p.m., 20 min, 4°C) and 500 μg–1mg proteins were used per immunoprecipitation. Proteins were captured with 2 μg CTSO antibody and protein G-sepharose Fast-Flow (Sigma). Immunoprecipitation with mouse serum was used as negative controls. The immuno-complexes were then washed with NETN buffer three times followed by separation on SDS-PAGE. Proteins were resolved by SDS–PAGE, transferred onto PVDF membranes and probed using the appropriate primary and secondary antibodies coupled to horse-radish peroxidase.
Total RNA was isolated from cultured cells with the QIAGEN RNeasy kit (QIAGEN Inc., Valencia, CA, USA), followed by qRT-PCR performed with the one-step Brilliant SYBR Green qRT-PCR master mix kit (Stratagene, La Jolla, CA, USA). Specifically, primers purchased from QIAGEN were used to perform qRT-PCR with the Stratagene Mx3005P real-time PCR detection system (Stratagene). All experiments were performed in triplicate with GAPDH as an internal control. Reverse-transcribed Universal Human Reference RNA (Stratagene) was used to generate a standard curve. Control reactions lacked the RNA template. The 2-δδcycle threshold method was used for statistical data analysis.
Drugs were dissolved in DMSO, and aliquots of stock solutions were frozen at −80°C. Cytotoxicity assays were performed in triplicate at each drug concentration. Specifically, 4000 breast cancer cells were seeded in 96-well plates and were cultured in base media containing 5% (vol/vol) charcoal-stripped FBS for 24 hours and were subsequently cultured in FBS-free base media for another 24 hours. Cells were then transfected with either control siRNA or siRNA targeting CTSO. Twenty-four hours after transfection the media was replaced with fresh FBS-free base media and the cells were treated with 10 μL of tamoxifen at final concentrations of 0, 0.5, 1, 2, 4, 6, 8, 12, 24, and 48 μM with or without 10 μM olaparib. After incubation for an additional 72 hours, cytotoxicity was determined by quantification of DNA content using CYQUANT assay (#C35012, Invitrogen) following the manufacturer’s instructions. 100μL of CyQUANT assay solution was added, and plates were incubated at 37°C for one hour, and then read in a Safire2 plate reader with filters appropriate for 480 nm excitation and 520 nm emission.
LCLs selected based on ZNF423 and CTSO genotypes were cultured in RPMI 1640 media containing 5% charcoal-stripped FBS for 24 hours and 5x104 cells were subsequently seeded in triplicate in 96-well plates and cultured in FBS-free RPMI 1640 media for another 24 hours before treatment. Cells were treated with 10 μL of tamoxifen at final concentrations of 0, 0.5, 1, 2, 4, 6, 8, 12, 24, and 48 μM with or without 5 μM olaparib. After incubation for an additional 72 hours, cytotoxicity was determined by quantification of DNA content using CYQUANT assay.
ZR75-1 cells were transfected with CTSO plasmid. After 72 hr, cells were lysed by NETN buffer. Cell lysates were incubated with control IgG or CTSO antibody at 4°C for 4 hr, and then incubated with protein G-sepharose Fast-Flow for 2 hr. After washing with NETN buffer three times, bound proteins were eluted, and size fractionated by 10% SDS-PAGE. Coomassie-stained gel slices covering the entire molecular weight range were processed for analysis by mass spectrometer following a standard protocol at the Harvard Medical School Taplin Mass Spectrometry Facility.
All data were presented as mean ± SD of at least three independent experiments. Statistical analysis was performed using SPSS22.0 and Prism 5 (GraphPad Software Inc., San Diego, CA, USA). Single-factor analysis of the variance test was used for comparisons among multiple groups, and a t-test was used for comparisons between two groups; P <0.05 was considered statistically significant.
|
10.1371/journal.pntd.0005647 | The socio-economic burden of snakebite in Sri Lanka | Snakebite is a major problem affecting the rural poor in many of the poorest countries in the tropics. However, the scale of the socio-economic burden has rarely been studied. We undertook a comprehensive assessment of the burden in Sri Lanka.
Data from a representative nation-wide community based household survey were used to estimate the number of bites and deaths nationally, and household and out of pocket costs were derived from household questionnaires. Health system costs were obtained from hospital cost accounting systems and estimates of antivenom usage. DALYs lost to snakebite were estimated using standard approaches using disability weights for poisoning.
79% of victims suffered economic loss following a snakebite with a median out of pocket expenditure of $11.82 (IQR 2–28.57) and a median estimated loss of income of $28.57 and $33.21 for those in employment or self-employment, respectively. Family members also lost income to help care for patients. Estimated health system costs for Sri Lanka were $ 10,260,652 annually. The annual estimated total number of DALYS was 11,101 to 15,076 per year for envenoming following snakebite.
Snakebite places a considerable economic burden on the households of victims in Sri Lanka, despite a health system which is accessible and free at the point of care. The disability burden is also considerable, similar to that of meningitis or dengue, although the relatively low case fatality rate and limited physical sequelae following bites by Sri Lankan snakes means that this burden may be less than in countries on the African continent.
| Snakebite predominantly affects poor people in the rural tropics. The effect that snakebite has on these populations, both economically and in terms of death and disability, is poorly understood. We used data from a national household survey of snakebite in Sri Lanka to estimate the burden of death and disability and to calculate the financial cost of a snakebite episode for the Sri Lankan health system and for Sri Lankan households. We found that the burden of snakebite was considerable, similar to that of common diseases like meningitis or dengue and that treating snakebite cost the Sri Lankan government over $10 million each year. Despite health care being free in Sri Lanka, almost 80% of households experienced additional costs and loss of income following a snakebite; such costs are disastrous for poor rural workers.
| Snakebite is a major public health problem in rural communities in Asia, Africa and Latin America [1]. The problem has been extensively studied from a bio-medical perspective but rarely from a socio-economic viewpoint. However, the negative social and economic consequences of snakebite are likely to be considerable. Firstly, snakebite is a problem of tropical low and middle income countries that are already facing the considerable dual burden of communicable and non-communicable diseases; snakebite mortality is strongly associated with low per capita Gross Domestic Product and a low Human Development Index [2]. In such settings, the ill-health associated with snakebite can exert a considerable burden on the economic development of these countries. Secondly, the victims are usually economically productive, young individuals in these communities whose future productive lifespan can be negatively affected by snakebite. Thirdly, most of those affected are rural daily wage earners employed in the informal sector that include farming and other labour intensive occupations, for whom snakebite results in a considerable opportunity cost for being away from work.
Very few studies have attempted to formally investigate the socio-economic burden associated with snakebite. A household survey in rural India demonstrated a significant reduction in medium and long-term family income due to snakebite, in addition to the immediate costs of a bite [3]. A West African study estimated years of life lived with disability (YLD) due to limb amputations resulting from snakebite using DALYs for 16 countries in the region [4]. Disability-adjusted-life years (DALYs) is a widely used metric for estimating disease burden based on strong economic and ethical principles [5]. The Global Burden of Disease (GBD) Study used DALYs successfully to describe disability and the burden of important diseases for the period 1990–2020 [6]. Since then, the method of GBD estimation has been improved in repeated attempts to estimate the global disease burden [7]. National estimates of the overall social and economic impact of snakebite have not been attempted for any country to date. This paper estimates the economic cost and disease burden of snakebite for Sri Lanka using data from a recent country-wide community based survey [8].
The study was approved by the Ethics Review Committee of the Faculty of Medicine, University of Kelaniya. Permission for conducting the study was obtained from District and Divisional level public administrators before data collection. Grama Niladharis of the sampled GN divisions were informed about the study through the public administration system. Written informed consent was obtained from the participants before data collection. The information sheet was read out to illiterate patients in the presence of a family member and a witnessed thumb print was used to signify consent.
A country-wide community based cross sectional survey was conducted between August 2012 and June 2013 [8]. The survey was designed to sample approximately 1% of the population of Sri Lanka distributed equally among its nine provinces. A Grama Niladhari (GN) division (the smallest administrative unit in the country) was defined as a cluster for data collection, and 125 clusters were allocated to each of the 9 provinces. Within each province the number of clusters was divided among the districts in proportion to each districts’ population. The clusters were selected using simple random sampling from the list of GN divisions available at the Department of Census and Statistics, Sri Lanka. 40 households were sampled consecutively from the randomly selected starting point in each cluster.
Information related to all residents of the sampled households was obtained by trained data collectors using an interviewer administered questionnaire. They were assisted by local field volunteers recruited from within the cluster. The respondent of each household was either the head of the household or a responsible adult present in the house. A two part structured questionnaire was used for data collection. The questionnaire was translated into Sinhala and Tamil and was pre-tested in a GN division within each province which was not selected for the study. Based on the findings of the pretesting, the questionnaire was fine-tuned prior to use.
In the first phase of data collection, the research assistant screened the households for snakebite within the previous 12 months and obtained socio-demographic data from the households. In the second phase of data collection, instruments were administered to the households which had reported snakebites within the previous 12 months in order to obtain details of the bite, clinical manifestations of envenoming, residual disability and deaths due to snakebite. Detailed information on the household costs of snakebite was recorded. Only systemic symptoms or signs were considered to reflect envenoming.
Data were double entered into databases created in Epidata software. Discrepancies were corrected by referring to the original data sheets. Data analysis was performed in SPSS version 22.
The median out-of-pocket cost of different cost elements were estimated based on the data reported by the victims or a household member for a number of different out of pocket costs. Victims or household members were also asked to estimate the number of days lost off work due to the snakebite and the amount of wages lost by the victim and the household due to the snakebite. The total sum spent by patients for a particular cost item and the proportion of patients that incurred that cost was applied to the estimated national incidence of snakebite to estimate the total annual out-of-pocket cost of snakebite for the entire country. The income lost by the patient or family members and the proportion of patients who lost income in different ways were applied to the estimated national incidence of snakebite to arrive at the total annual lost income due to snakebite for the entire country. The health system cost of snakebite was estimated based on cost data obtained from the cost accounting system maintained at the Teaching Hospital, Kurunegala, Sri Lanka. The average cost of a patient day in the medical ward excluding drug costs was obtained from this database. This cost included the cost of nursing and medical care, investigations and the hotel costs of maintaining a patient in a medical ward. The cost of a patient day amounted to LKR 3214.00 (USD 22.96). Ancillary treatments are rarely used in Sri Lanka and were therefore not included. The cost of a vial of anti-venom was obtained from the price list issued by the Medical Supplies Division, Ministry of Health, Sri Lanka [9]. The number of anti-venom vials used was estimated based on the national guidelines for management of snakebite and the assumption that the 30%, 65% and 5% of envenomed patients received respectively, 10, 20 and 30 vials of anti-venom during the management of envenoming. This assumption was based on the consensus arrived among five specialist physicians experienced in managing snakebite in different parts of the country. The median duration of hospitalization for a snakebite with and without envenoming was estimated based on the data reported by the households of victims. The costs were extrapolated for the country using the estimated national incidence of snakebite from the nationwide survey.
DALYs were calculated using the following formula: DALY = YLD + YLL. The template developed by the World Health Organization [10] was used for estimation of DALYs. Population data from the 2012 national census were obtained from the Department of Census and Statistics, Sri Lanka ([11]. For envenoming, the disability weight used for the higher estimate was 0.6 which is the accepted disability weight used for poisoning in the original GBD study [6]. We used a disability weight of 0.163 for the lower estimate based on the disability weight used for poisoning in the 2013 GBD study. In using the disability weight for poisoning, we have assumed that snakebite envenoming and poisoning are comparable in terms of the associated disability. The duration of an episode of snakebite with envenoming was considered to be 0.3 years. For snakebites without envenoming disability weights of 0.006 (lower estimate) and 0.108 (higher estimate) were used. These are the weights used for open wounds in the GBD studies [6,12]. The duration of illness for snakebite without envenoming was considered to be 0.04 years. The standard discount rate used was 0.03. Beta and constant values used for standard age weighting were 0.04 and 0.1658.
The total sample was 165,665 individuals living in 44,136 households approximately equally distributed among the nine provinces of the country. The survey reported 695 snakebites and five deaths. Envenoming had been observed in 323 bite victims. Residual physical disability following the bite was reported by 59 (8.5%) victims. For the entire country, the extrapolated number of snakebites, envenoming and deaths were 80,277, 30,458, and 429, respectively [8].
551 (79.3%) victims incurred an economic loss following a bite and 550 (79.1%) victims incurred out-of-pocket expenditure for healthcare. The median total out-of-pocket expenditure per snakebite episode (envenomed and non-envenomed) was USD 11.82 (Inter-quartile range 5–28.57). Details of out-of-pocket expenditure are given in Table 1 and include travel, food, costs of keeping carers with the victim during the hospital stay, fees for laboratory investigation, purchase of pharmaceuticals and medical products that were not available in the hospital and other unspecified direct costs. In addition, a cost was often incurred for religious and cultural rituals which were organized for 138 (19.9%) victims necessitating expenditure for 101 (14.5%) families. The median cost of conducting these activities was USD 7.14 (Inter quartile range 3.57–14.29). The annual estimated national direct out-of–pocket expenditure for snakebite was USD 1,981,699.
442 of the victims (63.6%) were employed, but only 134 (19.3%) reported that they had to stop work temporarily due to the bite. The median total income lost due to the bite by these victims was USD 28.57 (Inter-quartile range 17.14–56.07) (Table 2). Self-employed victims (n = 158, 22.7%) lost a median income of USD 33.21 (Inter-quartile range 17.86–54.46) either due to lost work or costs of a replacement. National annual estimated lost income was USD 910,259 and USD 844,142 for employed and self-employed victims respectively. At least one family member of 103 (14.8%) bite victims had lost at least one workday due to the bite. The median economic loss by these family members due to the bite was USD 28.57 (Inter-quartile range 14.29–81.07). This amounted to an annual estimated lost income of USD 101,037 for family members of victims. Overall, the estimated annual lost income amounted to a total of USD 1,855,438, meaning that the total annual economic burden on households nationally was USD 3,837,137.
The estimated annual health system cost of snakebite management was USD 10,260,651.53 (Table 3). This comprised approximately USD 6.3 million for the snakebite anti-venom and USD 3.9 million for the hospital management of patients (Table 2). Combining household and health system costs meant that the estimated total annual economic burden of snakebite was USD 14,097,789.
The total YLL due to snakebite was 4,765 for males and 4,853 for females. Total YLD ranged from 859–3,161 for males and 624–2,296 for females. The annual estimated total number of DALYs for envenoming and death due to snakebite ranged from a lower estimate of 11, 101 to an upper estimate of 15,076 per year. This comprised 5,624–7,927 DALYs for males and 5477–7,150 DALYs for females, equating to 0.5–0.7 DALYs per 1000 population for snakebite envenoming (Tables 4 and 5). The total estimated DALYs due to snakebites without envenoming ranged from 20–500 per year (S1 and S2 Tables).
Snakebite is a major neglected tropical disease. The relative lack of data on the burden of snakebite demonstrates the lack of attention to this condition which is confined to poor rural areas of tropical, low and middle income countries [2]. We have estimated the societal and economic burden of snakebite using nationally representative data generated from a large scale community survey conducted in Sri Lanka. We estimate that snakebite is responsible for up to approximately 15,000 DALYs per year. This is three times greater than the estimate from the Institute for Health Metrics and Evaluation for Sri Lanka and would suggest that the burden of snakebite in Sri Lanka is equivalent to that of meningitis and greater than that of dengue [13].
Given the numbers of snakebites in Sri Lanka, the estimated DALYs due to snakebite in this paper are relatively modest compared to those estimated from West Africa [4]. The availability of antivenom and supportive facilities in an accessible and free health system has led to a reduction in snakebite mortality over the last two decades in Sri Lanka to the point where case fatality rates are now only 1.5% in envenomed patients [8]. This situation contrasts with the limited access to healthcare facilities and poor availability of antivenom in many parts of Africa, where snakebite occurs in the poorest rural populations [2]. The current crisis in antivenom supply for Africa means that many patients die because they simply cannot be treated [14]. In addition, a major contributor to the burden calculation for West Africa was the disability that results from severe tissue damage and consequent amputation, an outcome that is uncommon following envenoming by venomous species in Sri Lanka [4,15].
High quality facilities and emergency care coupled with adequate supplies of anti-venom lead to the good outcomes following snakebite in Sri Lanka, but means the estimated health system costs of managing snakebite are considerable at over USD 10 million and account for 0.7% of total government health expenditure [16]. It is unlikely that these costs will reduce in the near future as there is no indication that the high incidence of bites is declining and improving health seeking behaviour means that western standard healthcare facilities are increasingly likely to be used.
Even more concerning is the economic burden that snakebite places on victims and their households. Few studies have previously attempted to estimate the effect of snakebite on this although one Tamil Nadu study demonstrated the ongoing adverse household consequences of expenditure on snakebite, requiring loans and selling of household assets to pay for treatment costs and the economic burden of snakebite upon households has also been noted in Bangladesh [3,17]. Our research demonstrated the substantial household costs of an episode of snakebite from both out of pocket costs and lost income. Some workers in the formal employment may be compensated by paid sick-leave, but only a few victims of snakebite have formal employment. The median household cost of an episode of envenoming was $12 and around a fifth of patients reported losing income of approximately $30. To put this in context, the annual per capita expenditure on health in 2014 was $127 [18] and the mean per capita income in the rural areas was $74 per month [19]. The national household economic burden of snakebite amounted to USD 3.8 million and it is highly likely in Sri Lanka that snakebite drives the same catastrophic costs for the poor as many other diseases [19,20].
There are clearly limitations to this study. Many of the estimates depend on assumptions about the duration of disability and recall of individuals about the costs that they incurred. There were challenges in the estimation of disability burden as there are no accepted disability weights for snakebite and so those for poisoning were used for envenomed individuals and assumptions regarding wounds were made for non-envenomed. Despite this, to our knowledge, this is the first ever comprehensive estimation of a national socio-economic burden from snakebite. Our results demonstrate the extent of this burden in Sri Lanka and highlights the considerable physical and economic impact of this disease, both upon the country and upon the lives of poor rural workers.
|
10.1371/journal.pgen.1005683 | The Sex Determination Gene transformer Regulates Male-Female Differences in Drosophila Body Size | Almost all animals show sex differences in body size. For example, in Drosophila, females are larger than males. Although Drosophila is widely used as a model to study growth, the mechanisms underlying this male-female difference in size remain unclear. Here, we describe a novel role for the sex determination gene transformer (tra) in promoting female body growth. Normally, Tra is expressed only in females. We find that loss of Tra in female larvae decreases body size, while ectopic Tra expression in males increases body size. Although we find that Tra exerts autonomous effects on cell size, we also discovered that Tra expression in the fat body augments female body size in a non cell-autonomous manner. These effects of Tra do not require its only known targets doublesex and fruitless. Instead, Tra expression in the female fat body promotes growth by stimulating the secretion of insulin-like peptides from insulin producing cells in the brain. Our data suggest a model of sex-specific growth in which body size is regulated by a previously unrecognized branch of the sex determination pathway, and identify Tra as a novel link between sex and the conserved insulin signaling pathway.
| Female-biased sexual size dimorphism is common in invertebrates, yet the mechanisms underlying increased female body size remain unclear. We uncovered a key role for sex determination gene transformer (tra) in promoting increased growth in females. Interestingly, we found that sex differences in body size are regulated by Tra in a pathway that is separate of the canonical sex determination pathway, and of other aspects of sexual dimorphism. Instead, Tra function in the fat body regulates growth in a non cell-autonomous manner by regulating the secretion of insulin-like peptides from the brain. This novel Tra-insulin link we describe may have implications for other sexually dimorphic phenotypes in Drosophila (eg. lifespan, stress resistance), many of which are also regulated by insulin.
| Drosophila is a well-established model to study the mechanisms that control animal growth [1, 2]. Drosophila body size is determined by various developmental cues that coordinate tissue patterning with growth, and by environmental cues such as nutrients and oxygen, that regulate whole body metabolism. One important, but often overlooked, determinant of size in Drosophila is sex–adult females are significantly, and visibly, larger than males [3, 4]. This sexual size dimorphism (SSD) arises due to differences in larval growth: males and females have a similar overall duration of larval development, but females achieve critical weight at a larger size and grow more during the terminal growth period [5]. While over two decades of genetic research have identified many conserved signaling pathways that link developmental and environmental cues to the control of tissue and body size [6–9], the genetic and physiological mechanisms that account for the larger female body size remain unclear.
In flies, sex is determined by the ratio of sex chromosomes to autosomes (X:A) [10]. In females, the X:A ratio is 1, and a functional protein is produced from the Sex-lethal (Sxl) locus [11, 12]. In males, the X:A ratio is 0.5, and no Sxl is produced. Sxl is a master regulator of female sexual development (eg. sexual differentiation, reproduction), and Sxl mutant females are smaller than wild-type females. This is due, in large part, to the sex-specific splicing of its downstream target gene transformer (tra) [13–16]. As a result of this Sxl-dependent splicing, a functional Tra protein is produced in females, but not males. Tra is a splicing factor, and has only two known direct targets: doublesex (dsx) and fruitless (fru) [17–22]. While Tra is thought to mediate most of Sxl’s effects on sex determination, the control of sex differences in body size is thought to be independent of the Tra/Dsx/Fru branch of the sex determination pathway [23]. Here, we identify for the first time a role for Tra as a key regulator of SSD in Drosophila. Further, we show that Tra’s effects on SSD are mediated by a novel pathway that is independent of dsx and fru, and of other aspects of sexual dimorphism.
Female and male Drosophila larvae show no difference in their rate of development [5]. However, by the end of larval life, female body size is approximately 30% larger than male body size (Fig 1A and 1B). These differences are not due to sex differences in food intake or feeding behaviour (S1A and S1B Fig). Although the prevailing view is that tra does not regulate sex differences in body size [24], one study showed that adult weight in tra mutant females was reduced compared to wild-type females [25]. However, this weight reduction can be explained by the lack of ovaries in tra mutant females. We therefore tested whether the decreased weight was due to an effect of tra on growth by measuring pupal volume in tra mutant animals. We found that body size was significantly reduced in tra mutant females compared to wild-type females (Fig 1C and 1D). Thus while wild-type females are 30% larger in body size than males, tra mutant females are only 10% larger than males. This suggests that tra contributes to establishing SSD in Drosophila. Body size was unchanged in tra mutant males, consistent with the lack of a functional Tra protein in males (Fig 1D). We also performed loss-of-function experiments with tra, using an RNAi transgene directed against tra’s splicing co-factor tra2 (UAS-tra2-RNAi). We found that ubiquitous expression of the UAS-tra2-RNAi transgene using the Act5c-GAL4 driver successfully transformed female animals into phenotypic males (S1C Fig), and led to a reduction in body size (S1D Fig).
We next examined whether lack of Tra expression in males could explain their smaller body size. Ubiquitous expression of a UAS-tra transgene using the daughterless (da)-GAL4 driver led to a significant increase in body size in males (Fig 1E and 1F). Interestingly, overexpression of Tra also stimulated growth in females, showing Tra has growth-promoting effects in both sexes. While previous studies have shown that high levels of Tra expression can cause artifacts such as lethality [26], this is the first report of an alternative splicing factor promoting body growth in Drosophila. To further confirm these Tra-dependent changes in body size, we measured adult weight. In order to ensure that tra’s effects on adult weight are not confounded by its effects on ovary or testis development, we weighed 5-day-old animals from which the gonads were removed by dissection. As with pupal volume, we found that tra mutant females, but not tra mutant males, were significantly smaller than controls (S2A Fig). Overexpression of UAS-tra caused an increase in body weight in males (S2B Fig). Together, these results suggest that male-female differences in body size are created in part by the presence of Tra in females, and the absence of Tra in males. This defines a new role for Tra in the regulation of sex differences in body growth.
As with body size, female cell size in the wing [27], and the fat body (Fig 2A) are larger. Since one previous study showed that wing cell size in tra mutant females was intermediate in size between female and male cells [25], we tested whether Tra expression could mediate cell-autonomous effects on growth by expressing either UAS-tra or the UAS-tra2-RNAi transgenes in the fat body (polyploid cells) and the wing disc cells (mitotic cells) of developing larvae. We found that flp-out-mediated mosaic expression of UAS-tra2-RNAi in female fat body caused a significant reduction in cell size (Fig 2B). Consistent with the lack of a functional Tra protein in males, similar UAS-tra2-RNAi expression in males did not affect cell size (Fig 2B). In contrast, overexpression of UAS-tra in fat body cells was sufficient to increase cell size in both sexes (Fig 2C). We next examined whether Tra expression could affect sex differences in another larval tissue, the wing disc. Using engrailed (en)-GAL4, we expressed either UAS-tra2-RNAi or UAS-tra transgenes in the posterior compartment of the wing, and measured compartment size. We found a significant reduction in compartment size in female wings when we knocked down Tra function by UAS-tra2-RNAi expression (Fig 2D). This effect was also seen when we used a second UAS-tra2-RNAi transgene (S3A Fig). In contrast, male compartment size was unaffected (Fig 2D). To determine whether this reduction in compartment size was due to a decrease in cell number or cell size, we counted wing hairs in a fixed area in the posterior compartment in females. Each wing cell secretes one hair, thus by counting wing hairs we can accurately determine how many cells are present in a specific area. We found that the number of cells in the counting area was significantly increased in the compartments expressing UAS-tra2-RNAi compared to controls (Fig 2E). This suggests that the reduction in compartment size is due to a reduction in cell size, rather than cell number. Indeed, the estimated cell number in the posterior compartment of the wing was not significantly altered by expression of UAS-tra2-RNAi (Fig 2F). Overexpression of UAS-tra using en-GAL4 caused no significant increase in compartment size in either males or females (S3B Fig). Together, these results demonstrate a cell-autonomous requirement for Tra in females to promote increased female cell size in both mitotic and endoreplicating cells. This result supports previous findings from early gynandromorph studies, where sex differences in cell size were regulated in a cell-autonomous manner [28]. More recently, Sxl expression in the wing disc was shown to promote growth [29]; however, altering either the X:A ratio or Sxl expression also affects the process of dosage compensation. Since Tra does not affect this process [30], our results demonstrate that sex differences in cell size can be uncoupled from dosage compensation.
An emerging literature in Drosophila has highlighted the importance of noncell-autonomous signaling in the control of body growth [31–34]. This signaling relies on organ-to-organ endocrine communication and is particularly important in controlling body growth in response to dietary nutrients. We therefore tested whether sex differences in overall body size may also involve non-autonomous effects of Tra function specific tissues. We used a number of tissue-specific GAL4 drivers to express the UAS-tra2-RNAi transgene in larvae, and measured pupal volume in these animals. We found that loss of Tra in the fat body caused a significant reduction in female body size (Fig 2G). This reduction in body size was also observed in females expressing a second UAS-tra2-RNAi transgene in the fat body (S3C Fig). Expression of UAS-tra2-RNAi in neurons, glia, ring gland or in muscle did not reproduce the female-specific effects of the fat body (S4A Fig), and male body size was unaffected by expression of the UAS-tra2-RNAi transgene in any tissue (S4B Fig). We next asked whether expression of UAS-tra in specific tissues was sufficient to drive an increase in body size. Tissue-specific expression of UAS-tra using a panel of GAL4 drivers did not significantly affect body size in wild-type females or males (S5A and S5B Fig). We then asked whether tissue-specific expression of Tra could rescue the body size defects of tra mutant females. We found that ubiquitous, or fat-specific, expression of Tra was sufficient to rescue a normal body size to tra mutant females (Fig 2H). Together, these results suggest that the decreased body size in tra mutant females is due to fat-specific loss of Tra function. Tra controls many aspects of sexual differentiation, including gonad and germline differentiation, and previous studies in C. elegans showed that the germline can influence body growth [35]. We therefore tested whether Tra expression in these tissues could explain its effects on female body size. However, we found that gonad- or germline-specific expression of a UAS-tra2-RNAi transgene using the c587-GAL4 or nanos (nos)-GAL4 drivers, respectively, caused no significant reduction in body size in females (S6A and S6B Fig). Similarly, body size was unaffected in males and females completely lacking a germline (the progeny of tudor1 homozygous mutant females crossed to wild-type males; S6C Fig). Decreased body size in tra mutant females is therefore not due to the presence of a male gonad or germline. Instead, our results suggest the sex of the fat body, as determined by Tra expression, controls body growth in a non cell-autonomous manner.
Tra is a splicing factor, and has only two known direct targets: doublesex (dsx) and fruitless (fru) [17–22]. Dsx is expressed in a handful of tissues throughout the body and in a restricted expression pattern in the central nervous system (CNS) in both males and females [36–40]. In females, Tra binding to dsx pre-mRNA causes a female-specific Dsx isoform to be produced (DsxF). In males, which express no functional Tra protein, a default splice in dsx pre-mRNA generates a male-specific isoform of Dsx (DsxM) [19, 20]. Tra binding to the pre-mRNA of transcripts from the fru P1 promoter causes the introduction of a stop codon, and no Fru P1 protein is expressed in females. In males, the lack of Tra leads to the use of a default splice in fru P1 transcripts, generating a male-specific Fru P1 protein (FruM) [17, 18]. FruM expression is limited to males in approximately 2000 neurons in the CNS and peripheral nervous system (PNS) [18, 41]. Importantly, dsx and fru are thought to mediate most, if not all, effects of Tra on sex determination and behaviour [42, 43]. We therefore tested whether either gene was required for Tra’s effects on growth. We first examined whether mutants lacking Dsx or FruM expression phenocopied any of Tra’s effects on growth. We found that dsx mutant animals (genotype dsx1/Df(3R)dsx15), had no significant difference in pupal volume compared to controls in either males or females (Fig 3A). We also examined the effect of Dsx knockdown in the fat body using a UAS-dsx RNAi line. Using the flp-out system, we found that mosaic expression of UAS-dsx RNAi led to reduced fat cell size in both males and females, suggesting that dsx regulates cell size in this tissue (Fig 3B). However, expression of UAS-dsx-RNAi throughout the fat body using r4-GAL4 did not recapitulate the non cell-autonomous reduction of body size observed upon Tra inhibition in this tissue (Fig 3C). While it seems counterintuitive that Dsx causes a reduction in size in the fat body, but does not affect body size, dsx expression is restricted to specific tissues in larvae [37, 39]. Thus while loss of dsx may affect cell size in a relatively small number of tissues, the expression may not be broad enough to cause a reduction in overall body size. Also, in the context of our findings with tra, loss of dsx throughout the fat body does not phenocopy the non cell-autonomous effects of loss of Tra on body size. Similar to dsx, we found that males lacking FruM expression, or females ectopically expressing FruM proteins [44], showed no difference in body size compared to controls (Fig 3D and 3E). We next asked whether dsx was required for Tra-induced growth. Using da-GAL4 to overexpress Tra in a dsx mutant background, we found that Tra’s ability to drive body growth was unaffected by loss of dsx (Fig 3F). Our results suggest Tra controls growth in a pathway that is independent of its effects on sexual differentiation and behaviour.
In Drosophila, the conserved insulin/insulin-like growth factor signaling (IIS) and Target-of-Rapamycin (TOR) pathways are two main regulators of tissue and body growth [7, 45, 46]. Both pathways play a central role in linking dietary nutrients to regulation of larval metabolism and growth [34, 47–49]. We therefore tested whether IIS/TOR also plays a role in creating sex differences in body size. We first measured pupal volume in males and females grown in either nutrient-rich food (which promotes high levels of IIS/TOR signaling), or in food with reduced nutrition (which inhibits IIS/TOR signaling). We found that sex differences in body size were abolished in low nutrient conditions (Fig 4A). Since previous studies have shown that IIS/TOR can act in separate pathways to activate downstream effectors [48, 50], we wanted to specifically inhibit the TOR pathway, and examine the effects on body size. When we grew larvae on food containing rapamycin, a specific TOR inhibitor, we found an overall reduction in body size in both sexes; however, the SSD between males and female remained at 25% (S7A Fig). Thus IIS, but not TOR, is required for male-female body size differences. This finding is consistent with a recent study that showed sex differences in adult body weight were eliminated in animals heterozygous for two hypomorphic mutations in the insulin receptor (InR) gene [5].
We next examined whether we could detect any sex differences in IIS activity during development. The serine/threonine kinase Akt is phosphorylated and activated downstream of IIS. Measuring the ratio of the phosphorylated active form of Akt (P-Akt) to total Akt therefore provides a read-out of IIS activity. When we compared male and female larvae collected 96 hr and 120 hr after laying (AEL) at 25°C, we found that females had a significantly higher ratio of P-Akt:Akt at 120 hr AEL (S7B and S7C Fig). To further confirm higher IIS activity in females during development, we used an antibody staining in the larval fat body to detect the subcellular localization of the transcription factor FOXO. When IIS activity is high, FOXO is phosphorylated by P-Akt, and is evenly distributed throughout the cytoplasm and nucleus. When IIS is inhibited, FOXO re-localizes to the nucleus, where it regulates expression of its target genes [51]. We found the nuclear:cytoplasmic ratio of FOXO was significantly higher in males than in females, further suggesting that males have lower levels of IIS activity (S7D and S7E Fig). We found no significant male-female differences in mRNA levels of previously described foxo targets such as InR, 4E-BP, or dilp6 (S7F Fig) [52–54]. This may be due to additional FOXO-independent factors required for their expression [55,56]. While the differences in IIS we report here are not as dramatic as seen with genetic or starvation-mediated perturbation of IIS, they are consistent with females having a modest increase in IIS activity compared to males.
To understand how females achieve a modest increase in IIS compared to males, we examined insulin-like peptide (ILP) expression in larval insulin-producing cells (IPCs). The IPCs express three Drosophila ILPs (dilps 2,3 and 5) [57, 58]. Nutrients have been shown to regulate both mRNA transcription and secretion of these dilps [33, 57, 58]. Thus in response to amino acid input to the fat body, an as-yet-unidentified secreted factor is released that acts upon the IPCs in the brain to trigger dILP2 and dILP5 release into the larval hemolymph. These dILPs bind to the insulin receptor on target cells to activate IIS and promote body growth. In contrast, when nutrient abundance is low, the fat-to-brain signal is reduced and secretion of dILPs is inhibited, leading to decreased systemic IIS and body growth. Given our finding that Tra function in the fat body is required for normal growth in females, we examined whether Tra expression in the fat body influences brain dILPs.
We first examined dILP transcript levels and release in wild-type males and females. Using qRT-PCR, we found that only dilp3 transcript levels were different between the sexes, where males had a significant increase in dilp3 compared to females (Fig 4B). We next wanted to determine whether we could detect any differences in dILP secretion. This can be assayed by immunostaining for dILP2 expression in the IPCs. When dILP secretion is high, dILP2 levels seen in IPC are low. Conversely, when secretion is decreased, dILP2 levels in the IPCs are higher [33]. When we compared males and females, we found that male IPCs had a significantly higher average pixel intensity with anti-dILP2 (Fig 4C). Since dilp2 transcript levels are not different between males and females (Fig 4B), this result suggests that dILP2 secretion is higher in females than males.
Given the importance of the fat body as a regulator of IPC dILP release, we tested whether male-female differences in dILP2 secretion occur as a result of Tra expression in the female fat body. Using r4-GAL4, we expressed the UAS-tra2-RNAi transgene to inhibit Tra function specifically in the fat body, and measured dilp transcript levels (in the larval carcass that was devoid of fat body), or ILP2 staining intensity in the IPCs. We found that loss of Tra in the fat body did not affect transcript levels of dilp2, dilp3 or dilp5 in either males or females (S7G Fig). However, the average pixel intensity of dILP2 staining in the IPCs was significantly higher in r4>tra2-RNAi females compared to control females (Fig 4D and 4E). Male dILP2 levels were unchanged by loss of Tra function in the fat (Fig 4E). These results suggest that Tra expression in the female fat body can enhance levels of dILP secretion compared to males, to control systemic insulin signaling and consequently body size.
To test this, we measured pupal volume in tra mutants in which we genetically increase IIS activity via heterozygous loss of PTEN, a known inhibitor of IIS. We found that loss of one copy of PTEN rescued the decreased body size in tra mutant females (Fig 4F). We next wanted to test whether a reduction in IIS could suppress the ability of Tra to drive growth. Using da-GAL4 to drive ubiquitous expression of the UAS-tra transgene, we measured body size in larvae heterozygous for null or hypomorphic alleles of the InR. This genetic inhibition of IIS blocked Tra-induced overgrowth (Fig 4G). Together, these results support a model of sex-specific growth in which Tra function in the female fat body stimulates the release of dILPs from the IPC. Higher dILP levels stimulate IIS activity to promote increased body growth in females. Overall, our results identify sex determination gene Tra as an additional regulator of the highly conserved IIS pathway.
In almost all animals, sex is an important determinant of body size [4]. While sex hormones have been shown to control the rate and duration of growth in mammals to achieve SSD, the mechanisms underlying male-female differences in growth in invertebrates are less clear [3, 59]. We therefore used Drosophila larvae as a model to study the mechanisms underlying SSD. We identified clear male-female differences in the control of cell and body size that precede the differentiation of adult sexual morphology (eg. sex combs, abdominal pigmentation, genitalia). Since the duration of larval growth does not significantly differ between male and female larvae [5], these results implicate the sex-specific regulation of growth as a key determinant of SSD in Drosophila. While previous studies showed that master sex determination gene Sxl contributes to sex differences in body size [23], it was unclear whether these effects on growth were mediated by Sxl’s regulation of the sex determination pathway, or the process of dosage compensation. Also, Sxl’s role as a master sex determination gene is not conserved in all insects [60], suggesting other genes may contribute to SSD in these other species.
In our study, we show that sex determination gene tra contributes to SSD in Drosophila. This suggests that the sex-specific regulation of growth is at least partly independent of dosage compensation, as Tra does not regulate this process [30]. Since tra’s role in sex determination is widely conserved in insects, many of which show SSD, the sex-specific regulation of growth by Tra may be a conserved mechanism to create dimorphic body size across many insect species [61, 62]. It is important to note, however, that in spite of our results demonstrating an important role for Tra in creating SSD in Drosophila, loss of Tra function in females does not fully ‘masculinize’ body size, as tra mutant females remain significantly larger than wild-type males. There must therefore be other genes that contribute to increased female body size. One obvious candidate is Sxl, where Sxl mutant females have a male-like body size. The tra-independent effects of sex on body size may therefore be regulated by Sxl. This could occur in one of two ways: first, by Sxl acting on targets in addition to Tra, or second, by the effects of Sxl on dosage compensation. A recent study by Evans and Cline [63] showed that one female-specific behaviour, ovulation, was controlled by Sxl in a tra-independent manner. This ‘tra-insufficient feminization’ branch of the pathway does not cause any misregulation of the dosage compensation pathway, providing strong evidence that additional, as yet unknown, targets of Sxl mediate its effects on ovulation. In the case of SSD, then, other targets of Sxl may explain the tra-independent effects of sex on size. In addition to potential targets other than tra, the effects of Sxl on SSD may alternatively be mediated by its regulation of dosage compensation. In females, the presence of Sxl prevents the activation of the dosage compensation complex, whereas absence of Sxl in males allows dosage compensation to be activated to promote male development [11]. Loss of Sxl in females causes the inappropriate activation of this complex. Thus the decreased body size of Sxl mutant females may be explained by the ectopic activation of the dosage compensation complex. In the future, it will be interesting to dissect the individual contributions of Sxl, tra, the dosage compensation complex, and additional Sxl targets, to the control of male-female differences in body size in Drosophila. Further, it will be interesting, where possible, to determine whether these genes perform similar roles in SSD in other insect species.
One key finding from our work is that sex differences in body growth are regulated by Tra independently of sexual differentiation, behaviour and reproduction. To date, most studies have shown that Tra’s effects on sexual development are mediated by its known targets dsx and fru [43]. Together, dsx and fru regulate most aspects of sexual development and behaviour. However, our data shows that Tra’s effects on body size are independent of dsx and fru. Combined with our data showing that masculinizing or feminizing the gonads or germline has no effect on body size, this shows that that sex differences in body size are not simply a consequence of sexual differentiation, reproduction and behaviour. Instead, SSD in Drosophila is regulated by Tra in a separate pathway, separate from other sex determination genes and aspects of sexual dimorphism. One possible explanation for SSD to be regulated separately of other aspects of sexual development is that while increased female body size is an important sexual trait, as it is related to fecundity [64], the inability to adjust body size in response to environmental factors such as low nutrition can compromise survival during larval life [47]. Therefore, unlike aspects of sexual dimorphism that must be fixed to permit reproduction (eg. gonad and germline differentiation, female neural circuits for egg-laying), body size must show a higher degree of plasticity. Indeed, studies have shown that male genital discs in Drosophila are less sensitive to growth perturbation than other imaginal discs [65]. We therefore propose that sexually dimorphic body growth is regulated independently from other aspects of sexual differentiation to allow body size to be co-ordinated with environmental conditions.
Another finding from our work is that Tra function in the fat body can regulate the growth of other tissues to influence body size in a non cell-autonomous manner. Previous studies have also identified non cell-autonomous interactions that determine the sex of the genital disc, the development of the male-specific muscle of Lawrence, or sexual dimorphism in the gonad [66–69]. Combined with our data that sex differences in body size are also regulated in a non cell-autonomous manner, this suggests that in Drosophila, like in mammals, some aspects of sex determination and sexual dimorphism are regulated in a non cell-autonomous manner. Our identification of sex differences in the secretion of dILP2 suggest that this conservation extends to the cell-cell signaling pathways that mediate growth, as sex hormones in mammals are known to control male-female differences in body size via regulation of the growth hormone (GH)/insulin-like growth factor 1 (IGF1) axis [59]. Several recent studies have shown that higher levels of circulating dILPs can increase body growth by augmenting IIS activity [70,71]. Our findings therefore suggest a model of sex-specific growth in Drosophila in which the sex of the fat body, as determined by the presence (females) or absence (males) of Tra, is one contribution to the sex differences in body size via regulation of dILP secretion. Higher dILP secretion in females leads to elevated IIS activity, and consequently an increase in body size. This model of increased female body size is supported by data that flies lacking all three IPC-derived dILPs (dilp2-3,5 triple mutants) show a 40% reduction in body weight in females, but no effect on body weight in males [72]. Similarly, female body size is more strongly affected than in males in animals with loss-of-function mutations in components of IIS such as chico or InR [73,74]. Together, these findings highlight the importance of sex as a critical determinant of dILP secretion, and IIS-mediated body growth.
During larval development, the fat body responds to a variety of extrinsic and intrinsic cues such as nutrients and hormones to control body growth. For example, in response to nutrient input, the fat body releases an as-yet-unidentified factor into the larval hemolymph [33]. This secreted factor acts in an endocrine manner to control the release of dILPs from the IPC in the brain. Our studies have identified sex as an additional factor that alters the function of the fat body to influence body growth in a non cell-autonomous manner. In particular, we identified a role for the function of sex determination gene Tra in the fat body as one factor influencing SSD in flies. Yet it is unclear how Tra function in the fat body influences the molecular and physiological properties of this tissue to influence body size. Tra is a member of the conserved family of SR proteins. These proteins play well-characterized roles in the regulation of alternative splicing, and have also been shown to influence other aspects of RNA metabolism, such as regulation of mRNA translation [75–77]. Tra may therefore act in two ways in the fat body to control dILP2 release: 1) via sex-specific splicing to facilitate production or secretion of the secreted factor(s), or 2) in a more general mechanism by influencing mRNA translation to elevate production or secretion of these fat-to-brain signals. Although the regulation of dILP secretion is an established mechanism to regulate body growth, the molecules that are released by the fat body to control dILP release are only beginning to be identified. For example, the cytokine-like molecule unpaired 2 (upd2), and the peptide hormone CCHa2 play roles in coupling fat body function to regulation of dILP secretion and body size [32, 78]. Similarly, Hedgehog was also identified as a factor that can control dILP secretion in an endocrine manner [31]. In adults, fat body-derived dILP6 or dawdle, an Activin-like ligand in the TGF-β superfamily, could both influence the secretion of IPC-derived dILPs [79, 80]. In addition, several neuropeptides and neurotransmitters have also been shown to regulate IPC activity and dILP release [81]. In the future, it will be interesting to determine whether Tra directly regulates any of these known secreted factors to control dILP2 release. However, an additional possibility is that tra does not directly regulate any secreted factors; instead, tra’s effects on growth may be mediated by effects on mRNA translation. Many studies have identified the regulation of mRNA translation in the fat body as a limiting factor for growth during development. For example, two studies identified significant effects of TOR and Myc in the fat body in promoting dILP release [33, 82]. TOR is an important regulator of mRNA translation, and Myc’s effects were thought to involve elevated levels of ribosome synthesis. A more recent study showed that stimulation of tRNA synthesis, and consequently mRNA translation, in the fat body could drive increased body growth [83]. Future studies will allow us to determine which of Tra’s molecular functions (splicing vs. mRNA translation) determine its contribution to fat body function and consequently growth. Given the increasing awareness of functional similarities between the fly fat body and mammalian liver/adipose tissue, our results suggest the intriguing possibility that the function of these important endocrine organs may be similarly regulated by sex to control systemic growth and physiology in mammals.
In addition to Tra’s non cell-autonomous effects on body size, we found that Tra also has cell- and organ-autonomous effects on size. While our data suggests that Tra’s effects on body size are independent of fru and dsx, since loss of neither gene affects overall body growth or non cell-autonomous growth, it is possible that Tra’s cell-autonomous effects on cell size are mediated by dsx. In the larval fat body, we identified a cell-autonomous requirement for dsx in both males and females to promote growth in fat body cells. We believe the reason that these cell-autonomous effects of dsx on cell size do not affect overall body size is due to the restricted nature of Dsx expression in larvae. Indeed, two studies showed that Dsx expression in larvae is limited to the fat body, CNS, gonads, some regions of the gut, and subsets of imaginal discs [37,39]. However, in spite of the lack of effect on overall body size, previous studies in Drosophila have identified a role for dsx in regulating organ size in other tissues. For example, expression of the male- or female-specific isoforms of Dsx (DsxM and DsxF, respectively) control the sex-specific growth of the genital disc via Wingless and Decapentaplegic signaling [84]. In addition, DsxF has been shown to promote sex-specific programmed cell death in both the larval ventral nerve cord, and in male-specific gonadal precursor cells [85–87]. Our findings identify an additional mechanism by which Dsx controls organ size: regulation of cell growth. While the molecular mechanism by which Dsx controls cell size is unclear, Dsx has been shown to control horn size in stag beetles by regulating tissue sensitivity to a circulating hormone, juvenile hormone [88]. Interestingly, a recent paper identified the insulin receptor (InR) and the ecdysone receptor (EcR) as potential Dsx targets [89]. Since both pathways have been shown to control fat body cell size [47, 82], Dsx may influence fat body cell growth by regulating tissue sensitivity to circulating dILPs or the steroid hormone ecdysone, integrating signals from both the primary sex-determining signal (X:A) and circulating hormones to control tissue growth. In the future, it will be interesting to determine whether the integration of sex and environmental cues is a general feature of Dsx-mediated tissue growth, or whether this mechanism is limited to specific tissues, such as the fat body.
In conclusion, our studies identify Tra as one regulator of sex differences in growth and body size. Moreover, we provide the first link between Tra and IIS in the control of sex differences in body growth. Interestingly, sexual dimorphism in phenotypes such as stress resistance, immune responses and lifespan have been noted in Drosophila [90–95]. These phenotypes are also affected by altering IIS [96–99]. Tra may therefore control sexual dimorphism in a wide variety of phenotypes via regulation of dILP secretion and IIS activity. Deregulation of insulin secretion and IIS activity have been implicated in diseases such as diabetes and cancer [45, 100]. Interestingly, sex differences in incidence have been previously reported for both diabetes and some forms of cancer [101, 102]. Thus future studies on the link between sex and insulin secretion/IIS activity may explain why one sex is predisposed to these diseases.
Larvae were raised on food at a density of 50 larvae per vial at 25°C [83, 103]. The following fly GAL4 stocks were used in this study: da-GAL4, r4-GAL4 (fat body), cg-GAL4 (fat body), elav-GAL4 (neurons), repo-GAL4 (glia), P0206-GAL4 (ring gland), Mef2-GAL4 (muscle), Act5c-GAL4 (ubiquitous), en-GAL4 (posterior compartment of the wing), nos-GAL4 (germline), c587-GAL4 (gonad). We used the following UAS lines: UAS-tra2-RNAi (TriP), UAS-tra2-RNAi (VDRC), UAS-tra, UAS-dsx-RNAi (TRiP). The following mutant strains were used: w1118, foxoΔ94, tra1/TM6B, Df(3L)st-j7/TM6B, dsx1/TM6B, Df(3R)dsx15, tud1/CyO::GFP, fruF/TM6B, fru4-40/TM6B, fruΔtra/TM6B, pten100;CyO::GFP;MKRS/TM6B. We used the following stocks for flp-out experiments: hsflp;;UAS-tra2-RNAi, hsflp;UAS-tra, hsflp;;UAS-dsx-RNAi, act>stop>CD2>stop>GAL4. Larvae were sexed using gonad size. Where gonad size could not be used to sex larvae (eg. dsx or tra mutants, da>UAS-tra, Mef2>UAS-tra, Act5c>tra2-RNAi or Act5c>UAS-tra), males with a GFP on the X chromosome (Ubiquitin-GFP) were crossed to the virgin females of the correct genotype. In the progeny of the cross, females were GFP-positive and males were GFP-negative [5].
Pupal volume was measured as previously described [82]. n>60 per genotype.
Measured as previously described [83, 103]. n>40 per genotype.
Five-day-old adult flies lacking gonads were weighed in groups of six in 1.5 ml tubes on an analytical balance. The gonads were removed prior to weighing by dissection; n>30 per genotype.
96 hr larvae were fed for the indicated amounts of time on yeast paste containing 0.05% Bromophenol blue. After feeding for the desired amount of time, ten larvae were isolated in a 1.5 ml tube, with eight tubes per sex collected in total. 250 μl of PBS was added to the tube and the larvae were homogenized with a micropestle. The lysate was cleared by centrifugation at 5000 rpm for 1 min, then the absorbance at 595 nm was measured in a spectrophotometer.
Total RNA was extracted from larval tissues, then DNase-treated and reverse transcribed using Superscript II, as previously described [83, 103].
Whole larval extracts were prepared as previously described [83, 103]. The P-Akt and total Akt antibodies were obtained from Cell Signaling (#4054 and #9272).
Anti-FOXO antibody was applied to fat bodies dissected from larvae 110 hr AEL (25°C) at a dilution of 1:500, as previously described [80].
Larvae were grown on rich food containing either DMSO or rapamycin, as previously described [83, 103].
All data were analyzed using R Studio using the code described below.
Student’s t-test:
a <- filename$genotype1
b <- filename$genotype2
t.test(a,b)
One-way ANOVA:
aov.PV <- aov(Pupal_Volume ~ Genotype, data = filename)
ls(aov.PV)
summary(aov.PV)
TukeyHSD(aov.PV)
Two-way ANOVA with interaction term:
int <- aov(Pupal_Volume ~ Sex + Treatment + Sex*Treatment, data = filename)
summary(int)
TukeyHSD(int)
|
10.1371/journal.ppat.1000379 | The NOD/RIP2 Pathway Is Essential for Host Defenses Against Chlamydophila pneumoniae Lung Infection | Here we investigated the role of the Nod/Rip2 pathway in host responses to Chlamydophila pneumoniae–induced pneumonia in mice. Rip2−/− mice infected with C. pneumoniae exhibited impaired iNOS expression and NO production, and delayed neutrophil recruitment to the lungs. Levels of IL-6 and IFN-γ levels as well as KC and MIP-2 levels in bronchoalveolar lavage fluid (BALF) were significantly decreased in Rip2−/− mice compared to wild-type (WT) mice at day 3. Rip2−/− mice showed significant delay in bacterial clearance from the lungs and developed more severe and chronic lung inflammation that continued even on day 35 and led to increased mortality, whereas WT mice cleared the bacterial load, recovered from acute pneumonia, and survived. Both Nod1−/− and Nod2−/− mice also showed delayed bacterial clearance, suggesting that C. pneumoniae is recognized by both of these intracellular receptors. Bone marrow chimera experiments demonstrated that Rip2 in BM-derived cells rather than non-hematopoietic stromal cells played a key role in host responses in the lungs and clearance of C. pneumoniae. Furthermore, adoptive transfer of WT macrophages intratracheally was able to rescue the bacterial clearance defect in Rip2−/− mice. These results demonstrate that in addition to the TLR/MyD88 pathway, the Nod/Rip2 signaling pathway also plays a significant role in intracellular recognition, innate immune host responses, and ultimately has a decisive impact on clearance of C. pneumoniae from the lungs and survival of the infectious challenge.
| Chlamydophila pneumoniae (C. pneumoniae) is a common intracellular parasite that causes lung infections and contributes to several diseases characterized by chronic inflammation. Toll-like receptors expressed on the cell surface detect C. pneumoniae and mount a vigorous defense, but it is not known how the cell defends itself once the pathogen has taken up residence as a parasite. We reasoned that cytosolic pattern recognition receptors called Nods (nucleotide oligomerization domain) that detect microbes that gain entry into the cell might be involved. Using mice genetically deficient in Nod1 and Nod2 or their common downstream adaptor (Rip2), we show that in lung infection, Nod proteins are indeed essential in directing a defense against C. pneumoniae. Mice with defective Nod/Rip2-dependent signaling exhibited delayed recruitment of neutrophils, blunted production of pro-inflammatory cytokines and chemokines, and evidence of defective iNOS expression and NO production. These impaired responses led to delayed clearance of bacteria, intense persistent lung inflammation, and increased mortality. By performing bone marrow transplantation experiments and direct transfer of cells into the lungs of mice, we demonstrated that intact Nod-dependent signaling in bone marrow–derived cells was critical in the defense against C. pneumoniae. Our results indicate that Nod proteins also play an important role in host defense against C. pneumoniae. Coordinated and sequential activation of TLR and Nod signaling pathways may be necessary for an efficient immune response and host defense against C. pneumoniae.
| Chlamydophila pneumoniae is a Gram-negative obligate intracellular pathogen that is widely prevalent [1], causes respiratory tract diseases such as pneumonia, sinusitis, and bronchitis, contributes to acceleration of atherosclerosis [2],[3], and is associated with development of chronic lung diseases such as asthma [4] and other disorders where chronic inflammation is a hallmark feature [5],[6]. C. pneumoniae infects various cell types such as epithelial cells, monocytes, macrophages, smooth-muscle cells and endothelial cells, and often resides intracellularly for indefinite periods [7].
C. pneumoniae induces a similar lung pathology in humans and rodents [8]. A mouse model of lung infection has been used to study the immunological mechanisms of host defenses. Host immune responses to C. pneumoniae proceeds in two stages; 1) an early response requiring IFN-γ to limit the growth of the intracellular bacteria, which plays a central role in the innate control of this infection, and 2) a later adaptive immune response that includes CD4+ and CD8+ T cells in bacterial clearance and protection [9]–[11]. While the primary immune response is aimed to clear the primary infection from the host and provide protection against reinfection with the same pathogen, generation of tissue injury also occurs and Chlamydial infections often recur or remain persistent and long-term consequences of recurrent or persistent chlamydial infections can be severe [10],[12].
Chlamydia is internalized by macrophages as well as by “non-professional” phagocytes, where it survives and replicates. C. pneumoniae elicits IFN-γ production in infected bone marrow-derived macrophages [13]. In such cells, IFN-γ synergizes with bacterial products to activate various bactericidal mechanisms, including inducible nitric oxide synthase (iNOS), which leads to production of NO [14],[15], which in turn inhibits chlamydial growth [14],[16],[17].
Molecular motifs derived from C. pneumoniae are detected by several pattern recognition receptors, especially Toll-like receptor 2 (TLR2) and TLR4 [18],[19]. TLR4 recognizes chlamydial components such as lipopolysaccharide (LPS) and heat shock protein 60 (cHSP60) [20]–[24], and the intact organism stimulates TLR2 and TLR4-mediated responses [25],[26]. TLR-mediated signaling triggered by C. pneumoniae-derived molecules instigates development of an inflammatory innate immune responses and TLR/MyD88 signaling plays an important role in host responses against C. pneumoniae infection [18],[19]. Studies from our laboratory indicate that MyD88-null mice with C. pneumoniae lung infections are unable to mount a sufficient early inflammatory response against the pathogen [18]. These mice show marked delays in recruiting PMNs, CD8+ and CD4+ T cells to the lungs, and fail to clear the pathogen, but then develop a severe, late-stage, and persistent inflammation characterized by increased IL-1β and IFN-γ production that leads to increased mortality [18]. In contrast, TLR4−/−, TLR2−/−, and WT mice—all of which can detect C. pneumoniae and can signal normally via MyD88, readily recovered from the infection and cleared bacteria normally, indicating that MyD88 is essential to an effective defense, but that TLR2 and TLR4 can both detect the pathogen and are therefore redundant [18],[19].
C. pneumoniae has a unique biphasic developmental cycle that occurs within the chlamydial inclusion, a membrane-bound vacuole that is trafficked to the peri-Golgi region, where it avoids fusion with lysosomes and destruction, and are able to replicate intracellularly [27],[28]. Chlamydia-mediated vesicular trafficking events transform the inclusion into a compartment from which chlamydiae can acquire nutrients and interfere with multiple host cell functions [29],[30]. While residing intracellularly, the pathogen presumably is not detected by the cell surface TLR2 and TLR4 receptors; hence, it is unclear how C. pneumoniae might be detected and held in check once it has been taken up by the cell. C. pneumoniae–infected macrophages can limit bacterial growth by expression of IFN-γ, which in turn is controlled by TLR4/MyD88-dependent pathway. However, since Chlamydia can also induce IFN-γ in the absence of TLR4/MyD88 signaling [31], a potential role for TLR-independent and intracellular recognition receptors, such as the nucleotide oligomerization domain (Nod) proteins, has been suggested [31].
Nod proteins and their adaptor molecule Rip2 also known as RICK or CARDIAK are key components of a family of cytosolic innate immune pattern recognition receptors [32]–[36]. Nod1 and Nod2 recognize molecules in the cytoplasm that originate from bacteria, including peptidoglycan (PGN), a component of bacterial cell walls, and the muramyl dipeptide (MDP) structure found in almost all bacteria [37]. Both Nod1 and Nod2 signal via the serine/threonine Rip2 kinase [34],[38],[39]. Once activated, Rip2 mediates activation of NF-κB and the subsequent production of inflammatory cytokines such as TNF-α and IL-6 [40]–[42]. Although some reports indicate that Nod/Rip2-mediated signaling does not induce IFN-γ [43], other studies show that combined TLR and Nod/Rip2 signaling together can lead to IFN-γ expression [44].
In the present study we show that the Nod/Rip2 signaling pathway is essential to detect intracellular C. pneumoniae and direct subsequent innate immune host defenses and bacterial clearance in a mouse model of pneumonia, in addition to the well-established role of the TLR/MyD88 pathway. Rip2−/− mice infected with C. pneumoniae displayed an impaired cytokine and chemokine release such as IFN-γ, KC and MIP2, and showed impaired iNOS mRNA expression and NO production, and delayed neutrophil recruitment, which led to delayed bacterial clearance, an intense late-stage and persistent lung inflammation and increased mortality.
Rip2−/− mice and WT controls were infected intratracheally with C. pneumoniae (1×106 IFU/mouse) and evaluated for lung inflammation by histopathological analysis. Tissue sections were obtained at 3, 5, 14, and 35 days after infection, fixation and histological staining (H&E) was performed, and sections were graded for degree of inflammation in blinded fashion as detailed in the Materials and Methods Section. As expected, C. pneumoniae–infected WT mice developed marked lung inflammation as expected by days 5 and 14 and cleared the inflammation and recovered to baseline by day 35 (Figure 1A and 1B). However, Rip2−/− mice developed significantly greater lung inflammation than WT mice by day 5, and day 14, which persisted until the end of the study period at day 35 (Figure 1A and 1B).
Innate immune responses, and particularly IFN-γ plays an important role in host defense against acute infection and in establishment of persistence of C. pneumoniae [10]. We, therefore, determined the production of cytokines such as IL-6, IL-12 p40 and IFN-γ levels in BALF and lung homogenates from infected Rip2−/− and WT mice on days 3, 5 and 14. Concentrations of IL-6, IL-12p40, and IFN-γ were significantly reduced in BALF of Rip2−/− mice at day 3 compared to WT mice (Figure 1C). However, by day 5 and day 14, IL-6, IL-12p40 and IFN-γ concentrations in the BALF and lung homogenates from Rip2−/− mice were significantly increased and exceeded levels in WT mice (Figure 1C). Thus, in addition to increased histopathological inflammation seen in Rip2−/− mice on days 5 and 14 and during the later stages, we observed an initial impaired and delayed kinetics in cytokine production in C. pneumoniae–infected Rip2−/− mice on day 3, which was also followed by a significant increased in cytokine production in the lungs on days 5 and 14 (Figure 1C). We next measured IL-6 and IFN-γ levels in the supernatant of infected bone marrow–derived macrophages (BMDM) and whole lung cells ex-vivo. C. pneumoniae infection-induced cytokine production ex-vivo (IL-6, and IFN-γ release) were significantly impaired in both Rip2−/− macrophages and whole lung cells compared to WT macrophages and whole lung cells (Figure S1A–S1D).
Our data show impaired cytokine production in Rip2−/− mice infected with C. pneumoniae early on day 3 following infection but a significant reversal and increase in cytokines and more severe and persistent lung inflammation by day 5 and 14 compared to WT mice (Figure 1). We hypothesized that this more severe and persistent lung inflammation was due to an inability of Rip2−/− mice to clear bacteria, which would then be expected to continue to provoke inflammation and cause the delayed increase in cytokine production. To test this hypothesis, we performed quantitative bacterial cultures in the lungs of mice at days 3, 5, and 14 post-infection. As anticipated, we observed significantly higher numbers of C. pneumoniae IFU in the lungs of Rip2−/− mice on days 5 and 14 compared to WT mice (Figure 2A). This could not be explained by higher baseline load of bacteria in Rip2−/− mice, since on day 3, bacterial numbers in lungs were similar between WT and Rip2-deficient mice (Figure 2A). Consistent with the bacterial clearance data, virtually all WT mice survived the infectious challenge, while Rip2−/− mice had significantly increased mortality, and less than half the Rip2−/− mice survived until the end of the experiment at day 35 (Figure 2B). Furthermore, the lungs from the Rip2−/− mice that succumbed to infection harbored an abundance of C. pneumoniae (data not shown), while those who survived cleared the bacteria but still manifested chronic lung inflammation at day 35 (Figure 1B). Collectively then, these data indicate that: 1) Rip2 importantly contributes to clearance of C. pneumoniae from the lungs; and 2) in the absence of Rip2, severe lung inflammation occurs and persists, but fails to effectively combat the infection.
Polymorphonuclear neutrophils (PMN) are crucial for innate host defense against bacteria and fungi. We previously reported that MyD88−/− mice infected with C. pneumoniae fails to recruit PMN into the lungs during early and late stages of the infection [18]. To investigate the PMN recruitment in Rip2-deficient mice, we infected Rip2−/− and WT mice with C. pneumoniae intratracheally, and compared total cells and PMN in BALF on day 3 and 5 following infection. Both PMN and total BALF cells in Rip2−/− mice were significantly lower compared to WT mice on day 3 following infection (Figure 3A and 3B). However, by days 5 and day 14 post-infection, PMN as well as total BALF cell counts in Rip2−/− mice increased markedly, and were significantly higher than in WT mice (Figure 3A and 3B). Assessment of neutrophil recruitment to the lung by flow cytometric analysis demonstrated similar results (Figure 3C). The percentage of neutrophils (defined by Gr1+ CD11b+ cells) in the lungs of Rip2−/− mice were reduced on day 3 of infection, but increased thereafter, and by days 5 and 14, significantly exceeded the neutrophil percentage of lung cells in WT mice (Figure 3C). We next sought to examine whether the chemokines associated with neutrophil recruitment in the lungs were also affected in the Rip2-deficient mice. Rip2−/− mice showed significantly lower concentrations of KC and MIP-2 in both BALF and lung homogenates compared with WT mice on day 3 after infection (Figure 3D). However, both KC and MIP-2 levels in BALF and lung homogenates significantly increased in Rip2−/− mice compared to WT mice by day 5 (Figure 3D). Collectively, these data indicate that Rip2 plays an important role in early cytokine and chemokine production and neutrophil recruitment to the lungs during the initial days after C. pneumoniae infection, and Rip2-deficiency leads to delayed bacterial clearance, which is followed by an exaggerated secondary response consisting of increased cytokine and chemokine expression, PMN recruitment, prolonged, severe histopathological inflammation in the lungs, and increased mortality.
Alveolar epithelial cells are the main cells infected in lung infection model [45], but C. pneumoniae also infects different cell types including macrophages, dendritic cells, endothelial cells, and PMNs [7]. To determine which cells in the lungs are infected by C. pneumoniae, we analyzed infected cell profiles by flow cytometry. C. pneumoniae was predominantly found in macrophages and neutrophils, but also in alveolar epithelial cells in the infected lungs (Figure 4A and 4B). Interestingly, in Rip2−/− mice, the number of neutrophils that contained C. pneumoniae was significantly increased compared to that in WT mice (Figure 4A). To address whether more bacteria are in Rip2−/− macrophages and neutrophils, we analyzed mean fluorescence intensity (MFI) per cells, which corresponds to relative bacterial number (Figure 4C). We observed a shifted histogram in Rip2−/− neutrophils. These data revealed that neutrophils are likely the main site of Chlamydial replication in lungs at day 5 after infection in Rip2−/− mice.
We hypothesized that a bactericidal factor produced by immune effector cells might be responsible for the failure of Rip2−/− mice to clear bacteria. NO produced after cell activation by IFN-γ is important for killing or inhibiting growth of microorganisms [46]. Both IFN-γ and iNOS play major roles in host resistance to chlamydial infection [10]. We therefore assessed the levels of iNOS in the lungs following C. pneumoniae infection. Rip2−/− mice demonstrated significantly impaired iNOS mRNA expression compared to WT mice from day 0 until day 5 in total lung cells examined ex vivo (Figure 5A). In addition, bone marrow-derived macrophages obtained from Rip2−/− mice, showed significantly diminished NO production following in vitro infection with C. pneumoniae compared to WT macrophages (Figure 5B). These results suggest that NO production plays a role in killing and clearance of C. pneumoniae, and also suggest that Rip2 signaling contributes to NO production in response to C. pneumoniae infection. Consistent with this interpretation, C. pneumoniae growth was significantly increased in WT macrophages in the presence of an iNOS inhibitor (L-NMMA) compared to control cells treated with an inactive form of the inhibitor (D-NMMA) (Figure 5C). In contrast, C. pneumoniae growth was not affected by treatment with the iNOS inhibitor in Rip2−/− macrophages (Figure 5C). Collectively, these results suggest that Rip2-deficient mice have impaired iNOS expression and NO production in response to C. pneumoniae infection, which likely contribute to the host immune response defect and delayed bacterial clearance from the lungs of Rip2−/− mice.
Our results thus far indicate that the Rip2−/− mice display an impaired host defenses, delayed bacterial clearance, and increased mortality following C. pneumoniae lung infection. Since Rip2 is utilized by both Nod1 and Nod2, we next wished to determine the role of these upstream receptors in C. pneumoniae infection. Nod1 was shown to play a role in C. pneumoniae-mediated activation of human endothelial cells in vitro [47], But it is unclear which Nod receptors detect C. pneumoniae in macrophages and during in vivo infection. Nod1 is ubiquitously expressed in mammalian cells, but the expression of Nod2 is mainly restricted to primary antigen-presenting cells and epithelial cells, and Nod2 is not expressed in endothelial cells [48]. Furthermore, our data (Figure 4C) indicates that C. pneumoniae mainly replicates in macrophages and neutrophils. To determine which Nod receptor recognizes intracellular C. pneumoniae in macrophages, we infected Nod1−/− or Nod2−/− BMDM with live C. pneumoniae and measured KC and NO levels in the supernatant. Nod1−/− and Nod2−/− macrophages produced significantly diminished KC and NO (Figure 6A and 6B). Consistent with this decreased NO production, bacterial viability was significantly higher in both Nod1−/− and Nod2−/− macrophages in vitro (Figure 6C). To determine if greater bacterial viability in the absence of Nod1 and Nod2 also occurred in vivo, we infected Nod1−/− and Nod2−/− mice with C. pneumoniae and examined bacterial clearance in the lungs. In agreement with the in vitro data, both Nod1−/− and Nod2−/− mice displayed delayed pulmonary bacterial clearance compared to WT controls, as reflected by significantly higher bacterial counts in Nod1−/− and Nod2−/− mice at 5 days post-infection (Figure 6D). These results are consistent with the conclusion that intracellular C. pneumoniae is recognized by both Nod1 and Nod2 in macrophages, and that signaling emanating from both Nod1 and Nod2 significantly contributes in host defenses against C. pneumoniae lung infection, at least in part by regulating production of NO and inflammatory cytokines and chemokines such as IL-12 p40, IFN-γ, KC and MIP2.
Based upon data in the previous section that showed involvement of macrophages and PMNs in the lungs, we hypothesized that the Nod/Rip2 signaling pathway in bone marrow (BM)-derived cells rather than non-hematopoietic stromal cells was primarily responsible for innate immune host responses and clearance of C. pneumoniae from the lungs. To test this notion, we generated chimeric mice using donor marrow from WT or Rip2−/− mice (Figure S2), then infected the mice intratracheally with C. pneumoniae. Five days after infection, lungs were harvested and quantitative bacterial counts were determined. WT recipient mice that were transplanted with Rip2−/− BM displayed significantly higher bacterial load in the lungs compared to control WT mice transplanted with WT BM (Figure 7A). Conversely, Rip2−/− recipient mice that had been transplanted with WT BM displayed lower bacterial counts than control Rip2−/− mice that had received Rip2−/− BM (Figure 7A). In all chimeric mice, we observed generally higher bacterial titers observed in the lungs, most likely due to inherently increased susceptibility secondary to the irradiation procedure itself, as has been previously reported by other investigators [49]. These data indicate that Rip2 in BM-derived cells primarily mediates host defenses against pulmonary C. pneumoniae infection. However, it is possible that airway epithelial cells also contribute and play a role in C. pneumoniae detection in the lung. In order to further elucidate the primary role of macrophages in C. pneumoniae infection, we performed intratracheal adoptive transfer of WT or Rip2−/− BMDM, simultaneously with C. pneumoniae infection (i.e. macrophages mixed with bacteria), and then determined the effect on local bacterial replication in lungs of WT or Rip2−/− mice. As anticipated, adoptive transfer of Rip2−/− macrophages plus C. pneumoniae into WT mice resulted in significantly higher bacterial counts compared to WT macrophages plus C. pneumoniae transferred into WT mice (Figure 7B). However, WT macrophages plus C. pneumoniae adoptively transferred into Rip2−/− mice rescued the Rip2 phenotype, i.e. restored bacterial clearance (Figure 7B) and neutrophil recruitment in the lungs (Figure S3), indistinguishable from control WT mice that received WT macrophages plus C. pneumoniae. We did not observe a defect in phagocytosis in Rip2−/− macrophages using labeled C. pneumoniae (Figure S4) [50]. Taken together, these findings indicate that the Nod/Rip2 signaling pathway in BM-derived cells play a dominant role in bacterial clearance of C. pneumoniae from the lungs.
Here we show that cytoplasmic Nod proteins are importantly involved in generating innate immune defenses against intracellular C. pneumoniae. We found that deletion of Rip2, that is essential for both Nod1- and Nod2-mediated signaling, delays neutrophil recruitment to the lungs, and suppresses expression of chemokines and cytokines that are essential to generate an effective host defense. Although an inflammatory innate immune response was delayed, by day 5 Rip2−/− mice infected with C. pneumoniae developed a more severe inflammation that persisted longer compared to WT mice, but nevertheless failed to clear the pathogen, and most infected Rip2−/− mice ultimately succumbed to the infection. The inability of Rip2−/− mice to eradicate the pathogen despite robust inflammation was associated with delayed kinetics of IL-12p40 and IFN-γ production and suppressed iNOS expression and NO production, all of which are critically important elements in the innate immune armamentarium [9],[11]. Experiments with bone marrow-derived macrophages demonstrated that both Nod1 and Nod2 were involved in sensing intracellular C. pneumoniae. Results from experiments with bone marrow chimeric mice confirmed that cells derived from hematopoietic lineages rather than resident stromal cells were essential in Nod/Rip2-mediated defenses against C. pneumoniae. This conclusion was corroborated by adoptive transfer of WT or Rip2−/− macrophages directly into the airways of infected mice. Collectively, our data demonstrate that proper functioning of the Nod/Rip2 cytoplasmic innate immune detection system critically determines whether the host can effectively resist and eradicate an infectious challenge to the lungs from C. pneumoniae.
Our results underscore previous reports that increasingly emphasize the central role Nod/Rip2 signaling can play in defending the host against intracellular invasion. Nod proteins mediate host defense against a variety of both Gram negative and Gram positive bacteria. For example, Nod2 senses the PGN produced by Staphylococcus aureus, and Rip2 limits S. aureus growth in macrophages [51]. Also, Nod2 detects the MDP structure found in almost all bacteria [37]. Nod1 is required for IkK and NF-κB activation in human colon epithelial cells infected with E. coli [52], participates in KC induction and impacts bacterial viability to Pseudomonas aeruginosa in mouse embryonic fibroblasts [53]. Nod2 triggers cytokine production by DCs in response to live M. tuberculosis, but is not essential to control M. tuberculosis airway infection [54]. While Rip2 and Nod1 deficiency increases susceptibility to Listeria monocytogenes [32] and Helicobacter pylori [55] respectively, neither Nod1 nor Rip2 deficiency had any significant effect on Chlamydia muridarum vaginal infection [56]. Hence, available data indicates that Nod1 and Nod2 selectively interact with specific pathogens and can play a critical role in host defenses, highlighted by increased susceptibility to specific pathogens in mice lacking these intracellular receptors. Accordingly, our data now indicate that Nod/Rip2 signaling is also essential in successfully combating C. pneumoniae lung infection. Indeed, in the absence of Nod/Rip2 signaling, lung infection with C. pneumoniae proved fatal to the majority of mice.
Clearly, one reason for the increased susceptibility of the Rip2−/− mice to C. pneumoniae lung infection may be the inability of Rip2−/− mice to rapidly recruit neutrophils to the site of infection. Delayed neutrophil recruitment in Rip2−/− mice appears closely linked to lack of sufficient chemokine and cytokine expression compared to infected WT mice. However, by days 5 and 14 post-infection, the percentage of neutrophils in the lungs of Rip2−/− mice increased and significantly exceeded those in WT mice. Evidently, this delayed early response to the infectious challenge allows the bacteria to gain the upper hand, but despite the significant increase in PMN numbers on days 5 and 14 post-infection, the Rip2−/− mice showed delayed bacterial clearance from the lungs and increased mortality. Interestingly, in Rip2−/− mice, the number of neutrophils in the lungs that contained C. pneumoniae was significantly increased compared to that in WT mice. Several obligate intracellular microbial pathogens develop mechanisms to evade destruction upon ingestion by PMN [57],[58]. Indeed, C. pneumoniae can infect and replicate in PMNs and these cells in turn can enhance replication and Chlamydial burden during infection [59].
However, the main reason Rip2−/− mice cannot clear the infection and most often succumb to the disease appears to be more closely tied to our observation that expression of IL-12p40, IFN-γ, iNOS and NO are suppressed in the absence of Rip2−/−. These data are consistent with previous reports, which similarly indicate that IL-12p40 [9], IFN-γ [9],[60] and iNOS [11],[61] are essential for effective host resistance against C. pneumoniae infection. Several studies showed that IFN-γ and IL-12 both play an important role in the innate control of C. pneumoniae infection [9],[11] IL-12 participates in resistance to C. pneumoniae, likely by enhancing IFN-γ mRNA [9]. In turn, IFN-γ produced by innate cells increases iNOS expression and NO release and controls the intracellular growth of C. pneumoniae [14],[16],[17]. Increased susceptibility of IFN-γR−/− mice is associated with diminished levels of iNOS mRNA accumulation in lungs, and iNOS−/− mice also show higher sensitivity to C. pneumoniae infection [11]. However, IFN-γR−/− mice shows even greater sensitivity to C. pneumoniae infection compare to iNOS−/− mice, suggesting the presence of both iNOS-dependent and -independent IFN-γ-mediated effector mechanisms [11].
IFN-γ can be produced by cells from both the innate and acquired immune system. The susceptibility of IFN-γR−/− mice largely exceeds that of RAG-1−/− mice, suggesting an important role for non-T cell-mediated IFN-γ-producing cells in the host resistance against C. pneumoniae infection [11]. Several studies show that, besides NK and T cells, myeloid cells such as macrophages, DCs and neutrophils can also express IFN-γ [62].
Previous studies have shown that MDP induced NO production in macrophages [63],[64]. Several other reports also suggest a link between Nod and NO production [64]. Our results indicating that Nod/Rip2 signaling stimulates iNOS expression and NO production suggest that one of the reasons why C. pneumoniae lung infection proves lethal to most Rip2−/− mice is because they fail to generate an effective NO-mediated defense by immune effector cells, and thus cannot eradicate the pathogen.
In addition to innate immunity, adaptive immune responses may also directly or indirectly diminish the levels of IFN-γ and IL-12 mRNA early after infection and thus may alter the quality of the protective host immune responses. A protective role for CD8 T cells is shown by the higher sensitivity and enhances severity of infection in CD8−/− mice [11]. In MyD88−/− mice we observed a delay in recruitment of CD4 and CD8 T cells into the lungs [18], while in the current study with Rip2−/− mice we observed primarily a delay CD4 T cell recruitment initially followed by significant increases in the presence of both CD4 and CD8 T cells later on day14 (Figure S5).
Our data with bone marrow chimeric mice clearly demonstrate that the cells responsible for Nod/Rip2-dependent defense against C. pneumoniae are hematopoietic in origin, and are not resident stromal cells. Additionally, adoptive transfer of WT macrophages was able to rescue the bacterial clearance defect in Rip2−/− mice. While our data do not completely rule out a potential role for other cell types during C. pneumoniae infection, the bactericidal effects of the Nod/Rip2 pathway appear to be predominantly of bone marrow origin with macrophages playing the largest role.
C. pneumoniae-induced IL-12p40 production in vivo involves both MyD88-dependent and MyD88-independent pathways [59], suggesting a TLR-independent, but Nod-dependent mechanism of recognition and activation. Indeed, recent studies suggest that in endothelial cells, Nod1 plays an important role in triggering C. pneumoniae-mediated inflammatory responses [47], and that Nod1 is involved in NF-κB activation by Chlamydia in epithelial cell lines [56]. Furthermore, a recent study by Buchholz et. al. concluded that C. trachomatis induced IL-8 responses are dependent on Nod1 and Rip2 signaling in Hela cells [65]. Our in vivo data showing that C. pneumoniae induced chemokine production in the lungs depends on Rip2 signaling is consistent with the in vitro observations by Buchholz et. al. These findings suggest that PGN fragments are synthesized by chlamydiae and are recognized by the host innate immune system. The genome sequence revealed that Chlamydophila is actually equipped with a full complement of PGN synthesis genes [66]. Chlamydia is sensitive to antibiotics like penicillin that inhibit PGN synthesis [67],[68], but clear-cut biochemical evidence for the synthesis of PGN in chlamydiae is missing [69],[70]. A recent study revealed the biochemical capacity of C. trachomatis to synthesize m-DAP and that the m-DAP synthesis genes are expressed as early as 8h after infection [71]. This paradox, known as the ‘chlamydial anomaly’ is still being debated in the light of genomic information [72]. However, prior studies and our current data suggest that chlamydial PGN released by bacteria must make their way across the inclusion membrane into the cytosol. One potential mechanism by which this could occur is through the proposed type III secretion system [73]. A similar mechanism of type IV secretion has been proposed for Nod1 signaling in H. pylori infection [55]. While the exact ligand(s) of C. pneumoniae detected by the Nod are yet to be identified, our data clearly indicate that both Nod1 and Nod2 recognize C. pneumoniae and play an essential role in host defenses against this microorganism. Our data differ from those obtained in experimental infections with C. trachomatis or C. muridarum genital tract infection, where Nod1 deficiency had no significant effect on the efficiency of infection, or pathology in vaginally infected mice, while Rip2-deficient mice had only slightly increased bacterial load and delayed bacterial clearance and mildly increased oviduct inflammation [56]. Such differences are not surprising, as the two organisms display only 5 and 10% homology at the DNA and protein levels, respectively, as also reflected in the different pathobiologies they cause [74].
In summary, we demonstrate that the Nod cytosolic pattern recognition receptors are essential for mounting an adequate defense against C. pneumoniae, that Nod stimulate chemokine and cytokine production and neutrophil recruitment in the early phase of infection, and that the cells responsible for the effects of Nod are bone marrow-derived cells, not stromal cells. Furthermore, we show that Nods stimulate IL12-p40, IFN-γ, iNOS and NO expression, and that these factors are key for surviving the infectious challenge. Since the TLR/MyD88 pathway is also critically involved in detecting and eradicating C. pneumoniae, our data highlight an emerging theme in host defenses: that divergent pattern recognition receptors that are seemingly unrelated and expressed in distinct compartments can nevertheless direct cooperative responses that successfully combat invasion by common pathogens such as C. pneumoniae. Coordinated and sequential activation of TLR and Nod signaling pathways may be necessary for efficient immune responses and host defenses against C. pneumoniae. While TLRs might be important for initial activation upon Chlamydophila contact, it is likely that Nod proteins play a role in the sequential and intracellularly triggered prolonged activation of target cells by intracellular Chlamydophila.
Rip2−/− mice, backcrossed ten generation to C57BL/6, were kindly provided by Dr. Genhong Cheng (University of California at Los Angeles, Los Angeles, CA, USA). C57BL/6 mice and Nod2−/− mice were purchased from Jackson Laboratory. Nod1−/− mice were kindly provided by Dr. Jeffrey Weiser (University of Pennsylvania, Philadelphia, PA, USA). Mice were maintained under specific pathogen-free conditions, and were used at 8–12 weeks of age. All experiments were done according to Cedars-Sinai Medical Center Institutional Animal Care and Use Committee guidelines.
C. pneumoniae CM-1 (ATCC, Manassa, VA) was propagated in HEp-2 cells as previously described [18]. HEp-2 cells and C. pneumoniae stocks were determined to be free of Mycoplasma contamination by PCR. Mice were intratracheally infected with C. pneumoniae by inoculating 100 µl of PBS containing 1×106 IFU of the microorganism. Bronchoalveolar lavage fluid (BALF) was collected with 0.5 ml of PBS containing 2mM EDTA. The lavage fluid was centrifuged, and the supernatant was used for chemokine and cytokine measurements. The pellet placed on glass slides, and stained by modified Wright-Giemsa staining (Diff-Quick; Fisher Scientific, Pittsburgh, PA, USA) to determine leukocyte subtypes based on their cellular and nuclear morphology. Lungs were homogenized with 1ml of sucrose-phosphate-glutamate medium and stored at −80°C.
To quantify C. pneumoniae progeny, HEp2 cells were inoculated with lung specimens or cell lysates as previously described [75]. Briefly HEp2 cells were infected with diluted lung homogenates or infected cell lysates. Cultures were centrifuged for 1h at 800× g, fed with RPMI1640 in the presence of cycloheximide (1 µg/ml), and incubated for 72h. Thereafter, Cells were washed with PBS, fixed with methanol for 5 min at room temperature and stained with FITC-conjugated Chlamydia genus-specific mAb (Pathfinder Chlamydia Culture Confirmation System; BIO-RAD, Hercules, CA, USA) according the manufacturer protocol. Inclusion bodies were counted by fluorescence microscopy.
Lungs were fixed in formalin buffer, paraffin-embedded, and hematoxylin and eosin-stained sections were scored by a trained pathologist blinded to the genotypes as previously described [18]. Briefly, the degree of inflammation was assigned an arbitrary score of 0 (normal = no inflammation), 1 (minimal = perivascular, peribronchial, or patchy interstitial inflammation involving less than 10% of lung volume), 2 (mild = perivascular, peribronchial, or patchy interstitial inflammation involving 10–20% of lung volume), 3 (moderate = perivascular, peribronchial, patchy interstitial, or diffuse inflammation involving 20–50% of lung volume), and 4 (severe = diffuse inflammation involving more than 50% of lung volume).
The chemokine and cytokine concentrations in the BALF, lung homogenates or culture supernatant were determined using by Duoset Mouse KC, MIP-2 (R&D systems, Minneapolis, MN, USA), OptiEIA Mouse IL-6 ELISA Set (BD Biosciences, San Jose, CA, USA) and Mouse IFN-γ ELISA, Mouse IL-12 p40 ELISA (eBioscience, San Diego, CA, USA). The assays were performed as described manufacturer protocol.
The lymphocytic makeup in the lungs after infection were analyzed by flow cytometry of lung homogenates. Briefly, lymphocytes were isolated by digesting the lung tissue at 37°C for 1h with HANKS' containing 100 µg/ml Blenzyme (Roche Diagnostics, Indianapolis, IN, USA) and 50 units/ml DNase I (Roche Diagnostics) and filtering through a 70 µm cell strainer (BD Biosciences). Erythrocytes were depleted by lysis buffer before staining. Isolated single cells were stained with following specific mAbs; CD16/32 (clone 93), Gr1 (clone RB6-8C5), CD11b (clone M1/70), F4/80 (clone BM8), CD11c (clone HL3), CD45 (clone 30-F11), CD4 (clone RM4-5) and CD8 (clone 53-6.7) were purchased from eBioscience as direct conjugates to FITC, PE or PECy5. Anti SP-C polyclonal Ab and PEcy5-conjugated donkey anti-Goat IgG F(ab') were used for Alveolar type II epithelial (ATII) cell staining (Santa Cruz Biotechnology, Santa Cruz, CA, USA). Cells were identified based on expression of following antigens: pulmonary macrophages (F4/80+ and CD11c+), DC (F4/80− and CD11c+), Neutrophils (Gr1+ and CD11b+), ATII cells (SP-C+, CD45- and CD16/32-), T cells (CD3+), B cells (CD19+). For intracellular Chlamydophila staining, cells were permeabilized using Cytofix/Cytoperm kit (BD Biosciences) and stained with FITC-conjugated anti-Chlamydia LPS mAb (Accurate Chemical and Scientific Corporation, Westbury, NY, USA). Flow cytometric analysis was performed by FACScan flow cytometer (BD Biosciences) and the data was analyzed by Summit (Dako, Carpinteria, CA, USA).
Total RNA was extracted from homogenized lung tissues by RNeasey mini kit (QIAGEN, Valencia, CA, USA) following the manufacturer's protocol. Total RNA preparations were subjected to reverse transcriptase-polymerase chain reaction analysis by Total cDNA was generated using the Omniscript cDNA synthesis kit (Qiagen), PCR analysis was performed using specific primers for mouse iNOS (sense: 5′-TGG GAA TGG AGA CTG TCC CAG-3′:antisense: 5′-GGG ATC TGA ATG TGA TGT TTG-3′), 1min at 94°C , 1 min at 58°C and 2 min at 68°C. Amplification of GAPDH served as a control.
Femora and tibiae of mice were rinsed with cell culture medium. Bone marrow cells were treated with red blood lysis buffer (eBiosciences), cultured in RPMI1640 medium containing 10% FBS and 10 ng/ml M-CSF (R&D system). Medium changed at day 3 and day 6. BMDM were harvested at day 9 and exposed to C. pneumoniae by centrifugation at 500× g for 30 min.
Nitrite levels in the culture supernatant were determined using the colorimetric Griess reaction (Sigma, St. Louis, MO, USA). Absorbance was measured with a plate reader at 540 nm. The concentration of NO2− was determined from standard curves constructed with serial concentrations of NaNO2.
Recipient WT (Ly5.1), WT (Ly5.2) and Rip2−/− (Ly5.2) mice were lethally γ-irradiated with 950 rads using a 137Cs γ-source and were reconstituted intravenously with 5×106 BM cells derived from respective donors 2–3h later. All mice were placed on Baytril (Bayer HealthCare LLC, Shawnee Mission, KS, USA) for 2 weeks following irradiation. 6–7 weeks after engraftment, mice were tested by FACS analysis with FITC-conjugated Ly5.2 Ab (clone 104, eBiosciences) and PE-conjugated Ly5.1 Ab (clone A20, eBiosciences) staining for chimerism.
Data are reported as mean values±S.D. Statistical significance was evaluated by Student's t test. In the case of survival study, Statistical significance was evaluated by Fisher's exact test. For multiple comparison test, Statistical significance was evaluated by one way ANOVA with Tukey's post-hoc test.
|
10.1371/journal.pcbi.1002071 | Generation of Diverse Biological Forms through Combinatorial Interactions between Tissue Polarity and Growth | A major problem in biology is to understand how complex tissue shapes may arise through growth. In many cases this process involves preferential growth along particular orientations raising the question of how these orientations are specified. One view is that orientations are specified through stresses in the tissue (axiality-based system). Another possibility is that orientations can be specified independently of stresses through molecular signalling (polarity-based system). The axiality-based system has recently been explored through computational modelling. Here we develop and apply a polarity-based system which we call the Growing Polarised Tissue (GPT) framework. Tissue is treated as a continuous material within which regionally expressed factors under genetic control may interact and propagate. Polarity is established by signals that propagate through the tissue and is anchored in regions termed tissue polarity organisers that are also under genetic control. Rates of growth parallel or perpendicular to the local polarity may then be specified through a regulatory network. The resulting growth depends on how specified growth patterns interact within the constraints of mechanically connected tissue. This constraint leads to the emergence of features such as curvature that were not directly specified by the regulatory networks. Resultant growth feeds back to influence spatial arrangements and local orientations of tissue, allowing complex shapes to emerge from simple rules. Moreover, asymmetries may emerge through interactions between polarity fields. We illustrate the value of the GPT-framework for understanding morphogenesis by applying it to a growing Snapdragon flower and indicate how the underlying hypotheses may be tested by computational simulation. We propose that combinatorial intractions between orientations and rates of growth, which are a key feature of polarity-based systems, have been exploited during evolution to generate a range of observed biological shapes.
| How do genes control the growth of cells into complex tissue shapes such as flowers, wings or hearts? A key requirement is that genes must be able to modulate growth along particular directions. Two mechanisms have been proposed for how this may work; one based on the directions of mechanical stresses in the tissue and the other on molecular signals that propagate and provide local polarities. Here we show how a polarity-based system has the advantage of being able to act in combination with growth rates to generate a wide range of shapes. By applying this system to the development of the Snapdragon flower, we show, by comparison of computational simulations with actual flower development, how a simple set of polarity controls may underlie the formation of complex biological structures.
| Although there have been many experimental and theoretical studies on patterns of gene activities and their establishment in animals and plants [1]–[6] much less is known about how patterns of activity are linked to tissue growth and deformation. Addressing this problem represents a challenge because final form is usually not a direct readout of locally specified properties, but depends on mechanical constraints from neighbouring regions. For example, if the margin of a leaf has a higher specified growth rate than the centre, a wavy edge will emerge. The wavy edge is not directly specified but is a feature that emerges through the interaction between patterns of specified growth and the mechanical constraints of tissue continuity [7]. In such cases we may distinguish between specified growth, which is the growth that would be attained if each region grew independently of its neighbours (i.e. in mechanical isolation), and resultant growth, which is the growth observed when mechanical constraints of neighbours are taken into account (i.e. mechanically connected tissue). In animal systems a similar distinction is made between an imposed active deformation, and an elastic passive deformation [8].
Resultant growth can be measured experimentally by tracking tissue deformations over time [9]–[13]. However, to understand the mechanisms by which resultant growth arises we need to know how genes influence specified growth. Where specified growth is isotropic, genes need to control a single parameter, the local rate of growth. However, in many cases specified growth may be anisotropic requiring orientations as well as rates of growth to be under genetic control. Controlling orientations of growth requires a local axis to be defined (i.e. axiality, represented as a field of lines). In this respect growth is similar to stress which also has axiality. This similarity has led to the suggestion that stresses provide the primary cues for orienting growth. According to such a stress-based axiality mechanism, gene activity influences stresses in the tissue, the orientations of which are transduced to influence molecular properties of cells such as the cytoskeleton. These in turn modulate growth orientations which may further feed back to influence the pattern of stresses [14]–[18]. Recent support for such mechanisms in plants have come from studies of the effect of stresses on microtubule patterns [19].
A different way of specifying orientations of growth is through differential concentrations of signalling molecules. The varying concentrations define a local (cellular) polarity which includes both axiality and directional components (represented by a field of arrows). The axiality component is then used to orient growth. In this polarity-based axiality system, genes influence the distribution of signalling molecules which define a coordinated field of polarities. Incorporation of mechanical constraints then leads to resultant growth, which may feed back to influence, for example, tissue polarity orientations. In support of this system, there is considerable evidence that polarity is prevalent in biological tissues and may modulate growth [20]–[24]. For example, planar cell polarity (PCP) systems have been described in animals and implicated in processes such as growth of wings in Drosophila and convergent-extension in vertebrates [25]–[27]. Similarly, the polarised distribution of auxin transporters (PIN molecules) has been shown to be important for outgrowths of primordia in plants [28].
Elements of the stress-based axiality system have recently been modelled [29], [30]. Here we describe a framework and software implementation for the alternative polarity-based axiality approach, which we call GPT-framework. ur software, called GFtbox, is a MATLAB application available from http://www.uea.ac.uk/cmp/research/cmpbio/GFtbox. This framework was developed with plant growth in mind, although it may also be useful for modelling animal systems where cell movement is limited. In accompanying papers we show how a biologically relevant model can be derived using the GPT-framework [31], [32]. This model of Snapdragon flower development is constrained by a range of experimental data including gene expression patterns, mutant phenotypes, clonal analysis, growth dynamics and changes in geometry. It provides a working hypothesis for how growth is specified and shows how reorientation of growth can account for key observations. In this paper we explore a series of simplified models which illustrate how growth and polarity may interact combinatorially during morphogenesis to generate a wide range of forms. The results highlight the value of being able to specify orientation independently of stresses in the generation of complex tissue shapes. In addition, we provide the theoretical foundations on which our modelling depends.
Modelling the genetic control of tissue growth requires the incorporation of gene regulatory networks and signal propogation within a growing, mechanically connected, tissue. In the GPT-framework, tissue is treated as a continuous sheet of material with two surfaces and a thickness, here termed the canvas. Biologically, the canvas may correspond to a sheet of cells, single cells or subcellular components (e.g. walls). Regulatory factors are distributed over the canvas and may interact and propagate, allowing particular patterns and local polarities to be specified. Regulatory factors can be classified into two types. Identity factors do not propagate within the canvas, while signalling factors can. The regulatory factors specify a growth tensor field which describes the specified rates of growth parallel and perpendicular to the local polarity. Elasticity theory is used to compute the resultant deformation of the canvas. This deformation modifies the relationships within the canvas and thus feeds back to influence the regulatory factors. Our implementation (GFtbox) is specialised towards tissues that grow as sheets, such as petals or leaves, but the basic concepts are also applicable to bulk three-dimensional and flat two-dimensional tissues.
In the results we study the interactions between tissue polarity and differential growth in the generation of shape through a series of models. For convenience each example has a setup phase during which the shape of the initial canvas and distribution of regional identities and signalling factors is established, and three components that form the model. (1) A Polariser Regulatory Network (PRN) controls the activity of various organisers from which tissue polarity information propagates. There are two types of organiser, termed organiser and organiser. As a convention, we show polarity pointing away from organisers, and towards organisers. Polarity propagation is implemented through a signalling factor called POLARISER (POL), the gradient of which defines local polarity. The PRN controls production and degradation of POL at organisers that anchor the polarity. POL may also be produced and degraded at a background rate throughout the canvas. (2) A gene regulatory network (GRN) controls the activity of identity or signalling factors encoded by genes. (3) A growth rate regulatory network (KRN) determines how identity or signalling factors influence specified growth rates parallel to, , and perpendicular to, , local polarity. The KRN also specifies the growth in thickness, . The specified growth rates for a region of the canvas are equivalent to the growth that would arise without the constraints of surrounding material (see Methods).
In the first time step the specified growth field is applied to the initial canvas which may then distort through mechanical interactions in the continuous material (modelled according to elasticity theory, see Methods). The result is a slight deformation of the canvas (resultant growth field) that takes the regions of identity factors with it. Where a region containing an identity factor expands, that region inherits the properties of the parent region, so maintaining boundaries. In such new volumes, the concentrations of signalling factors are interpolated from the parent surrounding regions and then further adjusted according to their production, dilution, propagation and decay rates. The deformed canvas and expression pattern provides the starting point for the next time step and the sequence is reiterated. To verify the computational correctness of GFtbox, results were computed for several situations where analytical solutions are possible (see Text S1). In the following we explore combinatorial interactions between polarity and growth through a series of simple cases. We first consider deformations in 2D.
A key feature of the polarity-based axiality system is that orientation and growth rates can be specified independently and then combined in various ways. This combinatorial aspect is unlike the stress-based axiality system where orientations can only be specified once stresses have been generated in the tissue. These stresses will depend on the pattern of specified growth rates and the geometry of the tissue. To illustrate the combinatorial interactions within a polarity-based axiality system we first model simple anisotropic growth (Case A) and differential isotropic growth (Case B) separately. We then combine them in different ways (Cases C-I). We use an initially square canvas marked with black discs (simulating cells that produce marked clones) and a grid to show the geometrical transformations [33]. In all Cases the total areal increase (accumulated growth) is the same. The state of the canvas before and after growth is illustrated in Figure 1 for each Case.
Case A: Uniform polarity field with spatially uniform anisotropic specified growth rates. A gradient of POL is established during the setup phase through two organisers, ( organiser ) and (organiser ) at the bottom and top boundaries respectively. The PRN involves promoting production of POL while promotes degradation of POL, forming a proximodistal gradient of POL (arrows). After the setup phase the POL gradient is frozen (fixed to the canvas so that the gradient deforms with the canvas). An identity factor is expressed uniformly throughout the canvas. The KRN is (the value of is indicated by the intensity of orange). The resultant growth transforms the square into a vertically stretched rectangle. The black discs become vertically oriented ellipses. The specified growth pattern underlying this transformation is straightforward to implement using the polarity-based axiality system. By contrast, an stress-based axiality system would require an additional step that generates vertically oriented stresses and thus an additional deformation. Moreover, the pattern of stresses would need to be maintained during growth unless there was a mechanism for fixing the axiality.
Case B: Spatially varying isotropic specified growth rates. Differential growth is achieved by promotion of specified growth rates towards the right side of the square. This involves establishing an identity factor during the setup phase that is most strongly expressed along the right edge from where it declines gradually. The KRN involves promoting the specified growth rates equally in all directions (, ). This leads to a gradient of locally isotropic specified growth that increases from left to right. The overall result is a curved fan. Curvature is not directly specified but arises through differential growth and mechanical constraints inherent in the canvas.
Case C: A combination of Cases A and B: uniform polarity field with spatially varying anisotropic specified growth rates. The PRN and KRN are the same as in Case A while the pattern of is the same as in Case B. That is, specified growth rate is oriented parallel to the POL gradient and increases towards the right. The result is a convex fan with much stronger curvature than Case B. Thus anisotropic specified growth, which on its own produces no curvature (Case A), reinforces the curvature arising through differential growth. In principle this reinforcement may arise from two causes. 1) Because there is no , the gradient in is greater than in Case B. 2) Because polarity is local, the directions of specified growth rotates with the canvas, enhancing curvature. To separate the contributions of these two components, we fix the direction of specified growth by using an external (global) frame of reference, as shown in Case D.
Case D: A combination of Cases A and B but using an external field to specify growth orientations. The gradient of POL is determined by an external frame of reference (y axis) instead of being embedded in the tissue. Biologically, external polarity information could be obtained from, for example, the effect of gravity. The result is a fan with reduced curvature compared to Case C. Note that ellipse orientations still deviate from the vertical because, even though growth is specified to be vertical, at each step mechanical constraints force the canvas to curve. The enhanced curvature of Case C over Case D reveals the contribution of orientations being specified internally (2) rather than externally (2). Another way of reducing curvature is by using a local polarity field that re-adjusts dynamically as the structure grows, as will be shown in Case E.
Case E: The same model as Case C but allowing POL to continue diffusing rather than being frozen after the setup phase. As with Case D, the resulting curvature is less than Case C, particularly near the extremities. This is because growth orientations turn less near the extreme positions of the canvas.
The previous Cases considered uniform polarity fields with differential growth. This raises the question of how non-uniform polarity fields may influence shape. We first consider these when combined with uniform growth rates (Cases F and G, Figure 2) and then with differential growth rates (Cases H and I, Figure 2).
Case F: Similar to Case A but setting a spatially varying polarity field. A gradient of POL is established during the setup phase by , which is expressed along the horizontal midline, and , which is expressed in the top, bottom, and right edges, increasing toward the right corners. The resulting POL gradient is shown by the arrows. The polarity field is frozen (fixed to the canvas) after the setup phase. The distribution of is spatially uniform as in Case A. Growth at the top and bottom edges is oriented by the organisers producing a strongly concave right edge. Thus, curvature is generated as a result of non-uniform specified orientations of growth. The curvature is even stronger if the polarity field is not frozen after the setup period, as shown in Case G.
Case G: Similar to Case F but allowing POL to continue diffusing rather than being frozen after the setup phase. The result is more concave than Case F. This is because of feedback between canvas geometry and the polarity field. We now look at the effect of introducing differential growth rates.
Case H: Similar to Case F but with a gradient of specified growth rate. The PRN is the same as Case F leading to a polarity field pointing to the right corners. The KRN and distribution of are the same as Case C leading to increasing values of towards the right edge. The result is intermediate between Case C and Case F because the diagonally specified growth orientations counteract the curvature induced by differential growth. Thus, unlike Case C where local specification of orientation reinforces tissue curvature, here it antagonises curvature. This effect is still stronger when the POL gradient is not frozen as shown in Case I.
Case I: Similar to Case H but allowing POL to continue diffusing rather than being frozen after the setup phase. The right edge grows to be almost vertical showing that an appropriate specified local polarity can antagonise curvature arising from differential growth (Case B).
The main conclusion to emerge from Cases A to I is that the ability to combine specified growth rates with separately specified orientations provides an effective control mechanism for generating shape transformations. The shapes that emerge reflect interactions between specified orientations, differential growth and mechanical constraints. Depending on the spatial distribution of organisers and the dynamics of polarity propagation, tissue polarity can reinforce or antagonise curvatures resulting from differential growth or may generate curvature even in the context of uniform growth. So far we have only considered combinatorial interactions within the plane of the canvas. We next consider deformations out of the plane.
We again consider a series of simplified cases (Figures 3 and 4) in which polarity and differential growth are treated separately (Cases J, K, O) and in combination (Cases L, M, N, P, Q). In each Case the up-down symmetry is broken by the centre of the initial canvas being slightly bowed upward. To simulate the presence of tissue beyond the boundaries of the initial canvas, the edges of tissue are prevented from moving or rotating out of the plane.
Case J: Spatially varying specified orientation with a uniform areal growth rate. The PRN involves an organiser (), expressed in the middle of the canvas (blue). An outer region is defined by which keeps POL levels at zero. This leads to a divergent polarity field near the centre (arrows). POL continues to diffuse after the setup phase (i.e. the gradient is not frozen). is spatially uniform as in Case A. The KRN involves anisotropic growth in the polarised region ( and ). By default, growth is isotropic where the POL gradient is zero (). The result is a small spike. As with Case F tissue curvature has arisen through variations in specified growth direction even when areal growth rate is uniform. However, in Case J the curvature occurs out of the plane as well as in the plane.
Case K: Spatially varying isotropic specified growth rates. During the setup phase is established in the centre of the canvas from where it declines in a graded fashion. As with Case B, the KRN has setting the specified growth rates, , . This leads to a gradient of locally isotropic specified growth rate that increases towards the centre. The result is a puffball-like central bulge exhibiting curvature both in and out of the plane of the canvas. The rounder shape compared to Case K illustrates the limitations of isotropic specified growth in creating elongated outgrowths. However, by combinging Cases J and K the outgrowth can be further exaggerated as shown in Case L.
Case L: Spatially varying anisotropic specified growth rates (combining Case J and K). The PRN and KRN are the same as Case J leading to radially directed growth. The distribution of is the same as Case K leading to increased anisotropic specified growth towards the centre. The result is a tall central spike with a sharp tip showing how differential growth and anisotropy act in combination. In many biological structures, such as a growing plant apex, protrusions have rounded tips rather than sharp points. This can be achieved by reducing growth in the central region, as shown in Case M.
Case M: Spatially varying anisotropic specified growth rates with a central region of no growth. This is similar to Case L, except that additional identity factor sets and to zero in a small central region. The final shape is a rounded projection similar to what might be observed in a plant apex. Such a model is also consistent with the observation that growth rates tend to be lower in the central region of plant apices [12].
We conclude that a range of outgrowths can be readily obtained by combining specified growth rates and orientations. As for the 2D cases, deformations lead to changes in orientations of the polarity field which feed back to influence further deformations. So far we have considered the effects of uniform and divergent polarity fields. A further elaboration is to combine these two as illustrated in Figure 4.
Case N: A uniform polarity field with spatially varying anisotropic specified growth rates (combination of Cases A and K). The PRN and KRN are the same as in Case A leading to a left-right polarity field and anisotropic growth. The pattern of is the same as in Case K leading to enhanced growth in the centre. POL continues to diffuse after setup. The result is a thin bulge with grooves at each end. As with Case C, the polarity field is acting as a modulator rather than generator of curvature (no curvature is produced by the polarity field when combined with uniform anisotropic growth, Case A). We next look at the effect of combining the polarity fields in Cases A and J.
Case O: Interacting polarity fields with spatially uniform anisotropic specified growth rates (combination of Cases A and J). The PRN and KRN are the same as Case A except that the additional organiser from Case J is included. The new organiser distorts the polarity field shown in Case N inducing a saddle point upstream. As result following growth, the canvas widens slightly in the centre and forms a central ripple. Thus, as with Cases F and J, some curvature arises even with uniform specified areal growth rates. Next we combine this polarity field with centrally increased specified growth rates.
Case P: A combined polarity field with spatially varying anisotropic specified growth rates (combination of Cases K and O). The PRN and KRN are the same as Case O while the pattern of is the same as Case K. The result is an asymmetric spur reflecting the interactions between tissue polarity and growth. The asymmetry arises because the POL gradient generated by the central organiser flows in the same direction as the background POL gradient on one side but in the opposite direction on the other, creating a region of counterflow (arrowed). Disorganisation of growth in the counterflow region reduces growth along the main axis of the tissue. Asymmetry induced in this way is a feature of simple polarity-based axiality systems that would not occur in simple stress-based axiality systems. The orientation of the spur can be reversed by using a organiser instead of a organiser in the centre as shown in Case Q.
Case Q: A combined polarity field with spatially varying anisotropic specified growth rates. The PRN and KRN and the pattern of are the same as Case P except that the central organiser is replaced by a organiser. This time the asymmetric spur points in the opposite direction to Case P because the counterflow region is on the other side.
We conclude that combining polarity fields provides a further richness by generating asymmetries. The above Cases illustrate some basic combinatorial interactions between polarity and growth. To see how the same principles may apply to a biological example, we consider a simplified model of the Snapdragon corolla tube.
To simplify the Snapdragon tube we assume the initial canvas comprises an initial cylindrical canvas closed at one end. As a first step we study locally isotropic specified growth (Case R) and then explore the effect of introducing specified anisotropic growth (Cases S and T).
Case R: Spatially varying isotropic specified growth rates. An early step in the development of the Snapdragon flower is arching over of the tube through differential growth. We simplify this process by restricting growth rates in opposite regions of the cylinder and also at the base. This is achieved by having a general background level of which is inhibited in the base by and is also inhibited by a diffusing signal which is generated along opposite sides of the cylinder by . In this Case, specified growth is isotropic, .
The result is a ballooned out bowl (Figure 5B,C) rather than an arched over tube. Some areas of the canvas near the base show anisotropic resultant growth, evident from elongated ellipses. This is shown more clearly Figure 5 C where the principal directions of resultant growth () are shown with short lines and the rate of anisotropic growth () is shown in magenta. As with curvature, resultant anisotropy is not specified directly but arises through the interaction between differential growth and mechanical constraints. However, the pattern and extent of resultant anisotropy is inconsistent with experimental observations of clones in the Snapdragon tube, which are highly elongated along the proximodistal axis [31]. To address this discrepency we introduce specified anisotropic growth through a polarity field as shown in Case S.
Case S: A uniform polarity field with spatially varying anisotropic specified growth rates. A gradient of POL is established through two organisers, and - at the base and rim respectively (Figure 5A, arrows). The KRN is the same as Case R except that the specified growth rate is now anisotropic, and . Compared to the output from Case R, the sides of the cylinder curve towards each other rather than ballooning outwards (Figure 6 B). Thus, introducing specified anisotropy has a major effect, leading to a more closed shape. It also generates much more elongated clones matching experimental observations. However, continuation of growth leads to the two sides arching further (Figure 6 C) rather than creating the elongated shape that is observed experimentally. (In our implementation which does not currently include collision detection the two sides grow through each other. For clarity we therefore only show one side in Figure 6.) To address this discrepancy, we exploit the potential to reorient growth within the GPT-framework by modulating the polarity field as shown in Case T.
Case T: Initially the same as Case S followed by reorientation of tissue polarity. There are two phases of growth, early and late. During the early phase the cylinder grows as in Case S. At the start of the late phase, the polarity field is modulated by restricting the spatial region of the organiser. This is achieved by activating an identity factor in the lateral regions of the cylinder which inhibits , restricting the distal organisers (cyan) to small regions at the apex of each arch (Figure 6D,E). The reoriention of polarity leads to vertical elongation of the arch rims, maintaining the closed shape, rather than the sides continuing to arch over. This captures an essential feature of Snapdragon corolla tube growth.
We model growth through the accumulation of a series of small deformations of the tissue (canvas). Stresses are generated during the process as the canvas is mechanically interconnected. This may lead to anisotropic resultant growth even when growth is specified to be isotropic (e.g. Case R). In principle, such resultant stresses could be used, through stress-based axiality, to orient all forms of anisotropic growth. However, this would mean that specified orientations of growth would be dependent on differential rates of growth, precluding the possibility of independent control. By contrast, we show how a polarity-based axiality system allows diverse forms to be generated through combinatorial interactions between specified orientations and rates of growth.
In this system, a key element in controlling growth orientations is the distribution of polarity organisers. These are of two types, or , allowing polarity fields to be anchored at both ends. Even when specified anisotropic growth is uniform over the canvas, a range of forms can be generated by varying the pattern of organisers. For example, starting from an initial square canvas it is possible to generate rectangles (Case A), concavities (Case F), small spikes (Case J) and ripples (Case O). In these Cases polarity was fixed after a setup period. Biologically, this would correspond to an initial period when polarity propagates across the tissue (when the tissue is small), followed by polarised cells maintaining their polarity and passing it on to their daughters. Another possibility is that polarity continues to propagate during growth leading to slight modifications of the resulting shape (compare Cases F and G).
The range of shapes may be greatly extended by combining polarity fields with differential growth rates. For example, tissue polarity may reinforce or antagonise curvature arising through differential growth (Cases C and I). Both aspects are incorporated into the growing Snapdragon tube - reinforcement of curvature during the early phase leading to arching over (Case S), followed by antagonism of curvature leading to straightening (Cases T). It is also straightforward to generate extended outgrowths and apices by combining a single organiser with enhanced growth (Cases L and M). A further feature of polarity-based axiality systems is the emergence of asymmetries through interactions between polarity fields. For example, asymmetric spurs may arise because of counterflowing polarity on one side (Cases P and Q). The asymmetry of the outgrowths in these Cases results from the underlying polarity interactions and would not have arisen from a simple system with only stress-based axiality.
In these examples only a few organisers are needed to achieve major shape transformations. To test whether the same simplicity might underly more complex biological transformations, we modelled growth of the Snapdragon flower [31]. This model is constrained by a range of experimental data. The expression pattern of the genes DIV, CYC, DICH and RAD are set according to experimental observations. The model has to not only account for the wild-type phenotype but also double (cyc, dich) and triple (cyc, dich, div) mutants. The model is also constrained by the observed changes in 3D shape determined by optical projection tomography at several developmental stages. In addition the pattern of growth rates and directions over each model petal need to be similar to those observed by clonal analysis. The model starts with an initial cylindrical canvas with five lobes and a proximodistal pattern of polarities established through two polarity organisers ( and ) (Figure 7 A). During the early phase of growth the ventral region of the tube arches over through differential anisotropic growth. To account for the observed pattern of clones a third organiser () is introduced (Figure 7 C). In the absence of this organiser the tube bulges out (Figure 7 F) similar to what happens in the simplified corolla with no reorientation of growth (Case S). However, with the introduction of the organiser the tube automatically straightens out during later stages, consistent with experimental observations. Thus, this biologically relevant case provides evidence for three organisers underlying major shape transformations and growth dynamics. In the Snapdragon model, the reorientation of growth is under the control of DIV, a gene that encodes a Myb-like transcription factor that affects flower shape and symmetry [31]. As well as its effect on organiser activity, DIV also influences growth rates. Thus, although rates and orientations of growth are specified separately in the model they can be regulated by a common gene.
The polarity-based axiality system has the flexibility to account for global shape changes, observed growth patterns and clones without invoking large numbers of polarity organisers. This alone does not demonstrate the validity of invoking tissue polarity for the control of growth orientations. Nevertheless, tissue polarity is commonly observed in animals, for example, polarised cell movements [27] and in plants where the polar distribution of molecules within cells, such as PIN auxin transporters, suggests that cell polarity is also common [23], [24]. It has also been proposed that an auxin concentration maximum at a vascular boundary in the root tip establishes a distal polarity organiser in the root [20]. The GPT-framework allows hypotheses on polarity-based axiality growth to be established that can be subjected to further tests such as mechanical or genetic perturbations. The Snapdragon model, for example, was evaluated against predictions of shapes of multiple mutants not used to build the model [31], [32]. The results showed a good, quantitative, fit between predicted and observed shapes. The model also makes important predictions about the location of polarity organisers. Polarity markers are predicted to show reversals (i.e. arrows pointing away or towards each other) at these locations.
In all our Cases we make the simplifying assumption that the tissue is linear over small deformations and has isotropic material properties. An elaboration of the GPT-framework would be to incorporate non-uniform properties, although this would also require these properties to be measured across the tissue during growth. The GPT-framework is consistent with current hypotheses regarding the mechanisms in which plant tissue grows under turgor pressure through the loosening and formation of bonds (Theorems 1 and 2, Methods). Loosening bonds in the cell wall allows the tissue to grow. If new material is inserted that restores the properties of the cell wall then the residual strain returns to zero (‘snip and fill’ [31]). Biologically this would require some form of feedback between resultant stresses (or strains) and cellular properties [34]. Feedback from stresses to microtubule patterns has been proposed [19], and this can be interpreted as reflecting the need to dissipate residual stresses rather than being the primary way of orienting specified growth. Cutting provides a convenient experimental way to evaluate the extent to which residual stresses accumulate or dissipate in a given biological system. Often they accumulate in certain regions in later developmental stages. For example, the dorsal and ventral petals of the adult Snapdragon flower press against each other holding the flower shut (not a part of the model in Green et al [31]). The observation that the accumulation of residuals varies systematically from region to region suggests that the process of dissipating or accumulating residuals is under genetic control. Stresses that are accumulated can be modelled with the GPT-framework and, to enable direct comparison with experimental results, the resulting shapes can be cut allowing the structure to spring into a new shape.
The GPT-framework assumes that regions (e.g. cells) in a tissue do not slide or move past each other. This is valid for plants [35], making them particularly appropriate for this approach. The GPT-framework may also be applicable to some aspects of animal development. For example, finite element models have been used to capture deformations during Drosophila ventral furrow formation driven by apical constriction and apicobasal elongation of cells [8]. Comparable deformations can also be generated using GPT-framework by using a posterior-anterior polarity field [36] and incorporating negative growth (contraction) on one side of the canvas (Figure 8). Although this model does not incorporate all biologically relevant features such as constraints of the external vitelline membrane, it illustrates the flexibility of the approach.
Clones generated in early wing development of Drosophila often stay as contiguous patches, indicating that connectivity is broadly maintained and extensive mixing of cells does not occur [37]. Greater cell mixing is observed for clones in developing mammalian tissues such as the heart or limb, although even in these cases cell movements are not sufficient to disrupt formation of clonal clusters or patches[38]. At the tissue scale it may therefore be reasonable to model many animal structures with the framework described here, particularly as orientated cell behaviours are thought to play a critical role [39], [40].
As well as multicellular tissues the canvas could represent a region of a plant or bacterial cell wall. By extension of the GPT-framework it may also be possible to capture the growth of compartments enclosed by a canvas (e.g. cells with their walls) or growth of a bulk solid. Thus, the GPT-framework provides a general approach that can be applied to growing tissues at many scales.
The GPT-framework with its assumption of tissue polarity as a key component of growth specification provides an economical way of generating diverse shapes and forms. We hypothesise that this combinatorial richness is not only computationally attractive but has also been exploited during evolution to generate a range of observed biological shapes.
Various mathematical and computational methods [21] have been used to model tissue growth. These range in scale from detailed modelling of individual sections of cell wall to larger scale models treating the tissue as a continuous substance. The physical properties have been studied in terms of mass-spring models, elasticity theory of thin shells, and elasticity theory for solid volumes. Elasticity theory described here subsumes both classical linear elasticity theory and elastoplastic or viscoplastic theory for modelling solid flow.
In mass-spring models tissue is represented as a set of point masses linked by springs. De Boer [41] combines mass-spring modelling with the L-system formalism of [42] to describe a two-dimensional model of cellular growth. In these models, and in those of [13], [43], [44], the springs correspond to sections of cell wall, and the masses are where three or more springs meet. Growth is modelled by changing the resting length of the springs. The new equilibrium configuration is then computed by iteratively finding a state of minimum energy. There are empirically-based rules for deciding when cells should divide. These models are mainly limited to two-dimensional problems, although they have also been used to model model axisymmetric three-dimensional solid problems such as root tip growth. A problem with mass-spring modelling of continuous tissue (i.e. above the cellular scale) is that it is not trivial to design the model so that on a large scale, realistic elastic properties emerge. For example, a regular grid of springs is not geometrically isotropic.
For tissues which take the form of curved surfaces, thin in comparison with their extent, one can use thin shell theory (c.f. sheets of cells [45]). This is the branch of elasticity theory dealing with the mechanics of curved surfaces [46]–[48]. It is the limit of three-dimensional bulk elasticity theory as the thickness of the sheet tends to zero while retaining its bending stiffness properties. For surfaces which are extremely thin in comparison to their area, this has advantages for numerical computation over describing them by the methods of solid volume elasticity theory. The rippled edges of leaves have been modelled by this method as the mechanical consequence of faster growth at the edges [49], [50]. (Cf. Text S1, Case 14 and Video S8 [7], [51], [52].)
A third approach is to model biological structures as three-dimensional solid objects [19], [53]. This can be appropriate when tissue thickness is sufficiently large to make the thin shell approximation unnecessary. The method is analysed theoretically by Goriely and Ben Amar [54], who consider the general problem of describing the growth of elastic substances resulting from local growth fields and, by alternating a phase of growth without movement (that is, insertion of new material) over a small time interval and then allowing elastic relaxation, they show how growth over an extended period of time can be modelled. The net result is a visco-plastic deformation. It is this approach that is taken in the GPT-framework, and it has been extended to model both the extent and orientation of anisotropic growth.
The following theory covers the local specification of growth, how to compute the resulting growth given the mechanical properties of the canvas, how to handle residual growth, and how modelling using the GPT-framework relates to modelling growth in terms of turgor pressure and modifications to the mechanical properties of the cell walls.
We distinguish two types of growth, specified and resultant. Resultant growth is the growth that can be directly observed by tracking or clonal analysis. Specified growth is the growth that would happen to an element of the canvas if it grew in isolation. Resultant growth emerges as result of specified growth in different regions interacting through connected tissue. This is illustrated in Figure 9. Panel (A) shows the initial state of a square tissue, divided into a number of small tiles. If we apply a radially increasing field of locally isotropic growth, then in (B) we have an exploded view of how this would affect each tile individually, if it were not attached to its neighbours. It is clear that without some further deformation, these tiles cannot fit together into a continuous tissue without gaps. This conflict between the specified growth field and the continuity of the tissue leads to an equilibrium compromise between the two shown at (C). It is mathematically determined by the partial differential equations of elasticity theory, and numerically computed by the finite element method, both of which we shall briefly summarise.
Suppose that at a given time, each point at position in a tissue is moving with an instantaneous velocity . The resultant growth rate in the neighbourhood of is the gradient of the velocity field with respect to . This is the second rank tensor field (a two-dimensional matrix at each point of the tissue) whose components are , where and range from 1 to 3. This velocity gradient tensor represents both the change of shape and size and the rigid rotational motion of the material in the neighbourhood of the point . These are respectively its symmetric and skew-symmetric parts: , where and . is called the resultant strain rate tensor field, and the resultant vorticity. The vorticity field describes the angular velocity at each point. When the vorticity component of a tensor field is zero, the field is called irrotational. To avoid subscripts we abbreviate the definition of to , where is the differential operator defined by . The rate of resultant growth of the material in any particular direction is the sum .
Because the resultant strain rate tensor at a point is symmetric, it can be diagonalised by suitably rotating the local frame of reference. The resulting three diagonal components are the principal rates of resultant growth, in three perpendicular directions. These are the eigenvalues of , and the principal growth directions are parallel to its eigenvectors. The growth directions and rates will in general vary over the tissue.
To explain how resultant growth may be calculated from specified growth, it is convenient to think in terms of small displacements rather than velocities, by considering the effect over a small time . This is also how the computational implementation (to be discussed below) works, iterating through time in small steps . “Small” here means small enough that first-order approximations apply. In time a velocity field produces a small displacement field , and a growth rate or strain rate tensor field produces an amount of growth or strain, which we shall denote by the same symbols as before.
At each point in the growing canvas, let be a specified strain tensor at that point, being the product of a strain rate tensor by a small time . This is the growth that would occur in a small region around in time if it were mechanically isolated from the rest of the tissue. Let be the displacement field that will result from this pattern of growth if the tissue remains in mechanical equilibrium, and the associated growth tensor field. Except in some special cases, such as uniform isotropic growth, will differ from . Even if the rotational component of is ignored, its strain component will still in general differ from : there may be no displacement field of which is the strain field. This is due to the constraint of physical continuity that we mentioned above. (For clarity, the amount of growth shown in Figure 9 has been made far greater than we would normally compute in a single time step.)
Physical continuity is expressed mathematically by the St. Venant compatibility constraints [55]. If is a strain field of the form , then it necessarily satisfies the following partial differential equation:This can be verified by substituting for and (somewhat laboriously) finding that all of the terms in the resulting sum of third derivatives of components of cancel out. It is a deeper result that the St. Venant conditions are sufficient for such a velocity field to exist.
If, on replacing by in the above equation, it fails to hold, then whatever deformation is applied, the material must remain in a state of frustration. There will be unrelieved residual strain given by . When the material is in mechanical equilibrium, the displacement field will be such as to minimise the energy contained in that residual strain. To calculate , we use the principle of virtual work: if the material is in equilibrium, and any additional infinitesimal displacement is applied, then it will do zero work against the stresses in the material ([56], ch. 2).
These stresses are given by a tensor field calculated from the strain and the elasticity properties by the constitutive equation of the material:(1)The subscripts all range over the spatial dimensions 1–3. is the elasticity tensor or stiffness tensor, a 4th rank tensor field representing the elasticity properties of the substance [57]. The work done by any small strain against any stress is , and the total work done for strain and stress fields is found by integrating this over the whole tissue.
This is the linear elastic constitutive model, which we are assuming to be valid for small strains. For some biological tissues this assumption may not be accurate, for example as noted in [58] for the mouse ventricle, which also notes that determining a more accurate constitutive model is experimentally challenging.
To avoid writing explicit summations, we shall adopt the notations that if and are second rank tensors and and are fourth rank tensors, then:
The work done by the strain against the residual stress is then:(2)where the integration is over the whole volume. For to be the equilibrium deformation we must have:(3)
Except for degenerate situations (such as the initiation of buckling [59]), this determines up to a rigid translation or rotation of the whole object.
We have omitted from equation 3 the possibility of external forces acting on the substance, since there are no such forces present in the applications used in this paper and the Snapdragon model [31]. Boundary conditions can also be applied which stipulate that some parts of the substance remain stationary. We describe how these are handled when we discuss numerical methods.
Both the specified growth field and the resultant strain field are by definition irrotational. However, the resultant growth field in general does include rotations. Leaving aside rigid rotations of the whole tissue, the relative rotations between different parts of the tissue are entirely determined by the irrotational tensor . That is, relative rotations are caused solely by differential local growth and the continuity constraints, not by any explicit specification: rotations are always resultant, never specified.
Since , the whole analysis carries back to the description in terms of velocities, strain rates, and growth rates.
In plants, specified growth rates are always positive, but in animal tissue this is not always so. Both positive and negative growth rates in any direction can be handled computationally without difficulty. Figure 1 shows a simple model in which the shape changes with negligible change of volume.
The residual strain is given by the tensor , which is the symmetric part of the residual growth tensor . Most of the examples in this paper discard the residual strain after each time-step of the simulation. In biological terms this is consistent with the observations of [19] that imply a feedback mechanism that acts to absorb stresses. To illustrate the effect of discarding or retaining residual strains we consider several cases in which we cut the canvas after growth or constrain the canvas during growth and then release the constraint. We contrast the effect of discarding residual strains (Cases U and V) with accumulating strains (Cases W and X). These are illustrated in Figure 10.
Case U: Dissipation of residual strain with a non-uniform pattern of growth followed by cutting. This is identical to Case C in which residual strain is dissipated at each step. As expected, cutting the canvas induces no further changes of shape as there is no accumulated residual strain. We next show the result of constraining the canvas so that it cannot grow followed by releasing the constraint.
Case V: Dissipation of residual strain with a non-uniform pattern of growth that is constrained, followed by release. The model is the same as Case C except that all the boundary points are fixed during growth (column 2). When these constraints are released the shape does not change (column 3) as there is no accumulated residual strain. We next consider the effect of accumulating the residual strain.
Case W: Accumulating residual strain with non-uniform pattern of growth followed by cutting. The model is the same as Case C except that residual strain is accumulated at each step. The result is very similar to Case C but there is a small accumulated residual strain (column 2, blue). Cutting and allowing the canvas to relax releases some of this accumulated strain leading to a curve along the line of the cut compared to the straight line in Case U.
Case X: Accumulating residual strain with non-uniform pattern of growth that is constrained, followed by release. The model is the same as Case V except that strain is accumulated during growth (blue shows accumulated strain). Releasing the constraints allows a shape to emerge similar to Case W uncut.
We may also illustrate the effect of retaining residual strain with a 3D example. For this we use the simplified Snapdragon tube (Case T), but allow residual strain to accumulate on one side. This is illustrated in Figure 11.
Case Y: This is the same as Case T, except that residual strain is retained on the right side. Figure 11 B shows how the resultant shape of the right side differs from the left. The residual strain is shown in blue (Figure 11 D). A further difference between the two sides is revealed by making vertical cuts and allowing the mechanical system to relax to a new geometry (Figure 11 C,E). As expected, cutting makes little difference to the left side as there is little residual strain. However, the right side springs apart revealing some of the stored residual strain.
In the above examples the release of residual strain by cutting involves large displacements and rotations of the material. However, our computational methods are based on the linear elasticity theory of small displacements, and never directly solve large-displacement problems. The deformation resulting from the release of residual strain is, therefore, computed incrementally, by iteratively applying a small fraction of the residual strain, computing the resulting small deformation, transforming the remainder of the residual strain according to the new orientations of every part of the tissue, and repeating until an equilibrium is reached.
The stiffness tensor is a fourth-rank tensor, which in three-dimensional space has components at each point. However, it satisfies certain symmetry properties which imply that it has at most 21 independent components. For isotropic materials, further symmetries imply that is determined by just two values: the bulk modulus and Poisson’s ratio . is the ratio of applied pressure to relative change in volume. We will see later (see TextS1, Equation (S1) et seq.) that its value is irrelevant for the calculations we require: it cancels out of the equations. When a block of material is compressed by external forces in one direction, Poisson’s ratio is the ratio of its transverse expansion to its longitudinal compression. In practice lies between 0 and 0.5. As the value approaches 0.5, while the bulk modulus is held constant, the resistance to unidirectional stretching and compression decreases towards zero. If the limit is instead approached by keeping the shear modulus constant, then the bulk modulus tends to infinity. In the former case, the material’s resistance to shears vanishes and it approaches the state of a liquid, while in the latter it approaches an incompressible solid with a finite elastic resistance to everywhere volume-preserving deformations.
However, since there are no applied forces (such as gravity) in our models, but only growth described as a change in the resting shape of the material, the difference is more apparent than real. The elasticity tensor computed from (the shear modulus) and (Poisson’s ratio) is equal to the tensor computed from and multiplied by . As mentioned above, any such factor in the elasticity tensor cancels out (Text S1, Equation (S1)), because all of the forces that we consider result from the material acting against itself. Both methods of computing the elasticity tensor for any value of less than 0.5 give identical solutions to the equation, solutions which are independent of or . At exactly 0.5 the equations become highly degenerate, and a different analysis is required to calculate the physical behaviour in the limit. Any value above 0.5 is physically impossible for isotropic substances, as it would imply that the volume increased under compression, violating conservation of energy. Few experimental determinations of Poisson’s ratio for living plant tissues have been made. They range from 0.18 to 0.4 for onion epidermis [60]-[62]. We find that the growth behaviour of a model is insensitive to the precise value of (also see Case 6 in Text S1), and have generally set it equal to 0.3 in our simulations. In our current models, for simplicity we have taken the elasticity properties to be uniform throughout the tissue and over the time of its development. However, elasticity that varies over the tissue and over time can also be described using the GPT-framework.
The analysis so far has assumed that the deformations to be computed are always small. Growth by large amounts can be computed iteratively, by growing in a series of small time steps, in each of which the growth causes only a small deformation. The result is to produce a plastic flow of the material over large time intervals, computed by the theory of small deformations of purely elastic material.
Plant growth is thought to occur from a transient reduction in the stiffness of cell walls allowing them to stretch under turgor pressure, new material being added to restore the stiffness [16], [17], [29], [63]–[67]. When the process is anisotropic, it may be because the cell wall fibres typically have coherent directionality, or because the weakening is distributed non-uniformly over the walls of a cell. Most studies have simplified the process by assuming that it is equivalent to increasing the amount of material in a region and then relaxing the shape. This is also the approach used in the GPT-framework : the specified-strain model. The simplification avoids the need to measure relative stiffness and consider turgor pressure. We show below how this approach is related to turgor-based systems.
Suppose the tissue has stiffness tensor field , and turgor pressure field . As a result, it must be in some state of strain . The resulting stress in the tissue is . The condition for the tissue to be in equilibrium is that for any small displacement , . Now suppose we change the rest state. For any small piece of the tissue, its rest state is the shape it would be in if the turgor were removed (and the mechanical linkage to the rest of the tissue ignored). is the transformation from that state to the state that it takes up under turgor. Changing the rest state means applying a strain . (The minus sign is due to the fact that we want positive values of to model an increase in the resting size, but the effect of increasing the resting size is to put the tissue into a state of compression, which is described by negative values of strain.) When turgor is reapplied, the resulting strain is . (To validly add strains like this, we are assuming that all of the strains are small.) The applied strain will produce some equilibrium displacement field . The condition for the new equilibrium is . We can subtract from this the original equilibrium condition, leaving , our original equations (2) and (3). This means that the effect of a specified strain field is independent of the turgor, and we can ignore the turgor in our calculations.
The following two theorems explore the relationship between the method of specifying the strain and the method of modulating the stiffness tensor.
Theorem 1 Let a tissue have a stiffness tensor field , a turgor pressure field and a strain field , such that the tissue is in mechanical equilibrium. Suppose that the stiffness field is then changed to , where is small compared with .
Let a second tissue of identical geometry have a stiffness tensor field and be in equilibrium under a strain field .
Then there is a specified strain tensor field such that the deformation of the second tissue resulting from applying is the same as the deformation of the first tissue resulting from the change in stiffness .
The strain field can be split into two parts: the strain due to turgor, which is (where is the compliance tensor, i.e. the inverse of ), and a residual strain . The residual stress field in the first tissue is the residual strain multiplied by , which is . For the tissue to be in equilibrium in this state, the work done by any infinitesimal displacement field against the residual stress must be zero. This work is where the integration is over the whole tissue. Recall that is the differential operator that computes the strain tensor field of a displacement field.
If the stiffness is changed to , a new equilibrium configuration will be established by a displacement field . The residual stress field is then , and in equilibrium we have for all . Subtracting the previous virtual work equation gives . is equal to . If we assume that is small in comparison with and is small in comparison with , then the last term can be omitted as being of second order, leaving an effective residual strain of .
In the second tissue, the residual stress is initially , and equilibrium implies that is zero for all virtual displacement fields . When the strain is applied, it will produce a displacement field , and a residual stress . As for the first tissue, to determine the equilibrium value of we need only consider the effective residual stress .
To prove the theorem it is sufficient to find a value for such that when is taken equal to , the residual strains and are identical at every point. Thus we require to satisfy:
Since is invertible, its inverse being a compliance tensor , we can immediately calculate , where is the identity matrix. This proves the Theorem.
If the first tissue of this theorem is a biologically accurate description of an increment of growth in terms of the tissue’s background stiffness , turgor , and change in stiffness , then the theorem tells us that we can find another description in terms of a specified strain which gives the same deformation. Furthermore, we have a free choice of the background stiffness . In particular, we can choose to be uniform and isotropic, and constant over time. However, the relationship between and is somewhat complex. When using specified strain to model the result of growth by stiffness modulation, we would like to obtain a closer connection between and , which we now proceed to do.
Firstly, if we take , then the expression for simplifies to , and we need no longer calculate .
Now suppose that is orthotropic. That is, at each point there are three orthogonal axes such that the change in stiffness is symmetric under a half-turn about each of them. These are called the principal axes of . Under certain extra conditions, we find that the principal axes of coincide with those of . Thus the same distribution of polarisation can be used for either description of growth.
Theorem 2 Under the conditions of Theorem 1, suppose that the following conditions hold:
1. , , and are everywhere isotropic.
2. is orthotropic.
Then by taking , the principal axes of the specified strain given by Theorem 1 coincide with those of .
We have seen already that if , then . and are isotropic by the first condition. By the second condition, is the stress associated with the isotropic strain given the change in orthotropic material properties . Such a stress has the same principal axes as . Multiplying by the isotropic compliance leaves the axes unchanged.
We now turn to how the specified growth tensor field is determined by concentration fields of growth factors, and how such concentration fields can be created by defining methods of production, consumption, diffusion, and interaction.
A specified growth tensor has three principal axes at right angles to each other. When the tissue is a curved canvas of finite thickness, we assume that although the two sides of the canvas may grow at different rates, they have the same directions of principal growth axes, one of which is always perpendicular to the mid-plane of the canvas, the other two being parallel to it. These axes and the corresponding growth rates are determined by concentration fields of factors. We assume that factor concentrations do not vary through the thickness of the tissue, and therefore represent them computationally by their values on a two-dimensional mesh of triangles, being the midplanes of the mesh of pentahedra used for the elasticity calculation.
In the GPT-framework factors can be classified into two types. Identity factors do not propagate within the canvas, while signalling factors can. The specified growth tensor at each point of the canvas is parameterised as follows. The specified principal directions of growth within the plane of the canvas are determined by the gradient of a signalling factor called POL. The specified rates of growth parallel to these directions on the two surfaces of the canvas are given by factors called and . Likewise, the specified growth rates perpendicular to the polarising gradient are given by factors and . The rate of growth of thickness of the canvas is specified by a factor .
Propagation of signalling factors may occur through a variety of mechanisms, such as diffusion or active transport. Here we implement diffusion which biologically may be a proxy for a variety of underlying mechanisms. The evolution of a concentration field is modelled by the following equation:(5)The four terms on the right hand side represent respectively diffusion (with a diffusion constant ), production at a rate , decay at a uniform rate , and the diluting effect of growth. Here is the volumetric rate of resultant growth. Our implementation handles dilution as a separate step and does not include it in the equation (see later). The values of , , and may vary in space. They may also vary in time, but we assume not rapidly, so that they can be assumed constant over a single timestep.
To obtain the concentration field at time , we can make a linear approximation and write a forward Euler equation:(6)
By methods similar to those for elasticity, we can discretise this relationship, representing the concentration distribution by its values at the vertices, and obtain an equation similar in form to equation (4): . and are calculated from the geometry of the mesh, the diffusion constant, the production and decay rates, and the current distribution . This set of equations can be solved to give the new concentration distribution. As for the calculation of displacements, the equations allow boundary conditions to be added stipulating that the concentration remains fixed at some nodes.
Unlike the case of elasticity, here has full rank and the solution is uniquely determined. The sizes of and are and respectively, where is the number of vertices of the triangular mesh. This is of the value of for the elasticity computation, resulting in a much faster solution. A comparison of a computed diffusion pattern with its analytical solution is considered in Text S1. In the case of morphogens which do not diffuse, it is not necessary to solve the diffusion equation, and the effects of production and decay can be calculated directly, vertex by vertex.
We compute diffusion separately from elastic deformation. In principle, the diffusion problem could be solved for a growing and deforming canvas, but over a short time interval only second-order effects arise from the interaction between growth and diffusion, except for the dilution effect mentioned above. When a material expands, the concentration of a physical substance spread through it must decrease in proportion. We make this correction as a separate step: after the diffusion and elasticity calculations, the concentration for each factor subject to dilution by growth is reduced at each point by the proportional expansion at that point, .
It is unrealistic to assume that cells can detect the directions of arbitrarily shallow gradients; these also pose numerical problems. There are various options for dealing with very shallow gradients. (1) Generate new sources or sinks for signalling factors as space is created through growth. This would enable patterns to be continually elaborated as the shape expands. (2) Fix the pattern before it becomes too shallow. In the particular case of tissue polarity, for example, polarity may be frozen when the magnitude of the POL gradient falls below a certain threshold. This would be equivalent to a cell becoming polarised when the tissue is small enough for gradients to be measurable, and then retaining its polarity when the gradient falls below the threshold of detectability. Alternatively, (3) the polarity can disappear, resulting in isotropic growth.
Case Z: Partitioning a canvas using a diffusing signal.
Figure 13 shows a hypothetical example of how diffusion and thresholds can lead to the canvas being partitioned into regions to create a new central region from a peripheral one. An identity factor is expressed at the rim of the disc-shaped canvas (blue in Figure 13 A,C). The expression of an identity factor is represented by the value and non-expression by the value . ( could represent a transcription factor expressed only at the rim.) Initially a signalling factor (blue in Figure 13 C) is present everywhere (in this model an initial value of ) and its rate of production is promoted by . Diffusion and decay cause the level of to drop in the centre until it reaches a steady state, bowl shape, as shown in Figure 13 C. Wherever drops below a threshold the identity factor is expressed (is set to the value ) so defining a new central region. In this example, can be considered as a regional organiser as it provides a source of the signalling factor, , that enables regions to be elaborated.
This patterning process could also be used to control the timing of particular events. For example, there can be a further factor, with a high propagation rate, which is generated where falls below a second threshold . When this threshold is reached, will propagate rapidly to activate or inhibit factors bringing in a new patterning phase.
Identity, signalling, and growth parameters may interact in many different ways. Rather than assume a fixed set of possible interactions, the software allows the user to write a general function to model interactions, the “interaction function”. The function is called on every iteration of the simulation, before the calculation of diffusion and growth. To simplify the task, a few standard functions are provided to model promotion and inhibition of one factor by another. These are:(7)(8)
We use boldface for vectors of values, one value per mesh vertex, and italic for scalar values. Multiplication and division of vectors are to be understood elementwise. and both tend to vectors of 1’s as the components of tend to zero. If factor is to be assigned the value of factor promoted by factor by an amount , one writes the MATLAB equivalent of (i.e. y = z.*pro(k,x);). If is to be inhibited by , then . It is convenient to express inhibition and promotion in this way because the overall effects of different factors (say ) that may be expressed in different regions can be obtained by multiplication (e.g. ).
The iterative loop of the simulation combines the regulatory and mechanical systems as follows.
|
10.1371/journal.pbio.1001970 | Adaptive Management and the Value of Information: Learning Via Intervention in Epidemiology | Optimal intervention for disease outbreaks is often impeded by severe scientific uncertainty. Adaptive management (AM), long-used in natural resource management, is a structured decision-making approach to solving dynamic problems that accounts for the value of resolving uncertainty via real-time evaluation of alternative models. We propose an AM approach to design and evaluate intervention strategies in epidemiology, using real-time surveillance to resolve model uncertainty as management proceeds, with foot-and-mouth disease (FMD) culling and measles vaccination as case studies. We use simulations of alternative intervention strategies under competing models to quantify the effect of model uncertainty on decision making, in terms of the value of information, and quantify the benefit of adaptive versus static intervention strategies. Culling decisions during the 2001 UK FMD outbreak were contentious due to uncertainty about the spatial scale of transmission. The expected benefit of resolving this uncertainty prior to a new outbreak on a UK-like landscape would be £45–£60 million relative to the strategy that minimizes livestock losses averaged over alternate transmission models. AM during the outbreak would be expected to recover up to £20.1 million of this expected benefit. AM would also recommend a more conservative initial approach (culling of infected premises and dangerous contact farms) than would a fixed strategy (which would additionally require culling of contiguous premises). For optimal targeting of measles vaccination, based on an outbreak in Malawi in 2010, AM allows better distribution of resources across the affected region; its utility depends on uncertainty about both the at-risk population and logistical capacity. When daily vaccination rates are highly constrained, the optimal initial strategy is to conduct a small, quick campaign; a reduction in expected burden of approximately 10,000 cases could result if campaign targets can be updated on the basis of the true susceptible population. Formal incorporation of a policy to update future management actions in response to information gained in the course of an outbreak can change the optimal initial response and result in significant cost savings. AM provides a framework for using multiple models to facilitate public-health decision making and an objective basis for updating management actions in response to improved scientific understanding.
| If the response to a disease outbreak is poorly managed, lives may be lost and money wasted unnecessarily. Lack of knowledge about the disease dynamics, and about the effects of our control strategies on those dynamics, means that it is difficult to do the best job possible managing such epidemiological problems. Here, we present an adaptive management approach that allows researchers to use knowledge gained during an outbreak to update ongoing interventions, thereby translating scientific discovery into improved policy. We explore the implications of adaptive management for foot-and-mouth disease outbreaks in livestock and for measles vaccination strategies in humans. In these two particular cases, planning to update management actions leads to the recommendation of a less aggressive initial approach than if changes in management are not anticipated. We demonstrate expected savings of up to £20 million in terms of lower livestock losses to culling in a foot-and-mouth outbreak based on the dynamics observed in the UK in 2001. Similarly, up to 10,000 cases could have been averted in a measles outbreak like the one observed in Malawi in 2010. Adaptive management allows real-time improvement of our understanding, and hence of management efforts, with potentially significant positive financial and health benefits.
| Improvements in public health and disease control may arise not only from novel technologies, but also through novel strategies for optimal selection and application of existing technologies [1]–[4]. Unfortunately, optimal decision making for management of epidemiological systems is often hampered by considerable uncertainty. The sources of uncertainty are myriad, but can be broadly classified into one of two categories [5]–[7]. Epistemic uncertainties are due to a lack of system or process knowledge (biological or ecological); importantly for decision makers, such uncertainties can be reduced through improvement of the state of information. Aleatory uncertainty, which includes environmental variation and other uncontrollable stochastic events, cannot generally be reduced through learning.
The implementation of epidemiological interventions under epistemic uncertainty usually takes place via one of two distinct approaches. Under non-outbreak conditions, the focus is on reducing uncertainty through research; efficacy and risks associated with novel technologies or strategies are typically inferred from extensive clinical trials [8]. While this experimental approach potentially allows for the strongest inference, it is unlikely to be rapid enough to inform dynamic decision making during a crisis. During a novel crisis, such as a disease outbreak or the emergence of a new pathogen, decisions are usually informed through retrospective analyses of prior crises, trials, and interventions [3],[9]–[12]. However, most relevant information about the dynamics of the current crisis comes from observation of the outbreak as it progresses [13]–[15]. Epidemic management practice does not currently incorporate this real-time information into ongoing decision making in any formal, objective way.
Ideally, we would like to learn while we act, rather than only before or after. In this way, we benefit from real-time feedback from the epidemic, including the response to intervention. Adaptive management (AM) is a structured, iterative, decision-making approach for dynamic problems that acknowledges uncertainty and aims to reduce this uncertainty in order to improve outcomes. AM has a robust history in both conservation and wildlife management [16]–[26], which face an analogous challenge to manage in the face of incomplete knowledge of the underlying system and its dynamics. AM determines an optimal state-dependent policy, given a set of management options, a reward (or cost) function, and one or more state dynamics models. In the face of an epidemic, reducing epistemic uncertainty is justified only when it leads to improved management; learning is not valued for its own sake. AM accounts for the future consequences of current actions by weighing the tradeoffs between short-term learning and long-term management gains; thus evaluation of the outcomes of interventions is an essential step.
Using an AM approach has several key advantages over existing approaches. First, science and policy-making are fully integrated rather than being conducted in a sequential manner; such integration prevents loss of information and reduces the subjectivity in decision making. The formalization of the entire process allows decision makers to take full advantage of the considerable literature on decision theory, with its array of tools for rigorous decision making [27]–[31]. This process requires decision makers to explicitly specify objectives and articulate the scientific uncertainties that impede management, thereby providing important insights into the decision problem from the outset. Uncertainty is addressed explicitly, in a synthetic manner, rather than being ignored or addressed in a piecemeal fashion. Thus, instead of making decisions that are contingent on different individual model formulations and assumptions, the AM framework suggests an optimal decision, or set of decisions, that integrates across all models. Finally, the choice of management actions can be updated in response to current events, in a formal and objective way, rather than being decided a priori and then only updated on an ad hoc basis when the weight of evidence demands a shift in tactics, if at all.
Despite its potential to improve management, there has been no formal application of planned learning with an explicit strategy for updating interventions (i.e., AM) in epidemiological systems (but see [32]–[36]). We here illustrate the potential utility of AM for two epidemiological case studies: management of foot-and-mouth disease (FMD), and vaccination strategies for measles outbreaks. We further use these case studies to illustrate a range of possible applications of AM in public health settings.
AM is used to make decisions in the face of uncertainty that would otherwise impede consensus. AM involves a sequence of steps (Table 1), including the statement of an objective (usually encapsulated in a reward [or cost] function), of possible management options, and of any uncertainties that hinder effective decision making (usually formulated as alternative state dynamic models). All possible model and action combinations are then evaluated in terms of their ability to achieve the stated objective. If all models agree about the best management action, despite disagreeing about the underlying uncertainty, then no further analysis is needed, and the decision can be made. However, if there is disagreement among models about the best action to take, it is possible to quantify how much learning about the “correct” model can be expected to improve outcomes. If the value of learning is sufficiently high, then an initial action can be chosen (on the basis of the highest expected benefit [or lowest expected cost] in light of model uncertainty), but AM plans for this action to be changed should information gained during early interventions reduce our uncertainty about the best model.
The value of AM in selecting an intervention can be evaluated using the expected value of perfect information (EVPI), which estimates the value to the decision maker of resolving one or more uncertainties prior to the implementation of specific decisions. EVPI was originally developed in economics [30], and has since been applied in ecological contexts [30],[37],[38] and in the development and evaluation of clinical trials [39]–[41] to identify key sources of uncertainty that limit management success and direct the allocation of research effort to most efficiently improve management outcomes. EVPI reflects a theoretical maximum achievable benefit [42]. Though managers often passively update interventions as new information comes to light, the potential to recover the EVPI is necessarily limited by the lack of a framework for real-time learning. This explicit structured decision-making framework is integral to AM, in which learning is valued insofar as it helps to maximize the proportion of the EVPI attained through informed interventions.
The EVPI calculates the objective value gained by learning before making a decision. It involves a comparison of costs (and/or benefits) assuming perfect information with costs (and/or benefits) assuming the current level of information. Understanding the value of perfect information can meaningfully quantify the value of undertaking an AM program. Formally, EVPI is the difference between the average of optimum values conditional on each model and the optimum of an average of values, where the expectation is taken over the weights associated with the alternative models:(1)Here, Cik is the cost associated with action i under model k, pk is the weight associated with model k (subject to the constraint that ), and indicates the optimum (in our case, the minimum) over all candidate actions (also see Table 2).
We proceed through the AM process for two case studies, using each to illustrate different aspects of value in a range of circumstances. We describe in detail both the set-up (i.e., pre-outbreak) and implementation phases of an AM approach (Table 1) to FMD outbreak response, and quantify the value of a formalized strategy to update management actions as real-time surveillance improves discrimination among models. We illustrate how structural uncertainty (uncertainty about the functional form or parameterization of models) can be characterized by a set of discrete competing models; specifically, we quantify the uncertainty about the spatial scale of FMD transmission. We further quantify the value of a formally adaptive approach to management as the proportion of the EVPI that could be attained and demonstrate that a formal plan to reduce uncertainty can affect the optimal initial intervention. We also explore policy robustness of management recommendations (for example, to scenarios of greater than specified severity, or to very different objectives). We then more briefly sketch the AM approach for measles vaccination planning, using this case study to illustrate the use of the EVPI framework to structure planning when decisions are limited by logistical uncertainties and constraints. This case study allows us to explore a continuum of uncertainty about management capabilities in the field. We further use this case study to explore how the choice of initial action is affected by the time required to monitor management consequences and implement more informed actions.
The disparate predictions of competing models are a barrier to the development of policy under traditional (non-adaptive) management approaches [47],[48]. Rather than conditioning on a single “best” model, AM incorporates and systematically seeks to reduce the scientific uncertainty that impedes success, by integrating over models that encapsulate all of the articulated uncertainties to produce an inclusive decision set. Our simultaneous consideration of three alternative parameterizations of the dispersal kernel of the Keeling and colleagues [13] model and all possible interventions illustrates the expected value of resolving uncertainty about the dispersal kernels in an FMD outbreak, possibly saving millions of pounds in lost livestock. Passive learning and ad hoc adaptation did occur during the 2001 outbreak (the initial DC strategy was altered to CP within about a month); thus there would be no additional logistical burden to an AM approach. Our results show that an AM approach could be employed to realize a good portion (32.85% for case 2) of the EVPI, and provides an objective justification for an initially less-severe culling regime by minimizing expected costs over the full epidemic, given the option to change management actions in response to the observed progression of the outbreak. As seen in the UK in 2001, FMD outbreaks can potentially cause significant economic and environmental damage, and there is substantial concern about the likelihood and potential impact of future FMD disease outbreaks, both in the UK and the USA. Using an AM approach could significantly reduce the burden of such an outbreak.
The AM approach to the measles outbreak response case study illustrates how management decisions can be framed in the context of both discrete and continuous uncertainty, here with regard to the population at risk and logistical capacity. In particular, our simulations show that the cost of uncertainty about the at-risk population is critically dependent on the logistical capacity to implement the optimal vaccination target. When daily vaccination rate is highly constrained, the optimal strategy is to conduct the smallest, and thus fastest, campaign; however, it is in this regime where the value of information is greatest—potentially reducing case burden by 12% (∼10,000 additional cases averted) if campaign targets can be updated based on the true susceptible population. Further, we illustrate the inherent trade-off between the benefit of updating vaccination targets conditional on assessment of the true susceptible population and the time required to make such an assessment. If vaccination targets can be rapidly adjusted to the outbreak setting at hand, then the optimal strategy is to implement the smallest, fastest initial age target—with the potential to realize nearly 100% of EVPI (which corresponds to 40%–60% fewer cases relative to the best static age target) if updated within 30 days. However, if the initial target cannot be updated (or only updated after a very long period of surveillance), then the optimal recommendation is to choose a broader age target, which averages risk over the alternative distributions of susceptible individuals.
The goal of AM is not to replace decision makers or to automate decision making. Modeling plays an important role in developing a mechanistic understanding of the processes that give rise to observed dynamics and that mediate the costs and benefits of management actions. With an improved mechanistic understanding of a system, inherent trade-offs in decision making can be understood and management can be optimized, in the classical sense of optimal control, relative to a given model. AM plays a role in the common situation where a mechanistic understanding cannot be resolved a priori; thus managers must choose among the potentially disparate recommendations of alternative models or parameterizations. In this setting, EVPI is a measure of the degree of consistency between model predictions with respect to management actions. Interpreting EVPI in the context of the full decision-space highlights the dependence of the recommended actions on the underlying models and focuses attention on the differences among models in terms of their recommendations (the management action to best achieve the objective) rather than in terms of the projections of the system states.
AM improves management outcomes in three ways. First, the outcome of management is quantified in terms of an objective function that can be expressed in terms of both desired biological and economic outcomes. Second, the potential benefits of future improvements to management are balanced against the short-term costs of learning [37] and the capacity to enact updated interventions. Third, the expected benefit of initial interventions is calculated in light of the ability to implement future changes; thus there is no a priori presumption of a “best” intervention, and management may change through time. Managers often “adapt” their actions on an ad hoc basis, but AM formalizes this process by assessing all models and management options simultaneously.
Our case studies demonstrate that AM has the potential to improve management outcomes for a variety of epidemiological systems. The FMD case study showcases the value of AM for improving management interventions as information accrues, rather than relying only on prior knowledge, and anticipates the value of information in choosing early intervention strategies, here via an EVPI analysis. In this example, a more moderate initial culling intervention is optimal for a broader range of parameter uncertainty when the ability to change is included in the analysis. The use of AM in the event of future FMD outbreaks, in the UK, the USA, or elsewhere, would likely also realize significant socio-economic savings. In the measles example, we illustrate that while the expected cost of an adaptive strategy is always less than that of a single fixed strategy, optimal vaccine targets and the additional benefit of an adaptive approach depend both on uncertainty about the age-distribution of the at-risk population and on the logistical constraints of implementing improved interventions. These examples, taken together, illustrate that AM explicitly values enhanced scientific understanding in terms of its capacity to improve management outcomes through selection of appropriate interventions.
AM is flexible and can easily accommodate alternative objectives, additional management options, other models, and multiple sources of uncertainty. For example, other costs, such as damage to the agricultural or tourism industries, could also be included in the FMD objective cost function. Similarly, entirely different objectives, for example minimization of epidemic duration in FMD, to reduce the time taken to return to disease-free status for trade purposes [49] are straightforward to consider (Figure 4; Text S1E). If new management options or models arise, they effectively trigger a return to the set-up phase of AM. For example, vaccination (with its own inherent uncertainties about how, and how well, the vaccine performs) was not implemented during the 2001 FMD epidemic, but is now part of the UK's contingency plan in the event of future outbreaks [4]. Similarly, in future outbreaks under markedly different situations (e.g., in the event of an outbreak in the USA) transmission uncertainty would be even more extreme, and would likely require the assessment of additional kernels (e.g., farm-to-farm contact networks), models, or management strategies [49]. Many management situations also have multidimensional uncertainties. For example, in our analysis of measles we independently examined daily vaccination rate and the rate at which age targets are updated. It is straightforward to weigh the relative value of reducing uncertainty in each of these different unknowns [42],[50]. AM can frame all of these novel aspects. Relative to the analyses presented here, these additional complexities can be readily incorporated by modifying the fundamental objective, or by expanding the value of information analyses (Table 2; Text S1B, S1C, S1E) to include the additional model and intervention combinations (and associated model weights).
Applications of AM are not limited to disease outbreaks. AM also has the potential to improve other disease management outcomes, such as routine and supplemental vaccination strategies, infectious disease surveillance, and clinical trials. AM can improve management outcomes in situations where management actions are taken repeatedly in time or space, system dynamics are influenced by management actions or by changing environmental conditions, and there is uncertainty (or disagreement) about the expected impacts of management. The potential for improvement may be limited by monitoring capacity or by the logistical or political capacity to enact changes. Nevertheless, even if a static intervention is optimal or the value of information is low, the AM approach provides a framework for incorporating predictive modeling into decision making that embraces scientific uncertainty. Thus, AM may yield significant rewards in terms of money or lives saved.
|
10.1371/journal.ppat.1004279 | Virus-Specific Regulatory T Cells Ameliorate Encephalitis by Repressing Effector T Cell Functions from Priming to Effector Stages | Several studies have demonstrated the presence of pathogen-specific Foxp3+ CD4 regulatory T cells (Treg) in infected animals, but little is known about where and how these cells affect the effector T cell responses and whether they are more suppressive than bulk Treg populations. We recently showed the presence of both epitope M133-specific Tregs (M133 Treg) and conventional CD4 T cells (M133 Tconv) in the brains of mice with coronavirus-induced encephalitis. Here, we provide new insights into the interactions between pathogenic Tconv and Tregs responding to the same epitope. M133 Tregs inhibited the proliferation but not initial activation of M133 Tconv in draining lymph nodes (DLN). Further, M133 Tregs inhibited migration of M133 Tconv from the DLN. In addition, M133 Tregs diminished microglia activation and decreased the number and function of Tconv in the infected brain. Thus, virus-specific Tregs inhibited pathogenic CD4 T cell responses during priming and effector stages, particularly those recognizing cognate antigen, and decreased mortality and morbidity without affecting virus clearance. These cells are more suppressive than bulk Tregs and provide a targeted approach to ameliorating immunopathological disease in infectious settings.
| By repressing immune responses against pathogens, regulatory CD4 T cells are double edged swords. On one hand, they ameliorate immunopathological disease, diminishing morbidity but they also potentially contribute to pathogen persistence. Tregs have long been thought to be primarily directed at self-antigens, but we and others recently demonstrated the presence of pathogen-specific Tregs in infected animals. As is true for all pathogen-specific Tregs, few details are known about how these cells suppress T cell responses, especially those responding to the cognate epitope. Here, using mice with encephalitis caused by neurotropic coronavirus, we analyzed and compared the very earliest steps in the priming, proliferation and differentiation of Treg and Tconv responding to the same epitope, thereby providing new, fundamental information about these processes. Further, we identify a new role for pathogen epitope-specific Tregs in an acute infectious disease with an immunopathological component. Compared to bulk Tregs, they have the advantage of specifically diminishing numbers and function of pathogenic CD4 T cells responding to the same epitope without suppressing the anti-virus T cell response. Their use in the context of encephalitis or other infections would allow targeting of pathogenic CD4 T cell responses without generally suppressing the protective components of the immune response.
| Regulatory T cells (Tregs), characterized by Foxp3 expression, have critical roles in suppressing pro-inflammatory immune responses, with ameliorating effects in autoimmune disease and deleterious consequences in the context of tumor clearance [1]. Tregs are also critical for the resolution of immune responses against pathogens. They are required for entry of immune cells into sites of inflammation in some viral infections [2], [3]. Additionally, in chronic viral infections, such as those caused by HIV, simian immunodeficiency virus, Friend virus and hepatitis C virus (HCV), Tregs contribute to pathogen persistence [4]. On the other hand, in acute viral infections caused by pathogens that include West Nile virus (WNV), herpes simplex virus (HSV) and mouse hepatitis virus (MHV), Tregs ameliorate acute disease [5]–[7]. If Tregs are depleted from mice infected with HSV or MHV, clinical disease is more severe [8], [9].
Until recently, Tregs were considered largely to recognize self antigens, but an increasing number of studies show that pathogen-specific Tregs are detected in infectious settings [10]–[14]. Further, these Tregs originate from thymus-derived pools, and are generally not generated by peripheral conversion from pathogen-specific effector CD4 T cell populations [10], [12], [13], [15], with the exception of Tregs specific for gut pathogens [16]. Studies of autoimmune diseases, such as diabetes mellitus, showed that adoptively transferred Tregs specific for an epitope at a site of inflammation were more suppressive than bulk populations of Tregs [17], [18]. Tregs specific for a M. tuberculosis (Mtb) CD4 T cell epitope are more suppressive than those that recognize a non-Mtb CD4 T cell epitope [19]. However, whether pathogen-specific Tregs are more potent than bulk populations of Tregs obtained from wild type mice has not been addressed in any infectious setting.
Mice infected with neurotropic strains of MHV develop acute encephalitis or acute and chronic demyelinating diseases [20]. Tregs are required to diminish immune-mediated disease in these mice. Thus, Treg depletion converted a nonlethal encephalitis to one with high mortality while transfer of bulk populations of Tregs to mice infected with a virulent strain of MHV prevented a lethal outcome [8]. In addition, transfer of naïve bulk populations of Tregs along with MHV-immune effector T cells to infected RAG1−/− (Recombination Activation Gene1−/−) mice resulted in less severe clinical disease and diminished cell infiltration when compared to mice that received only effector T cells [7]. More recently, we identified Tregs that recognized the immunodominant CD4 T cell epitope (M133) in the brains of mice infected with the neuroattenuated rJ2.2 strain of MHV as well as in the T cell precursor pool of naïve mice [12]. Tregs at sites of inflammation adapt to the milieu by expressing transcription factors such as T-bet (Th1-type), STAT-3 (Th17-type) or IRF4 (Th2-type) [21] and as expected, brain-derived M133-specific Tregs in infected mice expressed T-bet. T-bet-mediated expression of CXCR3 is necessary for Treg migration to inflamed tissues [22]. These cells expressed cytokines such as IFN-γ and TNF, in addition to IL-10 when stimulated with M133 peptide directly ex vivo. Further, we showed that IFN-γ production was not an in vitro phenomenon since IFN-γ expression by Tregs was detected directly ex vivo in the absence of peptide stimulation if mice were treated with brefeldin A prior to sacrifice [12].
As is true for all pathogen-specific Tregs, few details are known about how these cells affect T cell responses, especially those responding to the cognate epitope or whether these cells are more immunosuppressive than bulk Tregs. Addressing these questions directly in wild type mice is difficult because M133-specific Tregs comprise only a small fraction of total Tregs. To circumvent this problem, we developed a mouse transgenic for the expression of an M133-specific T cell receptor [23] and used Tregs (M133 Tregs) from these mice in the current study. The results show that Tregs function both in the draining lymph nodes (deep cervical lymph nodes, DCLN, and to a lesser extent, superficial cervical lymph nodes, CLN) and in the brain to ameliorate encephalitis severity.
In the natural infection, M133-specific Tregs were detected in the rJ2.2-infected brain [12]. However, adoptively transferred bulk populations of splenic C57BL/6 (B6) Tregs (bulk Tregs) were minimally detected in the infected central nervous system (CNS) [7]. To reconcile these disparate results, we co-transferred bulk Tregs and M133 Tregs derived from uninfected Foxp3gfp and M133 Tg-Foxp3gfp mice, respectively. Approximately 0.5–1% of the CD4 T cells in the blood of M133 Tg mice expressed Foxp3; 50% of Foxp3+ and >97% of Foxp3−CD4 T cells (Tconv) bound I-Ab/M133 tetramer (Figure 1A). Equal numbers of Violet-labeled bulk and M133 Tregs were transferred to B6 mice, some of which were infected with rJ2.2 24 hours later (Figure 1B). Tregs were identified in the lymphoid tissues of infected and uninfected recipient mice seven days after infection, using the gating strategy shown in Figure 1C. Equal numbers of bulk and M133 Tregs were detected in the spleen, cervical and deep cervical lymph nodes (CLN, DCLN) of uninfected mice and neither population proliferated substantially. In contrast, after infection, M133 but not bulk Tregs proliferated extensively in all three peripheral lymphoid tissues. Further, only M133 Tregs entered the infected brain to a detectable extent (Figure 1D). By day 7 post infection (p.i.), the ratio of M133 to bulk Tregs ranged from 10∶1 in lymphoid tissue to greater than 1000∶1 in the brain (Figure 1E). In addition to exhibiting less proliferation, transferred bulk Tregs in the DCLN were less activated when assessed by measuring expression of CD25, CTLA-4, CXCR3 and ICOS at day 4 p.i. (Figure 1F).
To identify the site of initial priming and expansion of M133 Tregs compared to Tconv, we first analyzed priming of M133 Tconv (Treg-depleted CD4 T cells) by transferring CFSE-labeled cells one day prior to infection (Figure 2A). Mice were sacrificed at days 3, 4 and 5 p.i. and analyzed for CFSE dilution (indicative of cell division) and for CXCR3 expression (required for T cell migration to the inflamed brain [24]) by M133 Tconv in the spleen, CLN, DCLN, inguinal lymph nodes (ILN) and brain. CFSE dilution occurred initially in the DCLN (Figure 2, B and C). Expression of CXCR3 was upregulated within 1–2 rounds of proliferation (Figure 2B). By day 4 p.i., M133 Tconv accounted for 10% of all CD4 T cells in the DCLN, a frequency higher than in any other lymphoid tissues that were examined (Figure 2D). While the numbers of M133 Tconvs decreased dramatically between days 4 and 5 p.i. in the DCLN (Figure 2E), they continued to increase in other lymphoid tissues. These results suggested that M133 Tconvs were primed in the DCLN and migrated to other sites. By day 5, cells that had undergone varying numbers of divisions were detected in all tissues except the brain, where only highly divided cells (complete loss of CFSE labeling) were present.
To confirm the DCLN as the site of proliferation of Tconv and determine whether Tregs were also primed at this site, we transferred Violet-labeled M133-specific Tconv and Tregs in a 1∶2 ratio to mice one day prior to infection (Figure 3A). This resulted in a 1∶1 ratio because only 50% of Tregs were M133-specific (Figure 1A). Cells were analyzed at days 2–4 instead of days 3–5 to capture early cell proliferation, since we had observed substantial proliferation of Tconv in the DCLN by day 3 p.i. when only Tconv were transferred (Figure 2B). By day 2 p.i, we observed five generations (4 divisions) of Violet-labeled Tconv in the DCLN but not other tissues. In contrast, Treg proliferation was not detected until day 3 p.i. with Violet dilution occurring initially in the DCLN (Figure 3B). By day 4 p.i., M133 Tconv and Tregs that had undergone extensive division were detected in the DCLN and CLN, with Tconv proliferation occurring to a greater extent than that of Tregs. This resulted in a dramatic increase in the ratio of Tconv to Tregs from day 3 to 4 p.i. in the DCLN and CLN (Figure 3C). Of note, only very few M133 CD4 T cells entered the brain at days 3 or 4 p.i., but remarkably, the majority of these cells were Tregs (Figure 3D). Tregs were able to enter the brain after fewer cycles of proliferation when compared to Tconv (Figure 3D and 4B). Exposure to M133 antigen was required for proliferation because neither M133 Tconv nor M133 Treg proliferated in mice infected with a rJ2.2 mutant (rJ2.2.MY135Q) in which epitope M133 expression was abrogated (Figure 3E). Similar to M133 Tconv, we detected CXCR3 expression on M133 Tregs within 1–2 generations of division (Figure 3F).
While these results showed that the DCLN were the initial site of priming, M133 Tconv and Treg proliferation was also detected in the spleen and inguinal lymph nodes. To distinguish priming at these distal sites and migration after initial priming in the DCLN, we treated mice with FTY720, a drug which inhibits cell egress from lymph nodes [25]. In these experiments, equal numbers of M133 Tconv and Tregs were transferred to B6 mice and mice were treated with the drug 0.5 hr prior to infection and then on a daily basis (Figure 4A). Proliferating M133 Tconv or Tregs were barely detectable in the spleen or ILN after FTY720 treatment at day 4 p.i., but proliferation continued unabated in the DCLN and to a lesser extent, in the CLN. No transferred cells were detected in the brain after drug treatment (Figure 4, B and C). Further, only highly divided cells were detected in the DCLN in treated mice whereas in the absence of FTY720, cells with only partial dye dilution were detected, suggesting that the latter were newly recruited. Overall, these data demonstrate that all priming occurs in the cervical lymph nodes, predominantly in the DCLN.
Since M133 Tregs and Tconv expanded at the same site, we reasoned that M133 Tregs would suppress Tconv activation and proliferation if both were present in the DCLN, and that M133 Tregs would be more suppressive than bulk Tregs since the latter did not proliferate. To examine these possibilities, we transferred M133 Tconv in the presence or absence of co-transferred M133 or bulk Tregs and sacrificed them at days 3–5 p.i. (Figure 5A). Bulk Tregs had no effect on numbers of M133 Tconv in the DCLN, spleen or brain at any time (Figure 5B). In contrast, the presence of M133 Tregs resulted in decreased numbers of Tconv in the spleen and brain at all times p.i., suggesting an effect on Tconv proliferation. However, the effects of transferred M133 Tregs were more complicated in the DCLN. M133 Tconv numbers in the DCLN were increased at day 3 and 5 p.i., but decreased at day 4.
To probe this in more detail, M133 Tconv activation and proliferation, based on Violet dilution, was analyzed with or without co-transferred M133 Tregs (Figure 6). M133 Tconv were gated as shown in Figure 6A. Activation was assessed by measuring levels of CD25 and CD69 on Tconv in the DCLN at day 3 p.i. (Figure 6B). Equivalent levels of CD25 and CD69 were detected on undivided M133 Tconv. However, even though expression of CD25 and CD69 followed similar kinetics in the presence or absence of Tregs, levels of both molecules were lower in the presence of Tregs after cells began to divide. These results suggest that the initial activation of M133 Tconv was not affected by the presence of Tregs, perhaps because of the lag in Treg relative to Tconv activation, shown in Figure 3. However, as M133 Tregs became activated and proliferated, they functioned to downregulate both molecules on Tconv. Of note, CD69 levels were lower on M133 Tregs than on M133 Tconv in mice that received both types of cells, while, as expected, CD25 levels were higher on M133 Tregs (Figure 6B, blue bars).
M133 Tregs inhibited M133 Tconv proliferation in the DCLN at days 4 and 5 p.i., when cells were examined directly ex vivo. The effect of Tregs on proliferation is most likely even greater than shown in the figure because a limitation of this assay is that dye dilution cannot be detected beyond 9 divisions. Remarkably, at day 3 p.i., Tconv that had undergone more divisions were detected in the presence of Tregs, resulting in a higher expansion index (EI) (Figure 6C). Of note, only the EI in the DCLN at day 3 p.i. is shown because the EI cannot be calculated when there are too few cells in the parent generation (DCLN and spleen at days 4 and 5 p.i.) or too few proliferated cells (spleen at day 3 in the presence of M133 Tregs). Further, Tconv numbers in the DCLN decreased from day 4 to day 5 when transferred alone, indicating significant cell egress, while numbers of these cells continued to increase in the presence of M133 Tregs (Fig. 5B). These results, in conjunction with decreased Tconv numbers in the spleen and brain at all times in the presence of M133 Tregs (Figure 5B and 6C), suggest that Tregs inhibited egress from the DCLN in addition to effects on proliferation. Of note, decreased Tconv numbers could reflect increased levels of apoptosis, but the fraction of Tconv apoptosis was the same in the presence or absence of transferred M133 Tregs in the DCLN at day 4 (Figure 6D).
In addition, levels of CXCR3 on Tconv in the DCLN were much lower at day 4 when Tregs were co-transferred (Figure 6E). Since T cell entry into sites of inflammation depends in part on CXCR3 expression, this decreased expression contributed to decreased numbers in the brain and perhaps also in the spleen. CXCR3 transcription is regulated by T-bet [26]; consistent with this, T-bet expression in Tconv was less elevated in the presence of transferred Tregs (Figure 6E). Thus, in the DCLN, M133 Tconv accumulation was determined by a balance of opposing effects on migration and proliferation, while in the spleen and brain, both effects contributed to decreased numbers of Tconv. Notably, CXCR3 and T-bet levels on M133 Tregs and M133 Tconv were similar when these cells were co-transferred (Figure 6E, blue bars).
M133-specific Tregs in the infected brain express T-bet and several cytokines (IFN-γ, TNF and IL-10) [12]. We next sought to determine whether this differentiation to a Th1 phenotype occurred after Tregs had entered the brain, or earlier in the DCLN. We transferred either M133 Tconv or Treg into mice prior to rJ2.2 infection and measured T-bet and cytokine expression in the DCLN at day 3.5 p.i. (Figure 7). T-bet was expressed by both types of cells at this early time post infection, after either one (Tconv) or three (Treg) cell divisions. Further, IFN-γ and IL-10 were expressed by both M133 Tconv and Tregs in the DCLN after direct ex vivo stimulation with M133 peptide. Tregs required more cell divisions to express IFN-γ than Tconv but conversely, IL-10 expression by Tregs occurred after fewer cell divisions than Tconv. Thus, Treg ability to express cytokines did not require entry into the infected brain, but rather the milieu in the draining DCLN was sufficient to trigger differentiation. These cells then migrated to the infected brain, already competent for cytokine expression.
To examine the effects of transferred M133 Tregs on clinical disease, we transferred 105 M133 Tregs (in the absence of M133 Tconv), bulk Tregs or bulk Tconv (control group) to infected mice one day prior to rJ2.2 infection. Consistent with priming in the DCLN, numbers of M133 Tregs increased dramatically from day 0 to day 3 p.i. in the DCLN but not the spleen. M133 Tregs were first detected in the brain at day 3, peaked at day 7 in both the brain and DCLN and then gradually declined in both organs as the infection resolved (Figure 8A). Transferred M133 Tregs, but not bulk Tregs improved survival and diminished weight loss (Figure 8, B and C) without affecting virus clearance (Figure 8D). M133 Tregs decreased the frequencies of M133-specific CD4 T cells in the brain (Figure 8E). In addition, transferred M133 Tregs inhibited the effector function of these cells, manifested by reduced IFN-γ expression per cell (Figure 8F). The transferred Tregs also diminished the frequency of CD4 T cells recognizing the subdominant CD4 T cell epitope (epitope S358) but did not change the CD8 T cell response to the immunodominant S510 and subdominant S598 epitopes (Figure 8E).
As expected, most of these transferred M133 Tregs expressed IL-10 or IL-10 and IFN-γ after direct ex vivo peptide stimulation (Figure 8G). IFN-γ was expressed at lower levels per Treg than per effector CD4 T cell (compare Figure 8, F and G). Of note, IFN-γ+ M133 Tregs expressed IL-10 at higher levels than those not expressing IFN-γ (Figure 8G) perhaps reflecting the expression of IFN-γ only after cells had extensively proliferated and were presumably highly activated (Figure 7).
To provide further support for the notion that M133 Tregs suppressed immune function in the CNS in addition to the DCLN, we assessed microglia activation after infection by measuring levels of MHC class II. Transfer of M133 Tregs resulted in decreased activation of microglia, as shown by diminished expression of MHC class II (Figure 8H) at day 7 p.i. Overall, these data indicate that M133 Tregs functioned in both the draining lymph nodes and site of infection, to diminish the effects of pathogenic CD4 T cells.
To further buttress the conclusion that M133 Tregs preferentially suppressed the M133 Tconv response, we performed an in vitro suppression assay (Figure 9). Because S358 and S510 TCR Tg mice were not available, we obtained CD4 and CD8 T cells from infected mice and labeled them with Violet. Since these cells had been activated in vivo, M133 Tregs, isolated from uninfected M133 TCR Tg mice were therefore pre-activated in vitro with anti-CD3 and anti-CD28 antibodies. In all cases, M133 peptide was included in the in vitro suppression assay to activate M133 Tregs. Consequently, in order to distinguish proliferation of M133 and S358 CD4 or S510 CD8 T cells, the latter cells were obtained from mice infected with virus lacking expression of the M133 epitope (rJ2.2.MY135Q). As shown in Figure 9B and C, M133 Tregs preferentially suppressed the proliferation of M133 Tconv. In agreement with the results shown in Figure 8E and our previous results [27], M133 Tregs inhibited the proliferation of S358-specific CD4 T cells more potently than that of S510-specific CD8 T cells.
Tregs specific for epitopes present at sites of autoimmune inflammation diminish disease more effectively than bulk Treg populations [28]. Here we extend these results to mice with coronavirus-induced encephalitis and show that Tregs specific for a virus-specific CD4 T cell epitope are highly activated and protective. M133 Tregs proliferate, express cytokines, particularly the immunosuppressive cytokine IL-10, and suppress M133 Tconv proliferation in the DCLN. Further, they inhibit Tconv migration from the site of priming, the DCLN. After transfer to rJ2.2-infected mice, M133 Tregs increase survival and decrease weight loss when compared to bulk populations of Tregs and decrease numbers and effector function of virus-specific CD4 T cells in the brain. Many of these effects reflect Treg function in the DCLN, but additionally, microglia activation in the brain was diminished after M133 Treg transfer. We showed previously that large numbers of bulk Tregs could tamper the immune response in the infected brain, by causing a dimunition in total numbers of infiltrating cells, without affecting the cellular composition [7]. However, the use of virus epitope-specific Tregs allowed for a targeted immune response with effects primarily on T cells responding to the same epitope, which are pathogenic in this case. It should be noted that treatment with pathogen-specific Tregs is not always beneficial. In mice infected with Mtb, adoptive transfer of pathogen epitope-specific Tregs resulted in delayed effector CD4 and CD8 T cell responses to non-cognate Mtb epitopes in the draining lymph nodes and lungs and subsequent increased bacterial load [19].
Virus-specific Tregs were primed in the same draining lymph nodes as effector T cells, but the kinetics of proliferation and cytokine production were different. The onset of Treg proliferation lagged by approximately one day, possibly because Treg proliferation was dependent upon IL-2 production by Tconv in the DCLN. This lag resulted in a lack of suppression of initial Tconv activation. However, even by 1–2 divisions, Tconv downregulated CD69 and CD25 expression in the presence of M133 Tregs (Figure 6B), suggesting decreased activation compared to mice that received only M133 Tconv. Notably, even after Treg proliferation began, the M133 Tconv/Treg ratio continued to increase in the DCLN and CLN, the sites of priming, but not the spleen or ILN (Figure 3C), suggesting preferential retention or proliferation and perhaps, recruitment of Tconv. Preferential M133-specific effector CD4 T cell expansion also occurs in the natural infection because the ratio of M133-specific Tregs/T effector cells in the naïve precursor pool is 0.15 but this decreases to 0.02 in the infected brain [12].
It is noteworthy that M133 Tconv proliferation appeared to increase in the DCLN at day 3 p.i. in the presence of co-transferred M133 Tregs. However, rather than enhancing proliferation at this time, we postulate that Tregs actually function to inhibit Tconv egress from the DCLN, resulting in the retention of proliferated M133 T conv. Consistent with this, fewer M133 Tconv were detected in the spleen at this time point after co-transfer (Figure 5B). These results are similar to those obtained from FTY720-treated mice (Figure 4), which showed that when T cell egress was blocked, proliferating cells accumulated in the draining lymph nodes but were not detected in distal lymphoid tissues. Tconv activation in the DCLN was also diminished in the presence of transferred M133 Tregs (Figure 6B). Tregs use a multitude of mechanisms to suppress Tconv and dendritic cell function, including the production of IL-10 [1], [29]. IL-10 was expressed by M133 Tregs within 4 divisions and likely contributed to the suppression of Tconv proliferation in the DCLN (Figure 7). Collectively, these results suggest that the relative numbers of virus-specific Tconv in the DCLN, spleen and brain of mice that received M133 Tregs, compared to those did not, reflect a complicated interplay between Treg effects on Tconv activation, proliferation and migration.
In contrast to these results, Tregs were required for optimal migration of inflammatory cells including dendritic cells to the site of infection in mice infected with genital HSV, lymphocytic choriomeningitis virus or respiratory syncytial virus. In their absence, virus clearance was delayed and survival decreased [2], [3]. In these studies, Tregs were depleted prior to infection, a time when very few Tregs are virus-specific [12], [15]. Together, the results from these studies and the present one suggest that non virus-specific Tregs are important for initial egress from lymphoid tissue to sites of inflammation, but virus-specific Tregs, which would be induced later in the infection, function to inhibit migration of virus-specific T cells to the infected site.
M133 Tregs also decreased the numbers of co-transferred M133 Tconv in the infected brain (Figure 5B). M133 Tregs, if transferred alone, suppressed the endogenous M133 and, to a lesser extent, S358-specific CD4 T cell responses (Figure 8E). Similar results were obtained when the suppressive ability of M133 Tregs was examined in vitro (Figure 9). Preferential suppression of the M133 Tconv compared to the S358 Tconv response may reflect antigen competition between endogenous M133 Tconv and transferred M133 Tregs. This suppressive effect was not generalized since the number and frequency of cells responding to CD8 T cell epitopes S510 and S598 in mice were not changed by the presence of M133 Tregs (Figure 8E) and the proliferation of S510-specific CD8 T cells in vitro was less inhibited by M133 Tregs (Figure 9). This lack of effect on CD8 T cell responses in the brain may explain why virus clearance is not changed in the presence of M133 Tregs (Figure 8D); virus clearance is largely CD8 T cell dependent [20]. The diminished numbers of M133-specific Tconv reflected decreased proliferation in the DCLN, but also likely resulted from decreased expression of CXCR3 on M133 Tconv when M133 Tregs were co-transferred (Figure 6E). Treg-mediated downregulation of CXCR3 expression on Tconv in draining lymph nodes was also observed in mice with autoimmune diabetes [30]. Further, transfer of M133 Tregs resulted in decreased expression of MHC class II on microglia, which potentially diminished their ability to activate T cells (Figure 8I). In addition, M133 Tregs entered the brain at earlier times than M133 Tconv (Figure 3D). Together, these results suggest that Tregs directly suppressed microglia activation, but it is also possible that decreased numbers of Tconv in the brain contributed to lower microglial MHC class II expression.
Tregs in the infected brain expressed IFN-γ, albeit at lower levels on a per cell basis when compared to Tconv (Figure 8, F and G). We and others have shown that the expression of IFN-γ does not abrogate the immunosuppressive functions of Tregs in vitro or in mice and is even required for optimal activity in some settings [12], [27], [31], [32]. However, Treg-expressed IFN-γ may contribute to the inflammatory milieu, although whether this is physiologically significant will require further work. It is also notable that while Tregs in the brains of rJ2.2-infected mice expressed IFN-γ, IFN-γ expression by Tregs has not been detected in the lungs or draining lymph nodes of mice infected with Mtb or influenza A virus [10], [13]. Although M133 Tregs ameliorated clinical disease in rJ2.2-infected mice, their immunosuppressive efficacy in the inflamed brain may have been suboptimal. In mice with experimental autoimmune encephalomyelitis, myelin-specific Tregs were detected in the brain and spinal cord, but their function was inhibited by TNF and IL-6 [33]; these two cytokines are expressed in the rJ2.2-infected CNS [20], [34].
Transfer of M133 compared to bulk Tregs resulted in improved outcomes (Figure 8, B and C) because these cells, unlike bulk Tregs, underwent proliferation, entered the infected brain at even earlier times than M133 Tconv (Figure 3D), and decreased the number of pathogenic M133-specific CD4 T cells at this site (Figure 8E). Further, the majority of M133-specific Tregs, whether endogenous or transferred, expressed IL-10 in both the DCLN and brain [12] (Figure 7 and 8G). These results suggest a potential role for virus epitope-specific Tregs as a therapeutic option in encephalitis and perhaps in other infections. Compared to bulk Tregs, virus epitope-specific Tregs would have the advantage of specifically diminishing numbers and function of pathogenic CD4 T cells responding to the same epitope without generally suppressing the anti-virus T cell response. Thus far, Tregs have been used clinically in patients with autoimmune disease and transplantation (e.g., [35], [36], but our results suggest a possible role in viral infections.
Specific pathogen-free 6 week old C57BL/6 (B6) and Thy1.1 and CD45.1 congenic mice were purchased from the National Cancer Institute. Mice transgenic for the expression of an M133-specific public T cell receptor (M133 Tg mice) were developed as previously described [23]. Foxp3gfp (B6.Cg-FOXP3tm2tch/J) mice, in which eGFP is expressed behind an IRES element, were purchased from Jackson Laboratories. Male mice were used in all experiments. Mice were maintained in the animal care facility at the University of Iowa.
A neuroattenuated variant of the JHMV strain of MHV, rJ2.2 (a recombinant version of the J2.2-V-1 virus) [37] and a mutated rJ2.2 (rJ2.2.MY135Q) in which the M133 epitope was engineered to abolish binding to the I-Ab antigen [38], were propagated in mouse 17Cl-1 cells and titered on HeLa-MHVR cells [39]. 6–7-week-old mice were inoculated intracerebrally with 600 PFU rJ2.2 or 2,000 PFU rJ2.2.MY135Q in 30 µl DMEM. After viral inoculation, mice were examined and weighed daily.
The following antibodies and streptavidin, purchased from BD Biosciences, eBiosciences and BioLegend, were used in this study: purified functional anti-CD3 (145-2C11) and anti-CD28 (37.51), B220-APC (RA3-6B2), CD4-FITC, -PE, -PerCP-Cy5.5, -PE-CY7, or -eFluor 450 (RM 4–5), CD8-APC (53–6.7), CD11b-APC or -eFluor 450 (M1/70), CD11c-PE (HL3), CD16/CD32-biotin (93), CD25-PE (PC61), CD45- PE (30-F11), CD69-PE (H1.2F3), CTLA-4-PE (UC10-4B9), CXCR3-PE or –APC (CXCR3-173), Foxp3-FITC, -PE, or -Alexa Fluor 647 (FJK-16s), I-A/I-E-PerCP-Cy5.5 (M5/114.15.2), ICOS-PE (7E.17G9), IFN-γ-PE or -APC (XMG1.2), IL-10-APC (JES5-16E3), T-bet-PE (eBio4B10), Thy1.1-PerCP-CY5.5 (OX-7), Thy1.2-PE or –APC (30-H12), streptavidin-APC, -eFluor 450 or -BV510. A PE Annexin V Apoptosis Detection Kit I was purchased from BD Pharmingen. Cells were analyzed using a FACSVerse or LSRII (BD).
Brains harvested after PBS perfusion were dispersed and digested with 1 mg/ml collagenase D (Roche) and 0.1 mg/ml DNase I (Roche) at 37°C for 30 min. Dissociated brain was passed through a 70-µm cell strainer, followed by Percoll gradient (70/37%) centrifugation. Mononuclear cells were collected from the interphase, washed, and resuspended in culture medium for further analysis.
Peripheral blood was collected from the orbital sinus of WT-Foxp3gfp or M133 Tg-Foxp3gfp mice using heparinized Natelson blood collecting tubes (Fisher Scientific, Pittsburgh, PA). 100 µl blood were incubated with equal volumes of RP10 media containing 5U heparin and 1.6 µg PE-conjugated I-Ab/M133 tetramers (obtained from the NIH/NIAID MHC Tetramer Core Facility, Atlanta, GA) for 1 hour at 37°C. Cells were then stained with anti-CD4-PerCP-Cy5.5 at 4°C, and red blood cells were removed with ACK lysing buffer.
To detect IFN-γ and IL-10 production by antigen-specific CD4 effector T cells (Tconv) and Tregs, mononuclear cells isolated from the LN or brain were stimulated with 5 µM peptides M133 or S358 (CD4 T cells) or 1 µM S510 (CD8 T cells) (Bio-synthesis, Inc, Lewiston, TX) in complete RPMI 1640 medium for 6 h at 37°C, in the presence of 1 µl/ml Golgiplug (BD) and antigen-presenting cells (CHB3 cells, B cell line, H-2Db, I-Ab). The rJ2.2-specific epitopes encompass residues 133–147 of the transmembrane (M) protein and residues 358–372 and 510–518 of the surface (S) glycoprotein [40], [41]. Intracellular expression of IFN-γ and IL-10 was assayed. A Foxp3 Staining Buffer Set (eBioscience) was used for Foxp3 or T-bet staining or when cells were analyzed for Foxp3 and cytokine expression simultaneously; otherwise, BD Cytofix/Cytoperm and Perm/Wash buffers (BD Biosciences) were used in intracellular cytokine staining assays.
For all adoptive transfers, lymphocytes were prepared from spleens and lymph nodes of donor mice. For transfer of bulk Tconv and Tregs, cells were purified from Foxp3gfp/Thy1.1 or Foxp3gfp/CD45.1 mice. For transfer of M133 Tconv and Tregs, donor M133 TCR Tg/Foxp3gfp mice were generated by crossing M133 TCR Tg mice with Foxp3gfp or Foxp3gfp/Thy1.1 mice. CD4+ T cells were negatively selected using a CD4 T Cell Isolation Kit II (containing a biotinylated antibody cocktail) and an AutoMACS (Miltenyi Biotech). Enriched CD4 T cells were further labeled with anti-mouse CD8-APC, B220-APC and streptavidin-APC. Tconvs (APC−GFP−) or Tregs (APC−GFPhi) were then sorted twice using yield mode followed by purity mode on a FACSDiva or FACSAria (BD). Cell purities were typically 99.4%–99.7%. Cells were then labeled with 3 µM CFSE (Invitrogen) or Violet dye (CellTrace Violet Cell Proliferation Kit, Invitrogen). Because about 50% and 97% of Tregs and Tconvs in M133 TCR Tg/Foxp3gfp mice were M133-specific (Figure 1A), Tregs and Tconvs were mixed at a 2∶1 ratio (1.5×105 and 7.5×104) in most co-transfer experiments to achieve a 1∶1 ratio of M133 Tregs to Tconv. Cells were transferred to 6–7 week congenic B6 mice 24 hours prior to rJ2.2 infection. The expansion index (fold-expansion of the overall culture) was obtained using FlowJo software, proliferation platform (Tree Star). In some experiments, recipient mice were treated intraperitoneally with 80 µg FTY720 (Cayman Chemical, Ann Arbor, MI) 0.5 hour prior to infection and once a day thereafter.
B6 (Thy1.2) mice were infected with rJ2.2 or rJ2.2.MY135Q at day -8. To obtain responder cells, CD4 and CD8 T cells were sorted from the brains of rJ2.2-infected (M133-specific CD4) or of rJ2.2.MY135Q-infected (S358-specific CD4 and S510-specific CD8) mice at day 0. At day -1, M133 Tregs were sorted from naïve M133 TCR Tg/Foxp3gfp/Thy1.1 mice as described above, stimulated with coated anti-CD3 mAb+anti-CD28 mAb (5 µg/ml each) for 24 hours and washed thoroughly before use. To set up the suppression assay (Figure 9A), 5×104 Violet-labeled (2 µM) responders were co-cultured with 2.5×105 irradiated CHB3 cells (2000 rad) and the indicated number of M133 Tregs per well, in the presence of 5 µM M133 peptide (for M133-specific CD4) or 5 µM S358+5 µM M133 peptide (for S358-specific CD4) or 1 µM S510+5 µM M133 peptide (for S510-specific CD8), in a 96-well round bottom plate. After 66 hours, cells were harvested and stained with anti-Thy1.2 and anti-CD4 or anti-CD8 antibodies. CD4 or CD8 responders were analyzed for Violet dye dilution by flow cytometry. The Division Index (DI, the average number of cell divisions) was obtained using FlowJo software (Tree Star, Inc.). The percentage of suppression by Tregs was calculated as follows: % suppression = 100% x [1−(DI of responders plus Tregs/DI of responders only)].
Data are expressed as mean ± SEM. Two-tailed, unpaired Student t tests were used to analyze differences in mean values between groups in most experiments. Log-rank (Mantel-Cox) tests were used to analyze differences in survival. P values<0.05 were considered significant. *P<0.05, **P<0.01, ***P<0.001.
|
10.1371/journal.pgen.1006334 | Dnmt3a Is a Haploinsufficient Tumor Suppressor in CD8+ Peripheral T Cell Lymphoma | DNA methyltransferase 3A (DNMT3A) is an enzyme involved in DNA methylation that is frequently mutated in human hematologic malignancies. We have previously shown that inactivation of Dnmt3a in hematopoietic cells results in chronic lymphocytic leukemia in mice. Here we show that 12% of Dnmt3a-deficient mice develop CD8+ mature peripheral T cell lymphomas (PTCL) and 29% of mice are affected by both diseases. 10% of Dnmt3a+/- mice develop lymphomas, suggesting that Dnmt3a is a haploinsufficient tumor suppressor in PTCL. DNA methylation was deregulated genome-wide with 10-fold more hypo- than hypermethylated promoters and enhancers, demonstrating that hypomethylation is a major event in the development of PTCL. Hypomethylated promoters were enriched for binding sites of transcription factors AML1, NF-κB and OCT1, implying the transcription factors potential involvement in Dnmt3a-associated methylation. Whereas 71 hypomethylated genes showed an increased expression in PTCL, only 3 hypermethylated genes were silenced, suggesting that cancer-specific hypomethylation has broader effects on the transcriptome of cancer cells than hypermethylation. Interestingly, transcriptomes of Dnmt3a+/- and Dnmt3aΔ/Δ lymphomas were largely conserved and significantly overlapped with those of human tumors. Importantly, we observed downregulation of tumor suppressor p53 in Dnmt3a+/- and Dnmt3aΔ/Δ lymphomas as well as in pre-tumor thymocytes from 9 months old but not 6 weeks old Dnmt3a+/- tumor-free mice, suggesting that p53 downregulation is chronologically an intermediate event in tumorigenesis. Decrease in p53 is likely an important event in tumorigenesis because its overexpression inhibited proliferation in mouse PTCL cell lines, suggesting that low levels of p53 are important for tumor maintenance. Altogether, our data link the haploinsufficient tumor suppressor function of Dnmt3a in the prevention of mouse mature CD8+ PTCL indirectly to a bona fide tumor suppressor of T cell malignancies p53.
| Global deregulation of cytosine methylation is an epigenetic hallmark of hematologic malignancies that may promote tumorigenesis by silencing tumor suppressor genes, upregulating oncogenes, and inducing genomic instability. DNA methyltransferase 3a (DNMT3A) is one of the three catalytically active enzymes responsible for cytosine methylation and one of the most frequently mutated genes in myeloid and T cell malignancies. Its role in malignant hematopoiesis, however, remains poorly understood. Here we show that Dnmt3a is a haploinsufficient tumor suppressor in the prevention of peripheral T cell lymphomas in mice. Our molecular studies identified a large number of genes deregulated in the absence of Dnmt3a that may be putative drivers of oncogenesis. We also show that downregulation of the tumor suppressor p53 is an important event in the development of mouse T cell lymphomas. Thus, this study establishes a novel mouse model to elucidate how epigenetic deregulation of transcription contributes to the pathogenesis of T cell lymphomas.
| DNA methylation is an epigenetic modification involved in transcriptional regulation of gene expression. Three catalytically active DNA methyltransferases—Dnmt1, Dnmt3a, and Dnmt3b—are involved in the generation and maintenance of DNA methylation in mammalian cells. Dnmt3a and Dnmt3b are classified as de novo enzymes due to their methylation activity during early embryogenesis [1], whereas Dnmt1 has a high affinity for hemi-methylated sites and functions in the maintenance of methylation marks during cellular division [2,3].
Recent studies suggest that all Dnmts may play roles in generating and maintaining DNA methylation. For instance, in mouse hematopoietic stem cells, Dnmt3a is responsible for maintaining DNA methylation in lowly methylated regions known as canyons [4]. In addition, Dnmt1 was shown to have cancer-specific de novo activity in a mouse model of MYC-induced T cell lymphomas [5], whereas Dnmt3a and Dnmt3b were primarily involved in maintenance methylation in tumors [6,7]. However, a deeper understanding of individual Dnmt’s activities in normal development and in cancer is still missing.
DNA methyltransferase 3a has emerged as a central regulator of hematopoiesis over the last several years. The interest in Dnmt3a was in particular fueled by recent findings of somatic mutations in human hematologic malignancies of myeloid and T cell origin [8,9]. Given the importance of DNA methylation for differentiation of hematopoietic lineages [10] along with critical roles of Dnmt3a in differentiation and self-renewal of hematopoietic stem cells [11,12], it is not unexpected that a disruption of Dnmt3a activity affects a variety of cell types and has the potential to transform hematopoietic lineages. For example, recent studies using the Mx1-Cre transgene to conditionally delete Dnmt3a in hematopoietic stem and progenitor cells (HSPCs) followed by transplantation into lethally irradiated recipients showed that a vast majority of mice develop myeloid disorders such as myeloid dysplastic syndrome and acute myeloid leukemia (69%) with rare occurrences of CD4+CD8+ double positive T-ALL (<12%) or B-ALL (<4%) [13]. In addition, both myeloid deficiencies and neoplasms were observed in mice transplanted with Dnmt3a-null bone marrow obtained from Mx1-Cre;Dnmt3afl/fl mice, altogether highlighting the importance of Dnmt3a in prevention of myeloid transformation [14,15]. However, the role of Dnmt3a in differentiation into hematopoietic lineages and molecular functions in normal and malignant hematopoiesis in particular remain poorly understood.
To elucidate the role of Dnmt3a in normal and malignant hematopoiesis we used the EμSRα-tTA;Teto-Cre;Dnmt3afl/fl;Rosa26LOXPEGFP/EGFP (Dnmt3aΔ/Δ) mouse model to conditionally delete Dnmt3a in all cells of the hematopoietic compartment. Using this model, we previously showed that long-term Dnmt3a-defficiency resulted in the development of a chronic lymphocytic leukemia (CLL) around 1 year of age [16, 17]. In addition, we previously reported that combined inactivation of Dnmt3a and Dnmt3b results in the development of CLL and peripheral T cell lymphoma (PTCL) [16], however the molecular basis of PTCL is poorly understood. Here we expanded on our previous studies by observation of a larger cohort of Dnmt3aΔ/Δ mice. These studies revealed that while ~60% of mice succumb to CLL, ~40% of mice develop CD8+ mature peripheral T cell lymphoma either in combination with CLL or as a singular disease. Furthermore, we found that loss of one allele of Dnmt3a is sufficient to induce CD8+ PTCL in 10% of Dnmt3a heterozygous mice with tumors retaining expression of the wild-type allele. Molecular profiling of methylation and gene expression identified promoter hypomethylation as a major event in tumorigenesis of PTCL, which was frequently accompanied by upregulation of gene expression. Furthermore, we identified downregulation of tumor suppressor p53 not only in Dnmt3a+/- and Dnmt3aΔ/Δ lymphomas but also in pre-tumor thymocytes, suggesting that p53 downregulation is likely relevant in the initiation/progression of lymphomagenesis.
Altogether, our data demonstrate that Dnmt3a is a haploinsufficient tumor suppressor in the prevention of CD8+ T cell transformation and highlight the importance of understanding of the roles of Dnmt3a target genes in disease pathogenesis.
We previously utilized quadruple transgenic mice, EμSRα-tTA;Teto-Cre;Dnmt3afl/fl; Rosa26LOXPEGFP/EGFP (designated hereafter as Dnmt3aΔ/Δ), to conditionally inactivate Dnmt3a in HSPCs as well as mature cells of all hematopoietic lineages (Fig 1A). In such genetic setting, Dnmt3a is deleted in only ~40% of hematopoietic cells due to restricted patterns of EμSRα-tTA expression and cells are marked by EGFP expression driven from a reporter gene [16,17]. In the remaining 60% of hematopoietic cells the EμSRα-tTA transgene is not expressed and, as a result, the Dnmt3a conditional allele is not deleted and the EGFP reporter is not expressed. PCR based genotyping confirmed Dnmt3a deletion in EGFP–positive but not EGFP-negative stem cells as well as T-, B-, and myeloid cells (Fig 1B). Furthermore, Dnmt3a deletion was specific to cells of the hematopoietic lineages and was not observed in solid tissues (S1A Fig). Previously, we observed a small cohort of Dnmt3aΔ/Δ mice and reported the development of a CLL-like disease with a median survival of 371 days [16]. Here we utilized a larger cohort of 42 Dnmt3aΔ/Δ mice to observe the phenotypic consequences of Dnmt3a inactivation. Consistent with our previously reported data, we observed the development of a CLL-like disease in 61% of mice with a median survival of 306 days (Fig 1C and S1B Fig). This disease was characterized by expansion of B220+CD19+CD5+IgM+ EGFP+ cells in the spleen, blood, bone marrow, peritoneal cavity (IP) and occasionally in the lymph nodes (S1C and S1D Fig). Interestingly, 12% of Dnmt3aΔ/Δ mice developed a different disease (MS = 295 days) characterized not only by splenomegaly but also by a significant enlargement of lymph nodes that was not observed in the CLL cases (Fig 1C, 1D and S1B Fig). Histological analysis of spleens showed near complete effacement of the red pulp by massively expanded white pulp (Fig 1E). Small- to medium-sized cells were EGFP+, expressed markers of mature T cells–CD3, CD5, TCRβ and CD8 –and were negative for the expression of CD4, TCRγδ, NK-1.1 and CD16 (Fig 1F, 1G and S1E Fig). To determine if Dnmt3aΔ/Δ PCTLs are clonal, we analyzed TCR-Vβ rearrangement by FACS analysis. All three Dnmt3aΔ/Δ lymphomas analyzed showed only one or two TCR-Vβ rearrangements, suggesting that they were clonally derived from a single cell following somatic recombination of the TCR-β locus (S1F Fig and S1 Table). These phenotypes are most consistent with those observed in human cytotoxic peripheral T cell lymphomas not otherwise specified (PTCL-NOS).
Dnmt3a was efficiently deleted in lymph node cells of terminally ill mice as determined by immunoblot analysis using anti-Dnmt3a antibody (Fig 1H). When lymph node cells from terminally sick Dnmt3aΔ/Δ mice were injected into the IP cavity of sublethally irradiated FVB mice, recipients developed a CD3+CD8+ PTCL similar to that observed in the donor mice (Fig 1I), suggesting that Dnmt3aΔ/Δ cells have tumorigenic potential. Tumor burden (scored as average weights of spleens in terminally ill mice) was on average higher in PTCL than in CLL (S1G Fig).
In addition to the development of distinct disease types in individual mice, in 29% of mice we observed the simultaneous development of CLL and PTCL with a median survival of 293 days (Fig 1C, S1B and S1H Fig). Interestingly, all CLL and PTCL cases presented with B220+CD19+CD5+ and CD3+CD8+ immunophenotypes, respectively, suggesting that normal B-1a B cells and cytotoxic CD8+ T cells are in particularly sensitive to cellular transformation in the absence of Dnmt3a. Altogether, these data suggest that Dnmt3a is a tumor suppressor gene in prevention of CLL and PTCL in mice.
We have recently reported that mice harboring a conventional knockout allele of Dnmt3a (Dnmt3a+/- mice) develop either CLL, myeloproliferative disorder or remain healthy by 16 months of age [17]. Here we expanded these studies by observing a larger cohort of 30 Dnmt3a+/- and 20 control Dnmt3a+/+ mice. Interestingly, we found that 3 out of 30 analyzed mice developed CD8+ T cell lymphomas, which were indistinguishable from those observed in Dnmt3aΔ/Δ mice (Fig 2A–2C). None of the control mice were affected by lymphoma and remained healthy during the observational period. Serial transplantation of Dnmt3a+/- lymphoma cells induced PTCL within 2 months in secondary and tertiary transplanted mice, illustrating their selective advantage to grow and induce disease (Fig 2D). Dnmt3a+/- lymphomas retained approximately 50% expression of Dnmt3a, suggesting that the remaining allele is expressed in fully transformed cells (Fig 2E). Like Dnmt3aΔ/Δ lymphomas, Dnmt3a+/- lymphomas were also clonal (Fig 2F and S1 Table). Importantly, sequencing analysis of cDNA generated from two independent Dnmt3a+/- PTCL samples revealed no mutations in the coding sequence of Dnmt3a (S1 File), demonstrating that the expressed Dnmt3a allele is in the wild-type configuration. Similarly, we did not find any mutations in the coding sequences of two genes that are commonly mutated in human T cell malignancies, Tet2 and RhoA, and their expression was not changed in Dnmt3a-deffcient lymphomas, suggesting that changes in the activity of these genes may not be involved in the transformation of T cells in this model (S2 File and S2 Fig). Altogether, these data suggest that Dnmt3a is a haploinsufficient tumor suppressor gene in the prevention of CD8+ T cell lymphomas and CLL in mice.
To determine the nature of deregulated molecular events during PTCL development in Dnmt3aΔ/Δ mice, we performed global methylation analysis using whole genome bisulfite sequencing (WGBS) and gene expression profiling by RNA-seq on CD8+ T cells isolated from Dnmt3a+/+ spleens, as this cellular population is immunophenotypically the closest normal counterpart of CD8+CD4- PTCLs. Methylation analysis revealed that 75% of 13,859,068 CpG dinucleotides were heavily methylated (≥76%), while only 6% were methylated at low levels (≤25%) (Fig 3A). The remaining 19% of CpG were methylated at intermediate levels (25% to 75%). Likewise, 44% of core promoters (-300 to +150 bp relative to transcription start site; TSS) were heavily methylated (≥76%), while 29% were lowly methylated (≤25%) (Fig 3B). Analysis of methylation across core promoter regions revealed that over 13,000 genes had a mean methylation value greater than 50%, suggesting that the majority of promoters in CD8+ T cells are heavily methylated (Fig 3C and S2 Table). A combined gene expression and methylation analysis revealed that the majority of genes with low levels of promoter methylation were expressed, whereas genes with high levels of promoter methylation were largely repressed, suggesting that promoter methylation correlates with gene expression (Fig 3D and S3 Table).
Ingenuity pathway analysis (IPA) of highly expressed genes in CD8+ T cells revealed the top subcategories of genes significantly associated with organismal survival, hematological system, tissue morphology, hematopoiesis, lymphoid tissue structure (Fig 3E), underlining their link to hematopoietic system. Altogether, these data reveal that a significant number of promoters are hypermethylated and inactive in normal CD8+ T cells, highlighting both the importance of DNA methylation in differentiation and potential for deregulation of these genes upon inactivation of DNA methyltransferases.
To determine the effects of Dnmt3a loss on the cancer methylome, we next performed WGBS on DNA isolated from Dnmt3aΔ/Δ PTCL cells. Out of ~14 million CpG dinucleotides analyzed, we observed decreased methylation in 1,263,413 (9%) CpGs and increased methylation in 155,977 (1%) CpGs (Fig 4A and S4 Table). By analysis of differentially methylated cytosines (DMCs) we found that the majority of DMCs were present in gene bodies and intergenic regions, where hypomethylation was 8 fold higher than hypermethylation (Fig 4B and S4 Table). Although the vast majority of changes in methylcytosine levels occurred in gene bodies and intergenic regions, we detected an overall decrease in methylation in both long and short promoter regions (-1500 to +500 bp and -300 to +150 bp relative to TSS, respectively) in Dnmt3aΔ/Δ PTCL relative to CD8+ T cell controls (Fig 4C and S5 Table). Likewise, analysis of differentially methylated regions (DMRS) found significant changes in the methylation of long promoters, with 500 hypomethylated DMRS and 50 hypermethylated DMRS identified in PTCL relative to CD8+ T cell controls (Fig 4D and S6 Table). Similarly, short promoters were hypomethylated (132) more so than hypermethylated (19) in PTCL (Fig 4D and S6 Table). Like with promoters, hypomethylated DMRS were 10-and 18-fold higher than hypermethylated DMRS in gene bodies and enhancers, respectively (Fig 4D and S6 Table). Extensive hypomethylation was also observed in repeat elements, with LINE elements showing the largest degree of hypomethylation (~4 fold) compared to hypermethyation (Fig 4E).
Next, we were interested in determining if differentially methylated promoter regions shared particular transcription factor binding motifs. This analysis revealed a significant enrichment for AML1, NF-κB, and OCT1 binding motifs in hypomethylated promoters (Fig 4F). This suggests a possible involvement of these factors in maintenance methylation performed by Dnmt3a. Similarly, AP-2rep, SOX5, and myogenin binding motifs were enriched in gene promoters hypermethylated in tumors, possibly implying role of these factors in cancer-specific aberrant methylation (Fig 4G). Further functional analysis will be required to test an involvement of these proteins in deregulated methylation in mouse lymphomas.
Locus-specific analysis revealed that hypo- and hypermethylated DMRS associated with promoters and gene bodies were relatively equally distributed across the genome, with the highest number of hypomethylated promoters present on chromosomes 11 and 5, lowest numbers on chromosomes X and 12 (Fig 5 and S6 Table). Interestingly, very few differentially methylated promoters were detected on the X chromosome, suggesting that Dnmt3a is dispensable for maintenance methylation in these areas of the genome (Fig 5 and S6 Table). Altogether, our data suggest that disease development in the absence of Dnmt3a results in decreased methylation across the genome with a significant number of gene promoters affected whose untimely activation may contribute to the malignant transformation of CD8+ T cells.
To determine whether the methylation landscape generated by WGBS is specific to the PTCL sample profiled or rather represents common changes that occur in Dnmt3a-deficient lymphoma, we validated hypo- and hypermethylated promoters using reduced representation bisulfite sequencing (RRBS) on additional normal CD8+ T cells and Dnmt3aΔ/Δ T cell lymphomas. This analysis confirmed hypomethylation of 90 gene promoters identified by WGBS in Dnmt3aΔ/Δ PTCL sample (S3A Fig, S3B Fig and S7 Table). In addition, 31 out of 38 gene promoters were confirmed to be hypermethylated by RRBS (S3A Fig, S3B Fig and S7 Table). The lesser extent to which hypomethylated promoters were confirmed by RRBS is not surprising in view of the inherent bias of the RRBS method which tends to underestimate the number of hypomethylated events in promoters with low CG content [18]. In fact, analysis of CpG content across DMRS revealed that hypomethylated promoters represent regions of lower CpG content when compared to hypermethylated promoters (S3C Fig). Altogether, these data are in good agreement with results obtained from WGBS, which show large scale promoter hypomethylation in Dnmt3a-defficient PTCLs.
To further validate data obtained by global methods on Dnmt3aΔ/Δ PTCL and to assess if methylation patterns are conserved in Dnmt3a+/- PTCL, we performed locus-specific methylation analysis using Combined Bisulfite Restriction Analysis (COBRA) for 11 selected genes in multiple independent tumor samples from Dnmt3aΔ/Δ and Dnmt3a+/- mice. Consistently with results obtained by WGBS, promoters of Coro2a, Cxcr5, Ikzf3, Il2Rβ, Jdp2, Lpar5, Oas3, Ppil1, Pvt1, RacGAP1, and Wnt8a were found to be hypermethylated in normal CD8+ T cells but hypomethylated in three independent Dnmt3aΔ/Δ PTCL samples (Fig 6). Furthermore, all 11 promoters were hypomethylated in Dnmt3a+/- PTCL samples, suggesting that loss of a single Dnmt3a allele is sufficient to induce patterns of promoter hypomethylation similar to those observed in Dnmt3aΔ/Δ PTCL samples (Fig 6). To determine if promoter hypomethylation observed in Dnmt3aΔ/Δ and Dnmt3a+/- PTCL occurs as a result of Dnmt3a inactivation in normal CD8+ T cells due to the lack of Dnmt3a’s de novo or maintenance activity, we analyzed promoter methylation in CD8+ T cells isolated from 8-week old Dnmt3aΔ/Δ and Dnmt3a+/- mice. For all 11 genes tested, promoters were hypermethylated in Dnmt3aΔ/Δ and Dnmt3a+/- CD8+ T cells to a similar degree as in Dnmt3a+/+ CD8+ T cells, suggesting that partial or complete inactivation of Dnmt3a does not affect the methylation status of these genes in normal CD8+ T cells during development (Fig 6). Altogether, our data demonstrate that changes in promoter methylation identified using WGBS likely represent tumor-specific events occurring in mouse PTCL driven by mono or bi-allelic loss of Dnmt3a.
To better understand deregulated molecular events in PTCL induced by mono or bi-allelic loss of Dnmt3a we performed global gene expression profiling of Dnmt3aΔ/Δ and Dnmt3a+/- PTCLs using RNA-seq. Comparison of gene expression patterns obtained from lymphomas to patterns obtained from normal CD8+ T cells revealed that Dnmt3aΔ/Δ and Dnmt3a+/- PTCLs shared strikingly similar expression profiles. In total, 737 (69%) overexpressed and 697 (79%) underexpressed genes were conserved between Dnmt3aΔ/Δ and Dnmt3a+/- PTCLs relative to CD8+ T cell controls (Fig 7A and 7B and S8 Table). We also identified 329 upregulated and 185 downregulated genes specific to Dnmt3aΔ/Δ PTCL, as well as 650 upregulated and 549 downregulated genes specific to Dnmt3a+/- PTCL (Fig 7A and 7B and S8 Table). Altogether, these data suggest that molecular events driving T cell transformation in Dnmt3a+/- and Dnmt3aΔ/Δ mice are likely conserved.
IPA of differentially expressed genes in PTCL identified 3 Inhibited Pathways common to both Dnmt3a+/- and Dnmt3aΔ/Δ PTCL (Tec Kinase signaling, Type I Diabetes Mellitus Signaling, 4-1BB Signaling in T Lymphocytes) and 1 commonly Activated Pathways (Cyclins and Cell Cycle regulation) in Dnmt3aΔ/Δ and Dnmt3a+/- lymphomas (S4 Fig). The top 5 categories for “diseases and disorders” were identical for both Dnmt3aΔ/Δ and Dnmt3a+/- tumors (Inflammatory response, Immunological disease, Connective tissue disorder, Inflammatory disease and Skeletal and muscular disorders), further illustrating the similarities between their molecular landscapes.
Comparison of methylation and gene expression revealed that 71 genes (14%) whose promoters were hypomethylated in PTCL were associated with overexpression (Fig 7C and 7D, referred to herein as HOT genes–Hypomethylated and overexpressed in TCL). In contrast, we detected only three genes—CD226, Fhit, and Emp1—whose hypermethylation correlated with underexpression, suggesting that most of the cancer-specific hypermethylation has little effect on gene expression and tumor progression (Fig 7C). Further analysis revealed 54 genes (7%) whose predicted enhancer regions were hypomethylated in PTCL and were also overexpressed, whereas only 3 genes with hypermethylated enhancers were downregulated (Fig 7C and S5 Fig). Altogether, these data demonstrate that hypomethylation affects gene expression on a broader scale than hypermethylation in mouse Dnmt3aΔ/Δ PTCL and thus may functionally contribute to a disease development.
To determine the extent of similarity between mouse and human disease on the molecular level, we compared gene expression signatures obtained from mouse PTCL to those derived from human PTCL. We utilized microarray data obtained on a set of five normal tonsil T cells and three human PTCL samples with predicted inactivating Dnmt3a mutations [19]. When we compared expression of genes deregulated in human PTCL to those genes deregulated in either Dnmt3a+/- PTCL, we identified 316 (28%) overexpressed and 415 (36%) underexpressed genes that were shared between human and Dnmt3a+/- PTCL (Fig 8A and 8B and S9 Table). Fewer genes (252 overexpressed and 239 underexpressed) were shared between human PTCL and Dnmt3aΔ/Δ PTCL, suggesting that the transcriptome of lymphomas induced by loss of a single Dnmt3a allele more so resembles human disease than those that arise do to full inactivation of Dnmt3a (Fig 8A and 8B and S9 Table). The extent of overlap in up-and downregulated genes was significant for all comparisons (P<0.01), suggesting that similar molecular events may drive PTCL in both species.
Because promoter hypomethylation resulted in upregulated gene expression in PTCL we next asked whether any of the HOT genes may have potential oncogenic functions in the development of T cell lymphomas and whether such genes are also hypomethylated and overexpressed in human PTCLs. One such candidate gene with oncogenic function in the T cell compartment—Jun Dimerization Protein 2 protein (Jdp2)—is a component of the AP-1 transcription factor that was reported to negatively regulate Trp53 and promote the development of T cell leukemia in mice [20]. Consistently with global WGBS data, Jdp2 was hypomethylated in Dnmt3aΔ/Δ PTCLs as determined by COBRA (Fig 9A). Likewise, with RNA-seq data, analysis of Jdp2 transcript levels by qRT-PCR confirmed overexpression of Jdp2 in Dnmt3a+/- and Dnmt3aΔ/Δ PTCL samples (Fig 9B). Next, we analyzed the methylation status of the JDP2 promoter in human CD3 T cells and PTCL samples and found that like in Dnmt3a-deficient mouse PTCL, the JDP2 promoter is hypomethylated in human PTCL relative to controls (Fig 9C). To determine whether overexpression of JDP2 occurs in human PTCL, we compared transcript levels in a set of 8 human PTCLs to normal CD3+ T cells. This analysis showed ~20–1,700-fold increase in JDP2 levels in human PTCL relative to normal T cells (Fig 9D). These data demonstrate that JDP2 promoter hypomethylation correlates with its overexpression in human PTCL. To evaluate the role of Jdp2 in tumor maintenance we used an shRNA construct to knockdown the levels of Jdp2 in a Dnmt3a-deficient MYC-induced PTCL cell line. However, decrease in the levels of Jdp2 in this cell line did not affect cellular growth, suggesting Jdp2 is not required for tumor maintenance in such setting (S6A and S6B Fig). Overall, our data indicate that methylation likely plays a role in the regulation of Jdp2 in mouse and human PTCL.
Because Jdp2 was reported to negatively regulate p53 transcript levels we analyzed Trp53 expression by qRT-PCR. Despite a 10-70-fold increase in Jdp2 levels in Dnmt3a+/- and Dnmt3aΔ/Δ PTCL samples, we did not observe any effects on Trp53 transcript levels (S7 Fig), suggesting that in this setting Jdp2 overexpression has no direct effect on p53 transcription. However, analysis of p53 protein levels showed downregulation of p53 in all Dnmt3a+/- tumors and 3/4 Dnmt3aΔ/Δ tumors, suggesting that Jdp2 may regulate p53 at the protein level or p53 downregulation in tumors occurs independently of Jdp2 overexpression (Fig 10A). Gene Set Enrichment Analysis (GSEA) using RNA-seq data from normal Dnmt3a+/- and Dnmt3aΔ/Δ PTCL revealed significant downregulation of the p53 pathway genes in both settings (Fig 10B and 10C and S10 Table). To determine when p53 is downregulated during lymphomagenesis, we next measured protein levels in spleens, lymph nodes and thymi isolated from 9 months old Dnmt3a+/+ and Dnmt3a+/- mice. At this age, Dnmt3a+/- mice did not show any sign of lymphomagenesis or cellular changes in the hematopoietic compartment (S8 Fig). Interestingly, whereas p53 levels in spleens and lymph nodes were similar between Dnmt3a+/+ and Dnmt3a+/- settings, p53 was downregulated in thymocytes of Dnmt3a+/- mice (Fig 10D). To determine whether downregulation of p53 occurs as a direct response to Dnmt3a monoallelic loss we analyzed p53 protein levels in splenocytes and thymocytes of 6 weeks old Dnmt3a+/+ and Dnmt3a+/- mice. This analysis revealed no apparent differences in p53 levels, suggesting that loss of one allele of Dnmt3a is insufficient to downregulate p53 at this time point (Fig 10E). Because we did not succeed in establishing CD8+ lymphoma cell lines, we utilized previously generated MYC-induced Dnmt3a+/+ and Dnmt3a-/- T cell lymphoma cell lines [7], along with MSCV-IRES-p53-GFP overexpressing both p53 and EGFP [21], to evaluate the role of p53 in lymphomagenesis. Overexpression of p53 induced selection against EGFP-positive cells in both Dnmt3a+/+ and Dnmt3a-/- T cell lymphoma cell lines, suggesting that exogenous p53 inhibited cellular proliferation in vitro (Fig 10F). This result suggests that low p53 levels are important for tumor maintenance in vitro and therefore downregulation of p53 in vivo is likely an important event in tumorigenesis. Altogether, these data suggest that downregulation of p53 is chronologically an intermediate event in lymphomagenesis and therefore likely a relevant in initiation/progression of lymphomagenesis and may be mediated by upregulation of Jdp2.
In this study we show that loss of Dnmt3a in HSPCs in EμSRα-tTA;Teto-Cre;Dnmt3afl/fl; Rosa26LOXPEGFP/EGFP mice not only results in the development of CLL as we reported previously [16,17] but also in the development of peripheral T cell lymphomas in ~40% of Dnmt3aΔ/Δ mice either alone or in combination with CLL. We further show that not only complete inactivation but also a reduction in Dnmt3a levels results in the development of PTCL in 10% of Dnmt3a+/- mice. Lymphomas that develop in both Dnmt3aΔ/Δ and Dnmt3a+/- mice are exclusively CD8+CD4- mature T cell lymphomas. Importantly, Dnmt3a+/- PTCLs retain expression of the Dnmt3a wild-type allele. Thus, consistent with heterozygous mutations of Dnmt3a found in human T cell malignancies, Dnmt3a is a haploinsufficient tumor suppressor gene in the prevention of mouse mature CD8+CD4- T cell lymphomas.
The growing number of various malignant phenotypes observed in the hematopoietic system with Dnmt3a-deficiency in mice raises questions about the nature of deregulated events induced by Dnmt3a inactivation. Because Dnmt3a is a methyltransferase, we were interested in finding whether genes deregulated in a methylation dependent manner could provide clues towards understanding the pathobiology of Dnmt3aΔ/Δ PTCLs. A combined analysis of global methylation and gene expression identified promoter hypomethylation as a major deregulated event in PTCL development in the absence of Dnmt3a with as many as 500 genes hypomethylated in lymphomas. Of these genes, expression of as many as 71 genes (14%) was upregulated in tumors. Since Dnmt3a has now been shown to be a tumor suppressor in the prevention of a number of hematologic malignancies in a variety of biological settings [13–17], it is therefore possible that promoter hypomethylation along with gene upregulation may be either a contributing factor or even the primary event driving the initiation/progression of tumor development. In such a scenario, proto-oncogenes are silenced in normal cells but are progressively hypomethylated and overexpressed resulting in cellular transformation. Analysis of data derived from Dnmt3aΔ/Δ lymphomas identified several putative drivers of T cell transformation whose promoters were hypomethylated and overexpressed in tumors (HOT genes). One such HOT gene is the Interleukin-2 receptor Il2rb, a component of the IL-2 signaling pathway that is important for the growth of T lymphocytes. Inappropriate activation of this pathway may promote unchecked proliferation of T cells, thus contributing to tumorigenesis [22]. The HOT gene, Stat1, participates in cytokine signaling in T cells and has been reported to be significantly overexpressed in human PTCL-NOS [23]. In a mouse model of v-abl-induced leukemia, Stat1-/- mice were partially protected from the development of leukemia, demonstrating that Stat1 possesses tumor-promoting activity [24]. Another HOT gene, Trim14 was demonstrated to have oncogenic function in tongue squamous cell carcinoma cell lines by activating the NF-κB pathway [25]. Whereas HOT genes represent good candidates to explain the tumor suppressor function of Dnmt3a, demonstration of a causative oncogenic role in initiation/progression of lymphomagenesis for any of these genes is challenging as it requires long-term in vivo experiments in mice. Thus, only future functional studies can address the importance of these genes in the pathogenesis of Dnmt3aΔ/Δ PTCLs.
An additional HOT gene with predicted oncogenic activity is Jun Dimerization Protein 2 (JDP2), which we found to be hypomethylated and overexpressed not only in mouse PTCL but also in human PTCLs. Jdp2 protein is a component of the AP-1 transcription factor complex that represses transactivation mediated by the Jun family of proteins and it plays a role in AP-1-mediated responses in UV-induced apoptosis and cell differentiation [26]. Jdp2 was reported to promote liver transformation as JDP transgenic mice displayed potentiation of liver cancer, higher mortality and increased number and size of tumors [27]. Importantly, Jdp2 was identified in a screen for oncogenes able to collaborate with the loss of p27kip1 cyclin-dependent inhibitor to induce lymphomas [28]. Altogether these data along with our findings suggest that upregulation of Jdp2 induced by loss of Dnmt3a might be a contributing factor to the development of PTCL. Because Jdp2 was reported to negatively regulate Trp53 on a transcriptional level and promote the development of T cell leukemia in mice [20] we tested whether Trp53 levels are affected in Dnmt3aΔ/Δ PTCLs. Despite a 15–70 fold increase in Jdp2 levels we did not observe any changes in Trp53 transcript levels, suggesting that in this setting Jdp2 overexpression has little effect on p53 transcription. However, western blot revealed decrease in p53 protein in the majority of tumor samples, suggesting that Jdp2 may regulate p53 by other mechanisms or p53 downregulation occurs through an independent pathway not involving Jdp2. Regardless of the mechanism by which p53 is downregulated in tumors, decreased p53 protein is likely contributing to CD8+ T cell transformation due to its strong tumor suppressor function in T cell compartment. For example, it was previously reported that Trp53-/- mice are highly susceptible to spontaneous tumor development, with the majority of mice developing immature CD4+CD8+ thymic lymphomas [29]. To the best of our knowledge, there are no studies in mice demonstrating the tumor suppressor function of p53 specifically in CD8+ T cell lymphomas. However, a loss of the region containing the p53 gene on chromosome 17 was observed in human primary cutaneous CD8+ cytotoxic T cell lymphoma, suggesting that low p53 levels could be involved in the pathogenesis of human CD8+ PTCL [30]. The fact that p53 was downregulated in thymocytes isolated from 9 month old, but not 6 week old, Dnmt3a+/- tumor-free mice indicates that p53 downregulation is chronologically an intermediate event in lymphomagenesis and this strongly suggest that this event is relevant in the initiation/progression of CD8+ PTCL. Consistent with downregulation of p53 protein levels, GSEA revealed suppression of p53 pathway genes in both Dnmt3a+/- and Dnmt3aΔ/Δ PTCL tumors, such as GADD45a, ZFP36L1, and KLF4. Studies using Gadd45a-/- mice found that ablation of Gadd45a in lymphoma-prone AKR mice decreased the latency and increased the incidence of T cell lymphomas, while deletion of Gadd45a on a p53 deficient background altered the tumor spectrum to heavily favor the development of T cell lymphomas [31]. Similarly, mice deficient for ZFP36L1 and ZFP36L2 displayed altered T cell development and readily succumbed to CD8+ T cell acute lymphoblastic leukemia [32]. KLF4 was identified to be mutated in pediatric T-ALL patients [33] and was shown to induced apoptosis in primary T-ALL cells [34].
These results suggest the downregulation of p53 target genes may contribute to T cell transformation in Dnmt3a-deficient mice. Altogether, these data indicate that downregulation of p53 is an important event during lymphomagenesis in Dnmt3a+/- and Dnmt3aΔ/Δ mice.
Promoter hypomethylation and p53 downregulation may not be the only relevant events involved in the development of PTCL in Dnmt3a-defficient mice. An additional DNA methylation change that could contribute to the development of PTCL in Dnmt3aΔ/Δ mice is promoter hypermethylation, as it has been linked to the inactivation of tumor suppressor genes [35,36]. Although such changes would not be linked to Dnmt3a directly as inactivation of this enzyme is an initiating event of tumorigenesis, promoter hypermethylation mediated by other DNA methyltransferase and subsequent gene silencing could still drive tumorigenesis. In particular, in our previous studies we observed upregulation of Dnmt3b in Dnmt3a-defficient MYC-induced T cell lymphomas, suggesting that such an event may result in aberrant de novo methylation [7]. Surprisingly, despite identification of 50 genes whose promoters are hypermethylated in PTCL relative to CD8+ T cell controls, only Fhit, CD226, and Emp1 were underexpressed. This raises a possibility that silencing of these genes contributes to PTCL development. Fhit is a predicted tumor suppressor gene that is frequently deleted in B cell malignancies, including Burkitt’s lymphoma and primary effusion lymphoma [37,38]. Furthermore, in vivo studied using Fhit+/− mice found that loss of a single allele of Fhit increased susceptible to carcinogen-induced tumor development in the esophagus and forestomach, further demonstrating the role of Fhit as a tumor suppressor [39]. CD226 is expressed on different hematopoietic cells including CD8+ T cells and contributes to their activation, expansion and differentiation but its deficiency in mice did not induce lymphomas, suggesting that this gene may not be a tumor suppressor gene [40]. Similarly, Emp1 overexpression correlated with enhanced cell proliferation and poor prognosis in B cell precursor ALL leukemia, suggesting an oncogenic function of this gene at least in some hematologic malignancies [41]. However, the possible role of Fhit, CD226, and Emp1 as tumor suppressors in CD8+ Dnmt3aΔ/Δ PTCL is unclear. Thus, the role of hypermethylation and silencing in disease development and progression in mouse Dnmt3aΔ/Δ PTCL will require further investigation.
One of the interesting findings presented here is the exclusive sensitivity of CD8+ T cell to transformation in Dnmt3a+/- and Dnmt3aΔ/Δ mice. This is not a consequence of impaired T cell development as we previously reported that loss of Dnmt3a does not affect the development of hematopoietic lineages [16]. Therefore, the reason as to why CD8+ but never CD4+ or CD4+CD8+ T cells become transformed in the absence of Dnmt3a is unclear at present. We speculate that the epigenome of CD8+ T cells is more dependent on Dnmt3a than other T cell types or CD8+ T cells may acquire genetic alterations that collaborate with epimutations more readily than other T cells. Of note, a differential sensitivity of T cell subtypes to transformation has been observed in response to infection by HTLV-1, which predominantly transforms CD4+ T cells, while HTLV-2 mainly transforms CD8+ T cells [42,43]. Further studies will have to clarify whether the methylome of CD4 cells is more resistant to the lack of Dnmt3a as well as the nature of events responsible for CD8+ T cell transformation.
Another interesting finding from our study is association of transcription factor (TF) binding motifs with regions hypomethylated and hypermethylated in Dnmt3a-defficient PTCL. Analysis of TF binding sites found motifs for three TFs—AML1, NF-κB, and OCT1 –that were enriched in hypomethylated DMRS, suggesting their potential role in maintenance methylation mediated by Dnmt3a. In such a scenario, interaction of these factors with Dnmt3a may determine which specific loci Dnmt3a is targeted to. Interestingly, the p50 subunit of the NF-κB transcription factor was reported to interact with Dnmt3a in a glioblastoma cell line [44]. Similarly, we also observed association of binding sites for Ap-2rep, SOX5, and myogenin with hypermethylated DMRS. Whether any of these transcription factors play role in aberrant promoter hypo- or hypermethylation remains to be determines.
Altogether, our data identify Dnmt3a as a critical tumor suppressor gene in the prevention of B- and T cell malignancies and link decreased Dnmt3a levels to decrease in p53, which may functionally contribute to the development of CD8+ PTCL. These data along with its documented role in prevention of myeloid malignancies defines Dnmt3a as a protector of the methylome critical for safeguarding normal hematopoiesis.
EμSRα-tTA and Dnmt3a2loxP/2loxP (Dnmt3afl/fl) mice were acquired from D.W. Felsher (Stanford University) and R. Jaenisch (Whitehead University), respectively. ROSA26EGFP and Teto-Cre mice came from the Jackson Laboratories. All transgenic mice were generated using standard crosses. All mice used in these studies were of the FVB/N background and were generated using standard genetic crosses. To obtain mice with a germline transmission of the Dnmt3a- allele, we crossed EμSRα-tTA;Teto-Cre;Dnmt3afl/fl mice with FVB mice, taking advantage of our observation that the EμSRα-tTA transgene is expressed in germ cells. To generate Dnmt3a+/- we subsequently bread out transgenes by crossing obtained mice with FVB mice. PCR-based genotyping of genomic DNA isolated from the tails was used to confirm genotypes. Dnmt3a+/- mice were harvested at the experimental end point of 16 months.
Human peripheral T cell lymphoma tissue samples were acquired from Cooperative Human Tissue Network, a National Cancer Institute supported resource (www.chtn.org).
All analysis was performed at the Flow Cytometry Facility at the University of Nebraska Medical Center. Single cell suspensions were generated from mouse organs and labeled with fluorescently conjugated antibodies (eBioscience). Data was collected using the LSR II (BD Biosciences) and analyzed using BD FACSDiva software (BD Biosciences). Immunopheotypic criteria for normal and malignant cellular populations analyzed by flow cytometry are as follows: Cytotoxic T cells (EGFP-negative CD3+CD8+ without population expansion), CD8+ peripheral T cell lymphomas (EGFP+CD3+CD8+ with population expansion), B-1a cells (EGFP-negative CD19+B220+CD5+ without population expansion), chronic lymphocytic leukemia (EGFP+CD19+B220+CD5+ with population expansion). Clonality was assessed using the mouse Vβ TCR Screening Panel (BD bioscience) which uses FITC-conjugated monoclonal antibodies to recognize mouse Vβ 2, 3, 4, 5.1 and 5.2, 6, 7, 8.1 and 8.2, 8.3, 9, 10b, 11, 12, 13, 14, and 17a T cell receptors.
Western blots were performed as previously described [6] with use of the following antibodies: Dnmt3a (H-295, Santa Cruz), γ-Tubulin (H-183, Santa Cruz), p53 (SC-6243, Santa Cruz), HDAC1 (ab7028, Abcam), and HSC-70 (SC-7298, Santa Cruz).
COBRA analysis was carried out as described previously [5,6]. Mouse and human bisulfite specific primers are shown in S11 Table.
The mouse mm9 MotifMap database containing 2,237,515 transcription factor motifs [52,53] (http://motifmap.ics.uci.edu/) was used to align transcription factors motifs present within promoters (-1500 to +500 TSS) that contained a significant hypo- or hypermethylated DMR using the Bedtools intersect routine [51]. For a control comparison 12 random sets of promoters (500 promoters each were used for the hypometylated controls and 50 for the hypermethylated) were selected from the UCSC known mm9 genes database using the Excel RANDBETWEEN function and sorting from highest to lowest number. The abundance of each transcription factor within the DMR promoters and random promoters were counted using the Excel COUNTIF function. P-values were calculated used a Wilcoxon sign rank test. Only P<0.05 were considered significant.
Splenic CD3+CD8+ T cells were isolated by FACS sorting from two Dnmt3a Δ/Δ mice with PTCL. Age-matched control T cells were FACS-sorted from spleens of FVB/N mice (n = 2). Genomic DNA was isolated using standard protocols. The RRBS libraries were prepared and sequenced at the Medical Genome Facility at the Mayo Clinic and ran on an Illumina HiSeq2500 sequencer. The Streamlined Analysis and Annotation Pipeline for RRBS data (SAAP-RRBS) was specifically designed to analyze RRBS data [54]. This software was used to align and determine the methylation status of CpGs associated with this type of restriction digest high throughput method. Sequences were initially aligned with genome mm9 then converted to mm10 using the UCSC Genome Browser Batch Coordinate Conversion (liftOver) utility. The methylation heat map was generated by taking the averages for all differentially methylated CpGs for a promoter (-1500 to +500 base pairs relative to the transcription start site). Promoters were only considered to be differentially methylated if one or more CpG sites showed a 30% change in methylation. RRBS data is available for download through the NCBI Gene Expression Omnibus (GSE78146).
RNA was isolated as previously described [6] from FACS sorted CD8+ T cells obtained from spleens of control FVB/N (CD8+CD3+) and Dnmt3aΔ/Δ terminally ill PTCL mice (EGFP+CD8+CD3+). Library generation was performed using the TruSeq mRNA kit. The resulting libraries were sequenced on the Illumina HiSeq 2000 platform using paired-end 100bp runs (SeqMatic, Fremont, CA). The resulting sequencing data was first aligned using TopHat and mapped to the Mus musculus UCSC mm10 reference genome using the Bowtie2 aligner [55]. Cufflinks 2 was used to estimate FPKM of known transcripts, perform de novo assembly of novel transcripts, and calculate differential expression [56]. For differentially expressed genes, we considered those genes with a fold change ≥ 2 and a p-value < 0.05 to be significant. RNA-seq data is available for download through the NCBI Gene Expression Omnibus (GSE78146).
qRT-PCR was performed as previously described [6]. Mouse real time primer sequences used in experiments presented are shown in S11 Table.
Data was compared in Microsoft Excel using Student’s t-test (p<0.05 considered significant) or other appropriate statistical comparison listed elsewhere in materials and methods.
H&E staining was performed using standard protocols by the University of Nebraska Medical Center Tissue Science Facility.
Microarray data was downloaded from the NCBI Gene Expression Omnibus. We compared gene expression of 5 normal Tonsil T cells samples (GSE65135) to 3 PTCL samples in which DNMT3A was reported to be mutated (GSE58445) [19]. Datasets were generated with Affymetrix U133 plus 2 arrays and analyzed using Affymetrix Expression Console and Transcriptome Analysis Console (v3.0). Data was analyzed using a one-way between-subject ANOVA to generate p-values and identify differentially expressed genes (p-value < 0.05 and fold change >1.5). Genes differentially expressed in human PTCL were compared to those genes identified as being over- or underexpressed in mouse Dnmt3a+/- or Dnmt3aΔ/Δ relative to CD8+ T cell controls (RNAseq, Fold change >2, p<0.05).
All differentially expressed genes (p<0.05, fold change >2, analyzed by Cufflinks V2.0) for Dnmt3a+/- PTCL relative to wild-type CD8+ T cells and Dnmt3aΔ/Δ PTCL relative to wild-type CD8+ T cells were imported into IPA software. Core analysis were performed to identify top ranking pathways and categories for differentially expressed genes. Activated and inhibited pathways (Z-score>1.5, p<0.05) common to both Dnmt3a+/- and Dnmt3aΔ/Δ PTCL are shown in S4 Fig. In Fig 3E, IPA core analysis was performed on highly expressed genes (FPKM ≥ 10) in wild-type CD8+ T cells and the top subcategories obtained in Physiological System, Development and Functions were displayed (P<0.05, for all subcategories).
TopHat/Cufflinks/Cuffdiff RNA-seq gene-level read_group_tracking file was converted to GCT expression dataset and matching phenotype model using the Read_group_trackingToGct module (http://www.broadinstitute.org/cancer/software/genepattern/modules/docs/Read_group_trackingToGct/1).Gene Set Enrichment Analysis (GSEA, http://www.broadinstitute.org/gsea/index.jsp) was used to test the relationship between RNA-Seq mRNA expression and the Hallmark Signature gene sets (http://software.broadinstitute.org/gsea/msigdb/genesets.jsp?collection=H). From this we concentrated our effort on the Hallmark p53 pathway gene set (http://software.broadinstitute.org/gsea/msigdb/cards/HALLMARK_P53_PATHWAY.html) that consisting of 180+ genes involved in p53 pathway and network.
HuSH 29-mer shRNA scrambled and shRNA Jdp2 in the retroviral vector pRFP-C-RS were purchased from Origene. Infections were performed as previously described [6], using a Dnmt3a-deficient MYC-induced CD4+ T cell lymphoma line [7]. Doubling time was calculated from each measured time point relative to the starting concentration of cells at Day 0. Each time point calculation of doubling time was considered a replicate measure and was averaged other measurements per experimental condition.
The pMSCV-IRES-EGFP (“Vector”) and pMSCV-bla-p53(WT)-IRES-EGFP (“p53”), a kind gift from Dr. Ute Moll [21], were transfected into the PhoenixEco packaging cells and retrovirus was produced. Transductions were performed as previously described [6], using wild-type Dnmt3a and Dnmt3a-deficient MYC-induced CD4+ T cell lymphoma lines [7]. The maximum percent of EGFP expressing cells per cell population was observed at 48 hours post-transduction and all subsequent EGFP data points were normalized to this time point. EGFP was measured periodically by flow cytometry on the LSRII available at the UNMC flow cytometry core facility. Cells were cultured in RPMI-1640 media supplemented with 10% FBS, 1% pen-strep-amphotericin B, 0.5% β-mercaptoethanol and split (1:3) to (1:5) every 3 days.
This study was performed in accordance with the guidelines established by the Guide for the Care and Use of Laboratory Animals at the National Institutes of Health. All experiments involving mice were approved by the IACUC (Protocol number: 08-083-10-FC) at the University of Nebraska Medical Center.
|
10.1371/journal.pcbi.1005640 | Speciation trajectories in recombining bacterial species | It is generally agreed that bacterial diversity can be classified into genetically and ecologically cohesive units, but what produces such variation is a topic of intensive research. Recombination may maintain coherent species of frequently recombining bacteria, but the emergence of distinct clusters within a recombining species, and the impact of habitat structure in this process are not well described, limiting our understanding of how new species are created. Here we present a model of bacterial evolution in overlapping habitat space. We show that the amount of habitat overlap determines the outcome for a pair of clusters, which may range from fast clonal divergence with little interaction between the clusters to a stationary population structure, where different clusters maintain an equilibrium distance between each other for an indefinite time. We fit our model to two data sets. In Streptococcus pneumoniae, we find a genomically and ecologically distinct subset, held at a relatively constant genetic distance from the majority of the population through frequent recombination with it, while in Campylobacter jejuni, we find a minority population we predict will continue to diverge at a higher rate. This approach may predict and define speciation trajectories in multiple bacterial species.
| Species are conventionally defined as groups of individuals that breed with each other, but not with those of other species. However, this does not apply to bacteria because, even if they reproduce clonally, DNA may be donated between distinct species. Nevertheless, bacterial species do exist, and a fundamental question is how they are created. We present a mathematical model to describe bacterial speciation. The model predicts that two groups of ecologically different bacteria, assumed to live in partially overlapping habitats, may evolve into genetically distinguishable clusters, without being able to proceed to full separation. Analysis of a divergent Streprococcus pneumoniae subgroup shows that such ‘satellite species’ exist and can be distinguished from more rapidly diverging clusters, like the one we detect in Campylobacter jejuni.
| Speciation in eukaryotes is well-studied [1], but the definition of bacterial species remains controversial due to recombination, which allows transfer of DNA between distant strains [2]. While recombination may maintain the genetic coherence of a species [3–5], theory suggests selection is necessary for diversification [6]. Bacterial populations generally comprise genetically and ecologically differentiated clusters [7–9], and several explanations have been offered for this [10–12]. For example, in the Ecotype Model [10], niche -specific adaptive mutations cause genome-wide selective sweeps that remove variability between isolates in the same the niche, resulting in genetically differentiated clusters in different niches. Recently, a model of ecological differentiation among sympatric recombining bacteria has been developed [13, 14]. In this model the differentiation is triggered by an acquisition of a few habitat-specific alleles through horizontal gene transfer. If recombination between habitats is limited, the result is gradual diversification, eventually creating genomically and ecologically distinct clusters. Unlike in the Ecotype Model, which assumes genome-wide sweeps, here the sweeps occur only at the habitat-specific genes, but the overall genetic differentiation happens more slowly because recombination unlinks the habitat specific genes from the rest of the genome. The resulting pattern has a small number of short regions with strong habitat association, while the majority of the genome is relatively uncorrelated with habitat, a pattern observed between two clusters of closely related Vibrio bacteria [13].
Fig 1 shows population structures in data sets with 616 Streptococcus pneumoniae [15] and 235 Campylobacter jejuni samples [16–18] (see Materials and Methods). Both include strains divergent from the rest of the population, providing us with an opportunity to investigate the early stages of bacterial differentiation. In particular, the S. pneumoniae data consist of 16 sequence clusters (SCs) of which one, SC12, differs from the rest, and has previously been characterized as ‘atypical pneumococci’ representing a distinct species [15, 19]. All other SCs are at the same equilibrium distance from each other, maintained by recombination, corresponding to the main mode in the distance distribution [4]. Two additional modes can be discerned: one close to the origin comprising the within SC distances, which may be explained by selection of some sort [4], and the other representing the broad division of the data into SC12 vs. rest, which indicates less frequent recombination between these two clusters. Whether SC12 is a nascent cluster, which will continue to diverge, is not known. It is also possible that the distance could be an equilibrium produced by the combination of mutational divergence and occasional recombination with the parent cluster. A similar minor mode is found in C. jejuni, in this case arising from a single divergent isolate shown in red. Whether this is an isolate from a cluster in the early stages of divergence is similarly unknown.
The goal to understand the population sub-divisions observed in Fig 1 motivated us to develop a model that could reproduce similar patterns. Previously models have been used to investigate the impact of homologous recombination on population structure [3, 20], the distribution of accessory genome [21–23], parallel evolution of the core and accessory genomes [4], migration and horizontal gene transfer [24], and gene sweeps and frequency dependent selection [25]. Our model is motivated by the fact that different species carry genetic differences that lead to physiological differences, and, consequently, to niche separation. However, the niche separation between different species may be incomplete, which means partial competition of the same resources and increased opportunities for interaction, as illustrated in Fig 2A. We take the model of sympatric differentiation [13, 14] as our starting point, and extend it in two ways. First, we introduce an explicit, controllable barrier for recombination between the two populations, and second, we derive an analytical approximation for the model.
An outline of our ‘Overlapping Habitats Model’ is shown in Fig 2B. Here the habitats represent different niches, and the key characteristic is the existence of two populations of different types of strains living in partially overlapping habitats. Recombination between the populations only occurs between individuals in the shared habitat, while migration enables strains to move between different parts of the habitat space. Notably, selection is implicit in the niche structure, in that there are regions of ecological space ‘private’ to each species where the other cannot survive. This habitat-specificity is assumed non-mutable and heritable, and could in practice be caused by a small number of genes. However, unlike [14], we do not model these explicitly, but rather focus on the consequences of that adaptation for the differentiation at the rest of the genome. This formulation facilitates predictions for the evolution of the population structure, given certain amount of habitat overlap, and, on the other hand, learning parameter values that result in a given population structure as an equilibrium.
As the basis of our model, we use a Wright-Fisher forward simulation of discrete generations, where each generation is sampled with replacement from strains in the previous generation. In our model, a strain is represented by a collection of genes, similar to [2], and we assume the genes are ‘core’, i.e., present in all strains. Genes are encoded as binary sequences of fixed length (500 bp). The model has in total four free parameters: mutation rate, homologous recombination rate, the proportion of habitat overlap, and migration rate. Mutations and recombinations take place between sampling of the generations. Mutations change one base in the target sequence, while recombination results in the whole gene of the recipient to be replaced by the corresponding gene of the donor. Recombination is allowed only between strains within the same habitat, and accepted with probability that declines with respect to increasing sequence divergence [26–28]. The habitat overlap parameter specifies the size of the shared habitat, and migration determines the rate with which strains move between the shared and private habitats (see below). In contrast with [2, 4], we simulate complete binary sequences, avoiding the need for additional approximations.
In detail, we simulate a population of strains of two types, A and B, that live in habitats a and b, respectively; however, part of the habitat space, denoted by ab, is shared, and both strain types can inhabit it. For simplicity, the habitat-specificity encoding genes are assumed implicit and not simulated in the model, and we further assume that strain types can not be changed by recombination or mutation. Migration of type A strains between habitats a and ab is achieved by sampling the next generation of strains in a, for example, from all type A strains such that strains in ab are sampled with a relative weight determined by the migration parameter. This corresponds to the assumption that strains within each habitat compete against each other and those trying to enter the habitat. In detail, the sampling scheme can be described as follows. We denote by Aa and Aab type A strains that are currently in a or ab environments; Bb and Bab are defined correspondingly. We sample strains for a with replacement from Aa and Aab such that the probability of sampling a strain x is equal to
Pr ( x ) = 1 | A a | + m | A a b | , if x ∈ A a , (1)
and
Pr ( x ) = m | A a | + m | A a b | , if x ∈ A a b , (2)
where 0 ≤ m ≤ 1 is the migration parameter. Value m = 0 corresponds to no migration, in which case Eqs 1 and 2 reduce to sampling the next generation for environment a from strains already in that environment. On the other hand, m = 1 corresponds to unlimited migration, and the next generation is sampled with equal probability from all type A strains in both environments a and ab. Strains for the b environment are sampled similarly from strains in b and ab environments. Finally, strains for the ab environment are sampled according to
Pr ( x ) =1 m | A a | + m | B b | + | A a b | + | B a b | , ifx ∈ A a b or x ∈ B a b , (3)
and
Pr ( x ) = m m | A a | + m | B b | + | A a b | + | B a b | , if x ∈ A a or x ∈ B b . (4)
Thus, if m = 0, the next generation of strains for the ab environment is sampled from strains already in the environment. In the other extreme (m = 1), the strains are sampled from all strains in both populations.
R-code for running and fitting the model, both simulation and the deterministic approximation (see below), is available as S1 Code.
We also derive a deterministic approximation of the Overlapping Habitats Model, which enables rapid prediction of the evolution of the population structure without simulating the actual sequences. The model is based on average distances between and within the different sub-groups of the whole population: Aa, Aab, Bab, and Bb (see the previous sub-section). In detail, let d be a vector comprising all 4 within and 6 between distances possible for the four groups. In S1 Text, we derive a function f that expresses how the average distances in the next generation, d*, approximately depend on the distances d in the current generation:
d * = f ( d ) . (5)
One of the main interests is to identify stationary points in the distance distribution, i.e., distances d, for which
d = f ( d ) (6)
holds.
We have implemented two methods to solve Eq (6). The first consists of using the update rule Eq (5) repeatedly until d converges, in which case the stationarity condition Eq (6) is satisfied. The second way to solve Eq (6) is to use a quasi-Newton method, implemented in the optim-function of the R software, to minimize the objective function h, defined as follows:
h ( d ) = | | f ( d ) - d | | 2 (7) = [ ∑ i = 1 10 ( f i - d i ) 2 ] 1 2 , (8)
where fi is the prediction for the ith element in the distance vector of the next generation, and di the current value of the corresponding element. In practice we have found useful a strategy of first running the Newton’s method, which is fast, followed by the robust sequential update procedure to confirm convergence.
Our strategy for fitting the Overlapping Habitats Model to a particular data set can be summarized as follows: we first assume the population structure observed in the data set represents an equilibrium, and use the analytical approximation, together with estimated values from the literature when available, to learn the remaining parameters so that the result is the observed equilibrium. Hence, we assume the patterns seen in data are relatively stable, but we also compare to a model that assumes more rapid divergence, and present a way to distinguish between these two (see Results). After fitting the model using the deterministic approximation, we run the simulation, which takes the stochasticity into account, to determine how easy it is to escape the equilibrium.
As discussed above, the S. pneumoniae data can be broadly divided into two sub-populations. To estimate the habitat overlap, we assumed the population structure, i.e., the within and between sub-population distances observed, represented an equilibrium, with values within = 0.01, between = 0.017. Multiple parameter combinations produced these distances (Fig 3). Therefore, to determine the remaining parameters, we set the recombination rate, r/m to a previously reported value r/m = 11.3 [15]. The proportion of diverging strains of the whole population was set to 5%, and migration to 0.5 (results were insensitive to these choices, see Fig 3 and Results). These specifications led to an estimate of 41% habitat overlap, and a mutation rate of 2.4 mutations per generation per gene in the whole population.
The parameters for the C. jejuni were estimated similarly. In detail, we assumed that the within population distance was 0.015 (the main mode) and the between distance 0.03 (the small separate mode). We fixed the recombination rate to a plug-in estimate of r/m = 49, derived from an estimate that 98 percent of substitutions in MLST genes in the species are due to recombinations [29]. We again set the proportion of the diverging strains to be 5% of the whole population. These specifications yielded an estimate of 24% habitat overlap, and a mutation rate of 3.8 mutations per generation per gene in the whole population.
For both data sets, we set the total number of strains simulated as 10,000 and the number of genes as 30. As each gene had length 500, this corresponded to the total genome size of 15,000 bp. The probability of accepting a recombination was assumed to decline log-linearly with respect to the distance between the alleles in the donor and recipient strains, according to 10−Ax, where x is the Hamming distance between the alleles. We used A = 18 for the parameter that determines the rate of the decline, according to empirical data [2]. Before computing the ecoSNP summaries (see below) we sampled subsets of simulated strains whose sizes matched the sizes of the clusters in the data sets.
Core gene alignments and the cluster annotation of the S. pneumoniae strains were obtained from [15]. As an additional data cleaning step, we removed all genes with alignment lengths less than 265bp, which corresponded to the 0.05th quantile of the lengths of the alignments of the core genes. This step was added to increase confidence in the genes detected. This left us with 1,191 core genes in the 616 pneumococcal isolates. More specifically, the genes are here clusters of orthologous groups (COGs), and we use these terms interchangeably.
The C. jejuni data consisted of 239 previously published genomes [16–18]. From the reference-based assemblies mapped to the NCTC11168 reference genome, we extracted 423 COGs using ROARY [30] with default settings. As a data cleaning step, we removed four isolates with significantly increased levels of missing data. Additionally, we removed COGs with alignment lengths less than the 0.05th quantile (225bp) of all lengths. This left us with 401 COGs in 235 isolates. The divergent isolate in Fig 1 differs from others in terms of its sampling location (New Zealand), and by being the only isolate sampled from ‘environment’ and having ST = 2381.
To investigate the impact of habitat structure on population structure, we simulated the model for 100,000 generations with two clusters, each with 5,000 strains. We varied the habitat overlap and migration, but used realistic mutation and recombination rates corresponding to the S. pneumoniae (see Materials and Methods). Fig 4 shows the evolution of the within and between cluster distances during the simulation. With the smallest overlap (Fig 4A and 4D), the limited interaction resulted in rapid divergence of the clusters, although within cluster distances reached an equilibrium as expected [2, 4]. With the largest overlap (Fig 4C and 4F) two clusters emerged, with the between cluster distance exceeding the within distance. However the clusters did not proceed to full separation, but rather maintained an equilibrium level of separation, and, furthermore, the between distances overlapped with the within distances, making clusters difficult to distinguish (Fig 4F). With an intermediate overlap (Fig 4B and 4E) the simulation still had periods of stationary behavior; however, now the clusters slowly drifted apart as a result of genes one by one escaping the equilibrium. To understand the equilibrium, we first note that if two clusters are very close, then recombination between them does not make them any more similar. If the clusters are very distant, the ability to recombine vanishes. The equilibrium, if exists, is located at an intermediate distance where the cohesive force of recombination equals the diversifying force of mutation.
Investigation of Fig 4 reveals that the deterministic approximation predicts the simulated within cluster distances with high accuracy. Also, with the smallest overlap, the deterministic approximation does not have a solution, immediately predicting the rapid divergence. However, we also see that the approximation has a tendency to underestimate the between cluster distances. The reason for this is that the deterministic approximation is based on average distances, and therefore does not account for variation in distances between specific donor and recipient alleles, whereas in the simulation distant recombinations, which have the biggest impact, are accepted less often. Therefore the approximation slightly overestimates the impact of recombination. Also, because the approximation is non-stochastic, it can not determine how easy it is to escape the equilibrium. Therefore, in our analyses of genomic data sets (see below), we first estimated the parameters with the deterministic approximation, and then ran the simulation with the learned values to produce the final prediction. S2–S4 Figs show additional results about the impacts of migration and recombination rates, and unequal cluster sizes, with similar conclusions. One interesting finding is that as long as migration is not extremely small (<0.01), its value has a negligible impact on the population structure (S2 Fig), motivating the use of a fixed value (migration = 0.5) in analyses of genomic data sets.
We next investigated whether the population divisions in the S. pneumoniae and C. jejuni data (Fig 1) are best explained by rapid clonal divergence, a stationary equilibrium, or some intermediate of these. To fit the Overlapping Habitats Model, representing the equilibrium or slow divergence, we assumed the distances between the divergent strains and other strains to be at equilibrium, and used a plug-in recombination rate estimate from the literature to compute the approximate overlap that would produce the observed level of separation (see Materials and Methods). For both data sets, a simulation with these parameters resulted in two separate clusters that were diverging slowly, with rates of 0.32 (S. pneumoniae) and 0.45 (C. jejuni) relative to the clonal divergence rates. This indicates the separation between the clusters, especially in the C. jejuni which also has a higher clonal divergence rate (see Model fitting), has exceeded the level where recombination could prevent the divergence. However, these results alone do not yet allow us to separate the two possible explanations: first, the clusters are in the process of slow divergence, as just described, or second, the clusters are in the process of rapid clonal diversification, and the distance between them just happens momentarily to be as observed.
A detailed comparison of the models’ outputs revealed a systematic difference in the ecoSNP distributions between the scenarios of clonal divergence vs. equilibrium or slow divergence, where ecoSNPs are defined, as in [13], as variants present in all strains of one cluster and absent from all strains of the other cluster. In particular, with rapid divergence and little recombination between the clusters, the ecoSNPs started to accumulate in all genes soon after the introduction of the recombination barrier (S5 and S6 Figs). On the other hand, under the equilibrium the majority of ecoSNPs were concentrated in only a few genes that already had escaped the equilibrium, while the majority of genes had no ecoSNPs at all during the whole simulation. For both data sets, the ecoSNP distribution supports the interpretation that the observed population structure is a result of equilibrium or slow divergence, rather than rapid clonal divergence (Fig 5). In the S. pneumoniae data the observed proportion of genes with no ecoSNPs is even higher than predicted by the overlap model, suggesting that previously published recombination rates may be underestimates. We note that while quantitatively the simulation output depended on the exact parameter values, qualitatively the conclusions regarding the main patterns were robust across a wide range of parameter values.
Here we have shown that certain combinations of niche structure and recombination may result in stable but distinct clusters, creating what might be termed ‘satellite species’, as seen in S. pneumoniae, and that these may be distinguished from dynamically diverging clusters using ecoSNPs, as shown by the analysis of C. jejuni. Having shown stable clusters are possible in nature, future work will be able to focus on determining the exact dimensions of the niches and candidate loci associated with them. We should also note that ‘niche’ is here an abstraction, similar to that proposed by Hutchinson [31], as the hypervolume in resource space where a species can survive, which we consider a proxy for physical connectedness. However, we extend this to be the portion of resource space where recombination is possible. In some cases recombination might occur without direct contact between the organisms, such as if mediated by diffusing DNA, and in this case the two will not be exactly equivalent. These simplifying assumptions are intended to help make a simple model, applicable to multiple species, that can be developed further in future work.
There are several differences between our model and previous work. Notably, selection is implicit in the niche structure, in that there are regions of ecological space ‘private’ to each species where the other cannot survive. This distinguishes the niches in question from purely geographic separation. The strict fitness threshold was selected for simplicity, and could be extended to a more realistic situation where strains have some probability of surviving in different niches, at the cost of introducing additional parameters to the model. We have chosen this approach as a way of implicitly modeling selection on already ecologically differentiated clusters (or species), because our interest is in the consequences of this ecological differentiation for parts of the genome that are not directly involved in niche specificity and are able to recombine. Rather than assuming niche specifying genes themselves cannot be recombined in reality, we suspect our model approximates the case where niche specificity is due to multiple loci, such that transfer of one (or a few) is not sufficient to alter a strain’s niche or a cluster’s trajectory.
Key parameters in our model are mutation rate, recombination rate, proportion of habitat overlap, and migration rate. Within-population distances were found informative about mutation rates, and values from the literature were available for recombination rates. To understand how the habitat overlap can be learned, we first note that if recombination between clusters happens freely at the same rate as within clusters, a certain equilibrium distance between the clusters is predicted. Observed distance greater than this suggests some additional barrier for recombination, and the extent of the barrier can be learned to produce that distance. The habitat structure can be interpreted as this additional barrier. In detail, the reduction in recombination between populations equals 1 − p, where p is the proportion of pairs that can recombine of all pairs (i.e. p = (|Aab||Bab|)/(|A||B|)). Notably migration does not affect the amount of genetic exchange between the populations, but only homogenizes each internally. Consequently, any non-negligible migration rate produced similar results. This finding motivated the simplification of our model by fixing the migration parameter. Eventually, after fitting the model, the ecoSNP distribution can be used to determine whether the fitted model, representing equilibrium or slow divergence, is better suited to explain the population structure than a model of more rapid clonal divergence.
The concentration of ecoSNPs in a few genome regions has previously been taken as evidence for gene-specific sweeps of habitat-specific adaptive alleles acquired through horizontal gene transfer [13]. Our results suggest a similar pattern may emerge without explicit selection on the loci affected, as a result of reduction in recombination due to habitat structure, which may allow a region to drift sufficiently far apart to reduce the ability for genetic exchange in the locus even further. This is followed by rapid diversification within the region concerned, while the rest of the genome remains at equilibrium. This recalls the concept of ‘fragmented’ speciation in which different parts of the genome speciate at different times [32], except here this was achieved without explicit selection on the diverging region. Eventually this results in highly divergent habitat-specific loci surrounded by regions with little habitat association. In practice this process could happen together with selection at the habitat-specific loci, as both processes have the potential to increase differentiation and create ecoSNPs between the clusters.
Despite its simplicity, the model adequately captured the main sub-divisions in two data sets. Nonetheless, much structure is not captured, for example the individual sequence clusters in the S. pneumoniae data. Our model does not contradict this additional structure, but instead shows that the individual sequence clusters can indeed be ecologically different, and still maintain the equilibrium distance between them, as a mere 60% of habitat overlap is sufficient for this (Fig 4). Nevertheless, the dense clusters observed in the data likely require some additional form of selection. While some alternatives are discussed in [4], we expect that in practice the within species dynamics will be governed by far more niches, with subtle distinctions leading to far more overlap, and we are actively working to extend the present work to handle this and see if it can at least qualitatively produce substructure like that we see in the pneumococcus. To conclude, our model provides means to characterize equilibrium structures and define speciation trajectories in bacterial populations and we believe it will be helpful when interpreting similar patterns in other data sets.
|
10.1371/journal.ppat.1004534 | Amphipathic α-Helices in Apolipoproteins Are Crucial to the Formation of Infectious Hepatitis C Virus Particles | Apolipoprotein B (ApoB) and ApoE have been shown to participate in the particle formation and the tissue tropism of hepatitis C virus (HCV), but their precise roles remain uncertain. Here we show that amphipathic α-helices in the apolipoproteins participate in the HCV particle formation by using zinc finger nucleases-mediated apolipoprotein B (ApoB) and/or ApoE gene knockout Huh7 cells. Although Huh7 cells deficient in either ApoB or ApoE gene exhibited slight reduction of particles formation, knockout of both ApoB and ApoE genes in Huh7 (DKO) cells severely impaired the formation of infectious HCV particles, suggesting that ApoB and ApoE have redundant roles in the formation of infectious HCV particles. cDNA microarray analyses revealed that ApoB and ApoE are dominantly expressed in Huh7 cells, in contrast to the high level expression of all of the exchangeable apolipoproteins, including ApoA1, ApoA2, ApoC1, ApoC2 and ApoC3 in human liver tissues. The exogenous expression of not only ApoE, but also other exchangeable apolipoproteins rescued the infectious particle formation of HCV in DKO cells. In addition, expression of these apolipoproteins facilitated the formation of infectious particles of genotype 1b and 3a chimeric viruses. Furthermore, expression of amphipathic α-helices in the exchangeable apolipoproteins facilitated the particle formation in DKO cells through an interaction with viral particles. These results suggest that amphipathic α-helices in the exchangeable apolipoproteins play crucial roles in the infectious particle formation of HCV and provide clues to the understanding of life cycle of HCV and the development of novel anti-HCV therapeutics targeting for viral assembly.
| In vitro systems have been developed for the study of hepatitis C virus (HCV) infection and have revealed many details of the life cycle of HCV. Apolipoprotein B (ApoB) and ApoE have been shown to play crucial roles in the particle formation of HCV, based on data obtained by siRNA-mediated gene knockdown and overexpression of the proteins. However, precise roles of the apolipoproteins in HCV assembly have not been elucidated yet. In this study, we show that infectious particle formation of HCV in Huh7 cells was severely impaired by the knockout of both ApoB and ApoE genes by artificial nucleases, and this reduction was cancelled by the expression of not only ApoE, but also other exchangeable apolipoproteins, including ApoA1, ApoA2, ApoC1, ApoC2 and ApoC3. In addition, expression of amphipathic α-helices in the exchangeable apolipoproteins restored the infectious particle formation in the double-knockout cells through an interaction with viral particles. These results provide clues to the understanding of life cycle of HCV and the development of novel antivirals to HCV.
| More than 160 million individuals worldwide are infected with hepatitis C virus (HCV), and cirrhosis and hepatocellular carcinoma induced by HCV infection are life-threatening diseases [1]. Current standard therapy combining peg-interferon (IFN), ribavirin (RBV) and a protease inhibitor has achieved a sustained virological response (SVR) in over 80% of individuals infected with HCV genotype 1 [2]. In addition, many antiviral agents targeting non-structural proteins and host factors involved in HCV replication have been applied in clinical trials [3], [4].
In vitro systems have been developed for the study of HCV infection and have revealed many details of the life cycle of HCV. By using pseudotype particles bearing HCV envelope proteins and RNA replicon systems, many host factors required for entry and RNA replication have been identified, respectively [5], [6]. In addition, development of a robust in vitro propagation system of HCV based on the genotype 2a JFH1 strain (HCVcc) has gradually clarified the mechanism of assembly of HCV particles [7], [8]. It has been shown that the interaction of NS2 protein with structural and non-structural proteins facilitates assembly of the viral capsid and formation of infectious particles at the connection site between the ER membrane and the surface of lipid droplets (LD) [9]. On the other hand, very low density lipoprotein (VLDL) associated proteins, including apolipoprotein B (ApoB), ApoE, and microsomal triglyceride transfer protein (MTTP), have been shown to play crucial roles in the formation of infectious HCV particles [10]–[12]. Generally, ApoA, ApoB, ApoC and ApoE bind the surface of lipoprotein through the interaction between amphipathic α-helices and ER-derived membrane [13], [14]. This binding of apolipoproteins enhances the stability and hydrophilicity of lipoprotein. However, the specific roles played by the apolipoproteins in HCV particle formation are controversial. Gastaminza et al. demonstrated that ApoB and MTTP are cellular factors essential for an efficient assembly of infectious HCV particles [10]. However, studies by other groups demonstrated that ApoE is a major determinant of the infectivity and particle formation of HCV, and the ApoE fraction is highly enriched with infectious particles [11]. In addition, Mancone et al. showed that ApoA1 is required for production of infectious particles of HCV [15]. However, the evidence of the involvement of apolipoproteins in HCV particle formation is dependent on knockdown data and exogenous expression of the apolipoproteins, and thus the precise mechanisms of participation of the apolipoproteins in HCV assembly have not been elucidated [10], [11], [16].
Recently, several novel genome editing techniques have been developed, including methods using zinc finger nucleases (ZFN), transcription activator like-effector nucleases (TALEN) and CRISPR/Cas9 systems [17]–[19]. DNA double strand breaks (DSBs) induced by these artificial nucleases can be repaired by error-prone non-homologous end joining (NHEJ), resulting in mutant mice or cell lines carrying deletions, insertions, or substitutions at the cut site. To clarify the detailed function of gene family with redundant functions, the generation of animals or cell lines carrying multiple mutated genes may be essential.
In this study, Huh7 cell lines deficient in both ApoB and ApoE genes were established by using ZFNs and revealed that ApoB and ApoE redundantly participate in the formation of infectious HCV particles. Interestingly, the expression of other exchangeable apolipoproteins, i.e., ApoA1, ApoA2, ApoC1, ApoC2 and ApoC3, facilitated HCV assembly in ApoB and ApoE double-knockout cells. In addition, the expression of amphipathic α-helices in the exchangeable apolipoproteins restored the production of infectious particles in the double-knockout cells through an interaction with viral particles.
First, we compared expression levels of apolipoproteins between hepatocyte and hepatic cancer cell lines including Huh7 and HepG2 cells (Fig. 1A and B). The web-based search engine NextBio (NextBio, Santa Clara, CA) revealed that ApoB, ApoH and the exchangeable apolipoproteins ApoA1, ApoA2, ApoC1, ApoC2, ApoC3, and ApoE are highly expressed in human liver tissues (Fig. 1A). On the other hand, the expressions of ApoA1, ApoC1, ApoC2, ApoC3 and ApoH in hepatic cancer cell lines were suppressed compared to those in hepatocytes (Fig. 1B). To examine the roles of apolipoproteins in the formation of infectious HCV particles, the effects of knockdown of ApoA2, ApoB and ApoE on the infectious particle production in the supernatants were determined in Huh7 cells by focus forming assay (Fig. 1C). The transfection of siRNAs targeting to ApoA2, ApoB and ApoE significantly suppressed the production of infectious HCV particles. This inhibitory effect is well consistent with the high level of expression of these apolipoproteins in the hepatic cancer cell lines, suggesting that the apolipoproteins involved in HCV assembly are dependent on the expression pattern in hepatic cancer cell lines, including Huh7 cells [20]. Therefore, we examined the effects of exogenous expression of the apolipoproteins highly expressed in the liver tissues on the infection of HCV in the stable ApoE-knockdown Huh7 cells (Fig. 1D). In contrast to the control-knockdown cells, expression of not only ApoE but also ApoA1, ApoA2, and ApoC1 rescued the infectious particle formation in the ApoE-knockdown cells (Fig. 1E), suggesting that various exchangeable apolipoproteins participate in the efficient production of infectious HCV particles.
To obtain more convincing data on the involvement of apolipoproteins in the production of infectious HCV particles, we established knockout (KO) Huh7 cells deficient in either ApoB (B-KO1 and B-KO2) or ApoE (E-KO1 and E-KO2) by using ZFN (Figure S1). Deficiencies of ApoB or ApoE expression in these cell lines were confirmed by ELISA and immunoblotting analyses (Figure S1). First, we examined the roles of ApoB and ApoE on the entry and RNA replication of HCV by using HCV pseudotype particles (HCVpp) and subgenomic replicon (SGR) of the JFH1 strain, respectively. The B-KO and E-KO cell lines exhibited no significant effect on the infectivity of HCVpp and the colony formation of SGR (Figure S2A and Figure S2B), suggesting that ApoB and ApoE are not involved in the entry and replication processes of HCV. To examine the role of ApoB and ApoE in the propagation of HCV, HCVcc was inoculated into parental, B-KO and E-KO cell lines at an MOI of 1, and intracellular viral RNA and infectious titers in the supernatants were determined (Figure S2C and Figure S2D). Although RNA replication and infectious particle formation in B-KO cells upon infection with HCV were comparable with those in parental Huh7 cells, E-KO cells exhibited slight reduction of particle formation, and the expression of ApoE in E-KO cells rescued infectious particle formation (Figure S2C, Figure S2D, Figure S2E). Next, to examine the redundant role of ApoB, the effect of knockdown of ApoB on HCV assembly was determined in parental and E-KO Huh7 cell lines (Fig. 2A). Knockdown of ApoB in E-KO cells resulted in a more efficient reduction of infectious particle production than that in parental Huh7 cells, suggesting that ApoB and ApoE have a redundant role in the formation of infectious HCV particles.
To further confirm the redundant role of ApoB and ApoE in the HCV life cycle, especially in the particle formation, 2 clones of ApoB and ApoE double-knockout (BE-KO1 and BE-KO2) Huh7 cells were established by ZFNs (Figure S3A and Figure S3B). The lack of ApoB and ApoE expressions was confirmed by immunoblotting and ELISA analyses (Figure S3C, Figure S3D, Figure S3E). The BE-KO cell lines also exhibited no significant effect on the infectivity of HCVpp (Fig. 2B) and the colony formation of SGR (Fig. 2C). Next, we examined the redundant role of ApoB and ApoE on the propagation of HCVcc. Upon infection with HCVcc at an MOI of 1, infectious titers in the supernatants of BE-KO1 and BE-KO2 cells were 50 to 100 times lower than those of parental Huh7 cells at 72 h post-infection, while the level of intracellular RNA replication was comparable (Fig. 2D and E). In addition, exogenous expression of ApoE in BE-KO (ApoE-res) cells rescued the production of infectious viral particles to levels comparable to those in parental Huh7 cells (Fig. 2F and G), suggesting that ApoB and ApoE redundantly participate in the particle formation of HCV.
It is difficult to determine the roles of ApoB in the particle formation of HCV, because ApoB is too large (550 kDa) to obtain cDNA for expression. However, previous reports have shown that expression of MTTP facilitates the secretion of ApoB [21]. To further clarify the roles of ApoB in the life cycle of HCV, we established knockout Huh7 cell lines deficient in MTTP (M-KO1 and M-KO2) and in both ApoE and MTTP (EM-KO1 and EM-KO2) by using the ZFN and CRISPR/Cas9 system (Figure S4A and Figure S4E). The lack of MTTP, ApoB and ApoE expressions was confirmed by immunoblotting and ELISA analyses (Figure Figure S4B, Figure S4C, Figure S4D, Figure S4F, Figure S4G, Figure S4H). As previously reported, the secretion of ApoB was completely abrogated in M-KO and EM-KO cells, while the mRNA levels of ApoB were comparable among Huh7, M-KO and EM-KO cells (Figure S4I). To examine the roles of MTTP in the assembly of HCV through the secretion of ApoB, HCVcc was inoculated into the Huh7, B-KO, M-KO, E-KO, BE-KO and EM-KO cell lines at an MOI of 1, and intracellular HCV genomes and infectious titers in the supernatants were determined (Fig. 3A–C). Although intracellular RNA replication in M-KO and EM-KO cells was comparable with that in Huh7, B-KO, E-KO and BE-KO cells (Fig. 3B), infectious titers in the supernatants of EM-KO cells were severely impaired as seen in BE-KO cells, while those of M-KO cells were comparable to those of parental Huh7cells (Fig. 3C), suggesting that MTTP participates in the HCV assembly through the regulation of ApoB secretion. To further confirm the roles of MTTP in HCV assembly through ApoB secretion, the effects of exogenous expression of MTTP in EM-KO cells on the infectious particle formation of HCV were determined. Immunoblotting and ELISA analyses revealed that exogenous expression of MTTP rescued the secretion of ApoB into the supernatants of EM-KO cells (Fig. 3D and E), while expression of ApoE or MTTP in both BE-KO and EM-KO cells exhibited no effect on the intracellular RNA replication (Fig. 3F). Although exogenous expression of ApoE rescued the infectious particle formation of HCV in both BE-KO and EM-KO cells, expression of MTTP rescued the particle formation in EM-KO cells but not in BE-KO cells (Fig. 3G), supporting the notion that MTTP plays a crucial role in the HCV assembly through the maturation of ApoB.
Next, to examine the roles played in HCV particles formation by other apolipoproteins highly expressed in the liver (Fig. 1A), the expressions of ApoA1, ApoA2, ApoC1, ApoC2, ApoC3 and ApoH in BE-KO1 cells were suppressed by siRNAs (Fig. 4A and Figure S5). While knockdown of ApoA1, ApoC3 and ApoH exhibited no effect, that of ApoA2, ApoC1 and ApoC2 significantly inhibited the release of infectious particles, which was consistent with the expression pattern of endogenous apolipoproteins except for ApoH in Huh7 cells (Fig. 1B), suggesting that not only ApoB and ApoE but also other exchangeable apolipoproteins participate in HCV particle formation. To confirm the redundant role of these apolipoproteins on the infectious particle formation, the effects of exogenous expression of these apolipoproteins on the propagation of HCVcc in BE-KO1 cells were determined. ApoA1, ApoA2, ApoC1, ApoC2, ApoC3, ApoE and ApoH were expressed by lentiviral vector in BE-KO1 cells (Fig. 4B upper panel). The expressions of ApoA1, ApoA2, ApoC1, ApoC2, ApoC3 and ApoE but not of ApoH enhanced extracellular HCV RNA, while they exhibited no effect on intracellular HCV RNA (Fig. 4C). In addition, the expressions of these exchangeable apolipoproteins enhanced the infectious particle formation in the supernatants of BE-KO1 cells (Fig. 4B lower panel). On the other hand, the expression of nonhepatic apolipoproteins, including ApoD, ApoL1, and ApoO, exhibited no effect on HCV particle formation in BE-KO1 cells (Figure S6). These results suggest that exogenous expression of not only the ApoE but also the ApoA and ApoC families can compensate for the impairment of HCV particle formation in BE-KO1 cells. Interestingly, specific infectivity (infectious titers/viral RNA levels in supernatants) was also enhanced by the expression of ApoA1, ApoA2, ApoC1, ApoC2, ApoC3 and ApoE, suggesting that these apolipoproteins participate in the infectious but not non-infectious particle formation of HCV (Fig. 4D). Previous reports have suggested that the expressions of Claudin1 (CLDN1), miR-122 and ApoE facilitate the production of infectious particles in nonhepatic 293T cells [16]. Therefore, the effects of exogenous expression of exchangeable apolipoproteins on particle formation were examined in 293T cells expressing CLDN1 and miR-122 (293T-CLDN/miR-122 cells). Exogenous expression of ApoA1, ApoA2, ApoC1, ApoC2, ApoC3 and ApoE, but not of ApoH by lentiviral vector facilitated the production of infectious HCV particles in 293T-CLDN/miR-122 cells (Fig. 4E). On the other hand, the expression of ApoE exhibited no effect on the propagation of Japanese encephalitis virus (JEV) and dengue virus (DENV) (Figure S7) in BE-KO1 cells. These results suggest that the exchangeable apolipoproteins and ApoB redundantly and specifically participate in the formation of HCV particles.
To examine the role of exchangeable apolipoproteins in the formation of other genotypes of HCV, the effect of exogenous expression of these apolipoproteins on the propagation of genotype 1b and 3a chimeric HCVcc, TH/JFH1 and S310/JFH1 viruses in BE-KO1 cells was determined (Fig. 5) [22], [23]. As seen in infection with HCVcc (JFH1), expression of ApoA1, ApoA2, ApoC1, ApoC2, ApoC3 and ApoE enhanced the formation of infectious particles of TH/JFH1 and S310/JFH1 chimeric viruses. These results suggest that ApoA1, ApoA2, ApoC1, ApoC2, ApoC3 and ApoE redundantly participate in the efficient formation of infectious HCV particles of genotypes 1b, 2a and 3a.
To determine the details of the assembly of infectious HCV particles in the BE-KO1 cells, intracellular infectious titers were determined in Huh7, BE-KO1 and ApoE-res cells by using the freeze and thaw method. Not only intracellular but also extracellular infection titers were impaired in BE-KO1 cells compared with those in parental and ApoE-res cells (Fig. 6A), suggesting that intracellular particle formation is impaired by deficiencies in the expression of ApoB and ApoE. Previous reports have shown that the recruitment of viral proteins around LD and redistribution of LD are essential for HCV assembly [24]. To clarify the roles of the exchangeable apolipoproteins on HCV assembly in more detail, we examined the intracellular localization of viral proteins, LD and ER in BE-KO1 and ApoE-res cells. The localization of core proteins around LD and the membranous-web structure forming the replication complex were observed in BE-KO1 cells upon infection with HCVcc, as reported in parental Huh7 cells (Fig. 6B, 6C and Figure S8). However, greater accumulation of core proteins and LD around the perinuclear region was detected in BE-KO1 cells in comparison with ApoE-res cells (Fig. 6C and 6D), supporting the notion that apolipoproteins participate in the infectious particle formation in HCV rather than viral RNA replication. Previous studies revealed that core proteins were mainly localized on the ER membrane upon infection with the genotype 2a Jc1 strain-based HCVcc (HCVcc/Jc1), and inhibition of capsid assembly and envelopment caused accumulation of core proteins on the surface of LD [25]–[27]. In ApoE-res cells, core proteins of HCVcc/Jc1 were mainly localized on the ER membrane, in contrast to the co-localization of core proteins of HCVcc (JFH1) with LD (Fig. 6E upper). However, core proteins were accumulated around LD in BE-KO1 cells infected with HCVcc/Jc1, as seen in those infected with HCVcc (JFH1) (Fig. 6E lower). These results suggest that apolipoproteins participate in the steps of HCV particle formation occurring after HCV protein assembly on the LD.
To further examine the involvement of apolipoproteins in the infectious particle formation of HCV, culture supernatants and cell lysates of BE-KO1 and ApoE-res cells infected with HCVcc were analyzed by buoyant density ultracentrifugation (Fig. 7A–B) [28]. Secretion of viral capsids in the supernatants was severely impaired in BE-KO1 cells in comparison with that in ApoE-res cells (Fig. 7A upper), in contrast to the detection of abundant intracellular capsids in both cell lines (Fig. 7B upper). Although peak levels of the core proteins and infectious titers were detected around 1.08 g/ml in both cell lines, the infectious titers in all fractions of BE-KO1 cells were significantly lower than those in ApoE-res cells, supporting the notion that apolipoproteins participate in the post-assembly process of HCV capsids which is required to confer infectivity. Next, to examine the involvement of apolipoproteins in the envelopment of HCV particles, lysates of BE-KO1 and ApoE-res cells infected with HCVcc were treated with proteinase K in the presence or absence of Triton X [26]. Protection of HCV core proteins from the protease digestion was observed in both cell lysates (Fig. 7C), suggesting that apolipoproteins are not involved in the envelopment of HCV particles. Collectively, these results suggest that exchangeable apolipoproteins participate in the post-envelopment step of HCV particle formation.
To determine the structural relevance of apolipoproteins involved in the HCV assembly, the secondary structures of the apolipoproteins were deduced by using a CLC Genomics Workbench and previous reports (Fig. 8A) [29]–[34]. Tandem repeats of amphipathic α-helices were observed in the apolipoproteins capable of rescuing HCV assembly in BE-KO1 cells, but not in those lacking this activity, suggesting that amphipathic α-helices in the apolipoproteins participate in the assembly of HCV. To examine the involvement of the amphipathic α-helices of the exchangeable apolipoproteins in the particle formation of HCV, we constructed expression plasmids encoding deletion mutants of ApoE and ApoC1, and then these deletion mutants were exogenously expressed in BE-KO1 cells by lentiviral vectors (Fig. 8B and C upper panels). The expression of all of the deletion mutants of ApoE and ApoC1 containing either N-terminal or C-terminal amphipathic α-helices rescued the particle formation of HCV in BE-KO1 cells (Fig. 8B and C lower panels), suggesting that amphipathic α-helices in the apolipoproteins play crucial roles in the production of infectious HCV particles. In addition, more abundant full-length and truncated ApoE were detected in the precipitates of the culture supernatants of cells infected with HCVcc than those of mock-infected cells concentrated by ultracentrifugation, suggesting that the amphipathic α-helices of apolipoproteins are directly associated with HCV particles (Fig. 8D and E). Taken together, the data in this study strongly suggest that exchangeable apolipoproteins redundantly participate in the infectious particle formation of HCV through the interaction between amphipathic α-helices and viral particles.
In this study, we demonstrated the redundant roles of ApoB and the exchangeable apolipoproteins ApoA1, ApoA2, ApoC1, ApoC2, ApoC3 and ApoE in the assembly of infectious HCV particles. The deficiencies of both ApoB and ApoE inhibited the production of infectious HCV particles in Huh7 cells, and exogenous expression of exchangeable apolipoproteins rescued the particle formation. cDNA microarray revealed that the expression patterns of exchangeable apolipoproteins in hepatic cancer cell lines are widely different from those in liver tissue. In previous reports, ApoE and ApoB were identified as important host factors for the assembly of infectious HCV particles [10], [11], and knockdown of ApoE and ApoB expression also inhibited the production of infectious particles in this study. Because ApoB and ApoE are major apolipoproteins in VLDL, several reports have suggested that the VLDL production machinery participates in the production of HCV particles. Furthermore, density gradient analyses revealed co-fractionation of HCV RNA with lipoproteins, with the resulting complexes being termed lipoviroparticles (LVP) [12], [35]. However, it has been reported that there is no correlation between secretion of VLDL and production of LVP [36]. In addition, exogenous expression of ApoE facilitated the infectious particle formation of HCV in 293T cells stably expressing CLDN1 and miR-122 [16], suggesting that ApoE-mediated particle formation is independent from VLDL production. Furthermore, exogenous expression of ApoA1, a major apolipoprotein of HDL, also facilitated the production of HCV particles as shown in Fig. 4E. These data suggest that the roles of the exchangeable apolipoproteins in HCV assembly are independent from the production of VLDL. MTTP plays crucial roles in the lipoprotein formation through the incorporation of triglyceride into growing lipoprotein and secretion of ApoB [21]. Although it has been shown that treatment with an MTTP inhibitor impairs the production of HCV particles [11], in this study, we found that knockout of MTTP abrogated the secretion of ApoB but not the production of infectious HCV particles. Collectively, these data suggest that exchangeable apolipoproteins redundantly participate in the infectious particle formation of HCV independently from lipoprotein secretion machinery.
Production of HCV capsids in the culture supernatants is impaired in 293T cells expressing miR-122 due to lack of ApoE expression, but envelopment of viral capsids is observed [37], suggesting that ApoE is involved in the post-envelopment step. Coller et al. suggested that ApoE is associated with de novo formation of HCV particles during secretory pathway based on an experiment using HCV possessing a tetracysteine-tag in the core protein [38]. In this study, ApoA1, ApoA2, ApoC1, ApoC2, ApoC3 and ApoE enhanced the formation of HCV particles in the post-envelopment step. These results suggest that a direct interaction between exchangeable apolipoproteins and enveloped particles in the ER lumen facilitates an efficient secretion of infectious HCV particles. Ultrastructural analysis of HCV particles has shown that large amounts of apolipoproteins, including ApoA1, ApoB and ApoE, bind to the surface of viral particles [39]. Interestingly, ApoE-specific antibodies were more efficient in capturing viral particles than α-E1/E2 antibodies, and significantly large numbers of gold particles reacting with ApoE were observed per virion than those with E2, suggesting that viral envelope proteins are masked by a large amount of apolipoproteins. The unique characteristics of interaction between apolipoproteins and HCV particles might be applied for visualization of entry and purification of HCV particles by using GFP- or affinity-tagged amphipathic α-helices of apolipoproteins. In the previous report, virocidal amphipathic helical peptides impaired the infectivity of viral particles [40]. There is a possibility that such peptide influences on the interaction between apolipoproteins and viral particles, and might be a new therapeutic approach.
In previous reports, the importance of the interaction between lipoprotein receptors and ApoE in the entry of HCV has been well established. Lipoprotein receptors including scavenger receptor class B type 1 (SR-B1) and low-density lipoprotein receptor (LDLR) are involved in HCV entry into the target cells [41], [42]. LDLR is thought to mediate cell attachment of HCV through an interaction with virus associated ApoE [43], [44]. SR-B1 also interacts with ApoE and hypervariable region 1 (HVR1) in the envelope protein of HCV [43]. In this study we have shown that exchangeable apolipoproteins including not only ApoE but also ApoA and ApoC facilitate the production of infectious HCV particles, and that exchangeable apolipoproteins directly associate with viral particles. Meunier et al. reported that ApoC1 associates intracellularly with viral particles during particle morphogenesis and enhances the entry of HCV through an interaction of the C-terminal region of ApoC1 with heparan sulfate [45]. Another group also showed that ApoC1 enhances HCV infection through the triple interplay among HVR1, ApoC1, and SR-B1 [46]. These results suggest that the interaction of HCV particles with apolipoproteins also participates in the entry through the binding of lipoprotein receptors including SR-B1 and LDLR.
Although the gene-knockout technique is essential to obtain reproducible and reliable data, and many knockout mice have been produced in various research areas, the development of experimental tools for HCV study has also been hampered by the narrow cell tropism [47], [48]. A humanized mouse model in which human liver cells were xenotransplanted into immunodeficient mouse was developed and provided an important platform for the analysis of pathogenesis and the development of antivirals for HCV [49]. However, the exogenous expression of human receptor molecules required for HCV entry and impairment of innate immunity are required for the complete propagation of HCV in mice [50]. Gene-knockout techniques using a CRISPR/Cas9 system composed of guide RNA and Cas9 nuclease that form RNA-protein complexes to cleave the target sequences [19] have allowed quick and easy establishment of gene-knockout mice and cancer cell lines [51], [52], and indeed, such MTTP-knockout cell lines were established also in this study. Recently, the high-throughput screening of host factors involved in several conditions was reported by using a CRISPR/Cas9 system [53]. Together, these novel genome-editing techniques are expected to reveal the precise roles of host factors involved in the HCV life cycle.
In summary, we have shown that apolipoproteins, including ApoA1, ApoA2, ApoC1, ApoC2, ApoC3, ApoE and ApoB, possess redundant roles in the assembly of HCV through the interaction of the amphipathic α-helices in the apolipoproteins with viral particles in the post-envelopment step. It is hoped that these findings will provide clues to the life cycle of HCV and assist in the development of novel antivirals targeting the assembly process of HCV.
The NextBio Body Atlas application presents an aggregated analysis of gene expression across various normal tissues, normal cell types, and cancer cell lines [20]. It enables us to investigate the expression of individual genes as well as gene sets. Samples for Body Atlas data are obtained from publicly available studies that are internally curated, annotated, and processed. Body Atlas measurements are generated from all available RNA expression studies that used Affymetrix U133 Plus or U133A Genechip Arrays for human studies. The results from 128 human tissue samples were incorporated from 1,067 arrays; 157 human cell types from 1,474 arrays; and 359 human cancer cell lines from 376 arrays. Gene queries return a list of relevant tissues or cell types rank-ordered by absolute gene expression and grouped by body systems or across all body systems. In the current analysis, we determined the expression levels of the apolipoproteins ApoA1, ApoA2, ApoB, ApoC1, ApoC2, ApoC3, ApoD, ApoE, ApoH, ApoL1, ApoL2 and ApoO in liver tissue. We used an analysis protocol developed by NextBio, the details of which have been described previously [20].
Expression profiling was generated using the 4 x 44 K whole human genome oligo-microarray ver.2.0 G4845A (Agilent Technologies) as previously described [54]. Raw data were imported into Subio platform ver.1.12 (Subio) for database management and quality control. Raw intensity data were normalized against GAP-DH expression levels for further analysis. These raw data have been accepted by GEO (a public repository for microarray data, aimed at storing MIAME). Access to data concerning this study may be found under GEO experiment accession number GSE32886.
All cell lines were cultured at 37°C under the conditions of a humidified atmosphere and 5% CO2. The human hepatocellular carcinoma-derived Huh7 and human embryonic kidney-derived 293T cells were obtained from Japanese Collection of Research Bioresources (JCRB) Cell Bank (JCRB0403 and JCRB9068), and maintained in DMEM (Sigma) supplemented with 100 U/ml penicillin, 100 µg/ml streptomycin, and 10% fetal calf serum (FCS). The Huh7-derived cell line Huh7.5.1 was kindly provided by F. Chisari. Huh7 cells harboring JFH1-based HCV-SGR were prepared according to the method of a previous report [54] and maintained in DMEM containing 10% FCS and 1 mg/ml G418 (Nakalai Tesque).
The cDNA clones of pri-miR-122, ApoA1, ApoA2, ApoC1, ApoC2, ApoC3, ApoE, ApoH, and AcGFP were inserted between the XhoI and XbaI sites of lentiviral vector pCSII-EF-RfA, which was kindly provided by M. Hijikata, and the resulting plasmids were designated pCSII-EF-miR-122, pCSII-EF-MT-apolipoproteins, and pCSII-EF-AcGFP, respectively. The deletion mutants of ApoC1 and ApoE were amplified by PCR and introduced into pCSII-EF. pHH-JFH1-E2p7NS2mt contains three adaptive mutations in pHH-JFH1 [55]. The pFL-J6/JFH1 plasmid that encodes the entire viral genome of the chimeric strain of HCV-2a, J6/JFH1, was kindly provided by Charles M. Rice [8]. pTH/JFH1 (genotype 1b) and pS310/JFH1 (genotype 3a) were used for the production of chimeric viruses [22], [23]. The plasmid pX330, which encodes hCas9 and sgRNA, was obtained from Addgene (Addgene plasmid 42230). The fragments of guided RNA targeting the MTTP gene were inserted into the Bbs1 site of pX330 and designated pX330-MTTP. The plasmids used in this study were confirmed by sequencing with an ABI 3130 genetic analyzer (Life Technologies).
Mouse monoclonal antibodies to HCV core, β-actin and Calnexin were purchased from Thermo Scientific and Sigma Aldrich, respectively. Mouse anti-ApoA1, ApoB, ApoC1, ApoE and ApoH antibodies were purchased from Cell Signaling, ALerCHEK Inc., Abnova, NOVUS Biologicals, and Santa Cruz Biotechnology, respectively. Rat anti-ApoA2 and Sheep anti-ApoC2 antibodies were purchased from R&D systems. Rabbit anti-NS5A antibody was prepared as described previously [54]. Alexa Fluor (AF) 488-conjugated anti-rabbit or mouse IgG antibodies, and AF594-conjugated anti-mouse IgG2a antibodies were purchased from Life Technologies.
A small interfering RNA (siRNA) pool targeting various apolipoproteins (siGENOME SMARTpool) and control nontargeting siRNA were purchased from Dharmacon, and transfected into cells using Lipofectamine RNAi MAX (Life Technologies) according to the manufacturer's protocol. A human shRNA library was purchased from Takara Bio Inc.
Upon transfection of pHH-JFH1-E2p7NS2mt or in vitro transcribed TH/JFH1, J6/JFH1 and S310/JFH1 RNA into Huh7.5.1 cells, HCV in the supernatant was collected after serial passages, and infectious titers were determined by a focus-forming assay and expressed in focus-forming units (FFU) [22], [23], [54]. To compare the localization of core protein, J6/JFH1 was used in Fig. 6E. Pseudoparticles expressing HCV envelope glycoprotein were generated in 293T cells as previously reported [5], and infectivity was assessed by luciferase expression using the Bright-Glo Luciferase assay system (Promega) and expressed in relative light units (RLU).
The lentiviral vectors and ViraPower Lentiviral Packaging Mix (Life Technologies) were co-transfected into 293T cells by Trans IT LT-1 (Mirus), and the supernatants were recovered at 48 h post-transfection. The lentivirus titer was determined by the Lenti-XTM qRT-PCR Titration Kit (Clontech), and the expression levels and AcGFP were determined at 48 h post-inoculation.
Cells lysed on ice in lysis buffer (20 mM Tris-HCl [pH 7.4], 135 mM NaCl, 1% Triton-X 100, 10% glycerol) supplemented with a protease inhibitor mix (Nacalai Tesque) were boiled in loading buffer and subjected to 5–20% gradient SDS-PAGE. The proteins were transferred to polyvinylidene difluoride membranes (Millipore) and reacted with the appropriate antibodies. The immune complexes were visualized with SuperSignal West Femto Substrate (Pierce) and detected by the LAS-3000 image analyzer system (Fujifilm).
Custom ZFN plasmids were designed to bind and cleave the ApoB, ApoE and MTTP genes and were obtained from Sigma Aldrich. Huh7 cells were transfected with in vitro transcribed ZFNs mRNA or pX330-MTTP by Lipofectamine 2000 (Life Technologies), and single cell clones were established by the single cell isolation technique. To screen for gene-knockout Huh7 cell clones, mutations in target loci were determined by using a Surveyor assay as previously described [56]. Frameshift of the genes and deficiencies of protein expression were confirmed by direct sequencing and immunoblotting analysis, respectively.
Protein concentrations of ApoB or ApoE in the culture supernatants were determined by using ELISA immunoassay kits (Alercheck Inc.) according to the manufacturer's protocol.
Total RNA was extracted from cells by using an RNeasy minikit (Qiagen) and the first-strand cDNA synthesis and qRT-PCR were performed with TaqMan EZ RT-PCR core reagents and a ViiA7 system (Life Technologies), respectively, according to the manufacturer's protocol. The primers for TaqMan PCR targeted to the noncoding region of HCV RNA were synthesized as previously reported [54]. Taqman Gene expression assays were used as the primers and probes targeting to apolipoproteins (Life Technologies). Fluorescent signals were analyzed with the ViiA7 system.
Cells cultured on glass slides were fixed with 4% paraformaldehyde (PFA) in phosphate buffered saline (PBS) at room temperature for 30 min, permeabilized for 20 min at room temperature with PBS containing 0.2% Triton after being washed three times with PBS, and blocked with PBS containing 2% FCS for 1 h at room temperature. The cells were incubated with PBS containing the appropriate primary antibodies at room temperature for 1 h, washed three times with PBS, and incubated with PBS containing AF488- or AF594-conjugated secondary antibodies at room temperature for 1 h. For lipid-droplet staining, cells incubated in medium containing 20 µg/ml BODIPY for 20 min at 37°C were washed with pre-warmed fresh medium and incubated for 20 min at 37°C. Cell nuclei were stained with DAPI. Cells were observed with a FluoView FV1000 laser scanning confocal microscope (Olympus).
The plasmid pSGR-JFH1 was linearized with XbaI, and treated with mung bean exonuclease. The linearized DNA was transcribed in vitro by using the MEGAscript T7 kit (Life Technologies) according to the manufacturer's protocol. The in vitro transcribed RNA (10 µg) was electroporated into Huh7 cells at 107 cells/0.4 ml under conditions of 190 V and 975 µF using a Gene Pulser (Bio-Rad) and plated on DMEM containing 10% FCS. The medium was replaced with fresh DMEM containing 10% FCS and 1 mg/ml G418 at 24 h post-transfection. The remaining colonies were cloned by using a cloning ring (Asahi Glass) or fixed with 4% PFA and stained with crystal violet at 4 weeks post-electroporation.
Intracellular viral titers were determined according to a method previously reported [10]. Briefly, cells were extensively washed with PBS, scraped, and centrifuged for 5 min at 1000× g. Cell pellets were resuspended in 500 µl of DMEM containing 10% FCS and subjected to three cycles of freezing and thawing using liquid nitrogen and a thermo block set to 37°C. Cell lysates were centrifuged at 10,000× g for 10 min at 4°C to remove cell debris. Cell-associated infectivity was determined by a focus-forming assay.
Correlative fluorescence microscopy-electron microscopy (FM-EM) allows individual cells to be examined both in an overview with fluorescence microscopy and in a detailed subcellular-structure view with electron microscopy. Cells infected with HCVcc were examined by the correlative FM-EM method as described previously [57].
Culture supernatants of cells infected with HCVcc were concentrated 50 times by using Spin-X UF concentrators (Corning), and the intracellular proteins collected after freeze-and-thaw were applied to the top of a linear gradient formed from 10–40% OptiPrep (Axis-Shield) in PBS and spun at 32,000 rpm for 16 h at 4°C by using an SW41 Ti rotor (Beckman Coulter). Aliquots of 10 consecutive fractions were collected, and the infectious titer and density were determined.
The proteinase K digestion protection assay was performed as described previously [37]. Briefly, cells were extensively washed with PBS, scraped, and centrifuged for 5 min at 1000× g. The cell pellets were resuspended in 500 µl of PBS and subjected to three cycles of freezing and thawing using liquid nitrogen and a thermo block set to 37°C. The cell lysates were centrifuged at 10,000× g for 10 min at 4°C to remove cell debris. The cell lysates were then incubated with 50 µg/ml proteinase K (Life Technologies) in the presence or absence of 5% Triton-X for 1 h on ice, and the digestion was terminated by addition of PMSF (Wako Chemical Industries).
The data for statistical analyses are the average of three independent experiments. Results were expressed as the means ± standard deviation. The significance of differences in the means was determined by Student's t-test.
|
10.1371/journal.ppat.1005865 | A Single Nucleotide Polymorphism in Human APOBEC3C Enhances Restriction of Lentiviruses | Humans express seven human APOBEC3 proteins, which can inhibit viruses and endogenous retroelements through cytidine deaminase activity. The seven paralogs differ in the potency of their antiviral effects, as well as in their antiviral targets. One APOBEC3, APOBEC3C, is exceptional as it has been found to only weakly block viruses and endogenous retroelements compared to other APOBEC3s. However, our positive selection analyses suggest that APOBEC3C has played a role in pathogen defense during primate evolution. Here, we describe a single nucleotide polymorphism in human APOBEC3C, a change from serine to isoleucine at position 188 (I188) that confers potent antiviral activity against HIV-1. The gain-of-function APOBEC3C SNP results in increased enzymatic activity and hypermutation of target sequences when tested in vitro, and correlates with increased dimerization of the protein. The I188 is widely distributed in human African populations, and is the ancestral primate allele, but is not found in chimpanzees or gorillas. Thus, while other hominids have lost activity of this antiviral gene, it has been maintained, or re-acquired, as a more active antiviral gene in a subset of humans. Taken together, our results suggest that APOBEC3C is in fact involved in protecting hosts from lentiviruses.
| The human APOBEC3 gene family consists of seven cytidine deaminases that mutate viral genomes. Compared to the other six human APOBEC3s, APOBEC3C has poor activity against viruses as well as endogenous retroelements, and its function remains poorly understood. Here, we report that although most humans express a version of APOBEC3C that only weakly blocks HIV, there is a polymorphism found in African populations that drastically enhances its anti-HIV activity. Furthermore, we demonstrate that the more active variant more efficiently deaminates cytidines in vitro and, in contrast to the common variant, forms dimers in solution. This polymorphism is absent in other hominids (chimpanzees and gorillas) but reverted or was maintained in some humans. Thus, while many humans have a “hole” in their innate defense against retroviruses, an ancient human polymorphism has restored this antiviral gene in some populations.
| The APOBEC3 locus encodes seven cytidine deaminase proteins that inhibit endogenous retroelements, lentiviruses such as HIV-1, and other viruses [1]. The APOBEC3 locus arose through duplication events on chromosome 22[2] of cytidine deaminase domains, resulting in single domain APOBEC3s (APOBEC3A, APOBEC3C, and APOBEC3H) and double-domain APOBEC3 genes (APOBEC3B, APOBEC3D, APOBEC3F, and APOBEC3G). In order for APOBEC3 proteins to restrict lentiviruses such as HIV-1, they are packaged into virions, brought to a target cell, and deaminate cytidines on ssDNA during reverse transcription, resulting in cytidine to uracil mutations in the viral genome. APOBEC3 proteins exert selective pressure on primate lentiviruses, which have evolved to encode a protein, Vif, which targets APOBEC3 proteins for proteasomal degradation.
Over long evolutionary time-periods, Vif-mediated antagonism of APOBEC3 proteins in populations infected with a lentivirus selects for polymorphisms in the population that have acquired mutations in APOBEC3 that allow for escape from Vif but maintenance of antiviral activity [3]. Lentiviruses, in turn, select for Vif alleles that target these APOBEC3 variants, leading to further adaptive evolution of APOBEC3 genes through selection for mutations that allow that host to evade viral infections. As such, enrichment of the rate of nonsynonymous mutations (dN) compared to the rate of synonymous mutations (dS), called positive selection (defined as dN/dS>1), is a common signature of antiviral genes [3]. APOBEC3 genes involved in blocking viral replication are expected to exhibit signatures of positive selection. Specifically, APOBEC3s involved in lentiviral restriction should have signatures of positive selection at the Vif:APOBEC3 interface [4].
There is considerable variation in the antiviral activity of each of the seven human APOBEC3 paralogs. APOBEC3G potently inhibits vif-deleted-HIV-1 (Δvif) [5]. Human APOBEC3D, APOBEC3F, and APOBEC3H also inhibit HIV-1 (Δvif), but to a lesser extent than APOBEC3G [5–8]. In contrast, APOBEC3A and APOBEC3B do not potently block HIV infection of T cells [5–7, 9], which are the primary target of HIV (although a target-cell effect has been reported in monocytes for APOBEC3A) [10]. Instead, APOBEC3A and APOBEC3B drastically inhibit replication of endogenous retroelements such as LINE-1 elements, as well as some DNA viruses [11–16]. In studies that compare the ability of the seven human APOBEC3s to restrict lentiviruses and endogenous retroelements, the only APOBEC3 that has weak activity against both lentiviruses and endogenous retroelements is APOBEC3C [5, 6, 12, 16–22].
For another APOBEC3 gene, APOBEC3H, the most common human variant does not block HIV infection although other human haplotypes exist that potently restrict lentivirus replication [23]. In fact, one haplotype of APOBEC3H restricts HIV-1(Δvif) nearly as potently as APOBEC3G [23] and has been shown to impact clinical outcomes in HIV-1+ patients [24–26]. Thus, we considered the possibility that while the common human haplotype of APOBEC3C encodes a protein with little antiviral activity, other variants of APOBEC3C may, in fact, encode more potent anti-lentiviral proteins. Moreover, the Vif protein of HIV-1 targets human APOBEC3C for proteosomal degradation [27]. In addition, APOBEC3C mRNA is highly expressed in the major HIV-1 target cells, activated T cells[28]. Thus, the high expression of APOBEC3C in HIV target cells and the antagonism of APOBEC3C by HIV-1 Vif are consistent with the hypothesis that APOBEC3C may have an overlooked role in combating lentivirus infection.
In this study, we found that APOBEC3C has evolved under positive selection in primates, in a manner that suggests that APOBEC3C has played a role in blocking primate lentiviruses. This provided motivation to determine if there are naturally occurring variants of APOBEC3C that potently block lentivirus replication. In humans, only one APOBEC3C coding variant is present at a frequency above 1% and this is a serine to isoleucine change at position 188, here called APOBEC3C I188 [29]. We show that the polymorphism APOBEC3C I188 is present at about 10% frequency in diverse populations throughout Africa, and thus did not recently arise in a particular subpopulation of humans, but is an ancient allele that has likely been circulating in humans for much of human history. Moreover, we show that the APOBEC3C I188 single nucleotide polymorphism (SNP) has about 10-fold more potent anti-lentiviral activity than the common human APOBEC3C variant and has greater in vitro cytidine deaminase specific activity. The greater activity of APOBEC3C I188 in turn correlates with its ability to dimerize. Moreover, construction of a forced dimer of APOBEC3C S188 also gains enhanced antiviral activity to a level comparable to APOBEC3G. We also show that the APOBEC3C I188 allele is likely the ancestral state since all sequenced Old World monkeys and some great apes carry isoleucine at position 188. However, gorillas, chimpanzees and most humans carry the S188, the apparent loss of function allele. Taken together, our results suggest that APOBEC3C is involved in protecting hosts from lentiviruses, and we speculate that some humans may be afforded some level of additional protection from lentiviruses by a more active antiviral version of this protein.
In studies that compare the antiviral activity of the seven APOBEC3 paralogs, APOBEC3C consistently has poorer restriction activity than the other paralogs [5, 6, 18, 20, 21]. However, we reasoned that if APOBEC3C is in fact a bona-fide restriction factor then we would expect that the gene has an evolutionary signature of positive selection [3]. We performed positive selection analyses of twenty-two APOBEC3C sequences derived from eighteen primate species with sequences representing diverse clades of catarrhines, a subdivision of primates including old world monkeys and apes (Fig 1A). Among these, multiple sequences were obtained from African green monkeys, because we chose to include three subspecies (vervet, tantalus, and sabeus). The sequences were aligned and tests for positive selection were conducted using maximum likelihood ratio tests comparing M8 (a model that allows positive selection across the gene) to M8a (a model that disallows positive selection). Our results indicate APOBEC3C shows a gene-wide signature of positive selection (p<0.0008) (Fig 1B).
We next analyzed individual lineages to determine which branches of the APOBEC3C tree have signatures of positive selection. Branch analysis identified two branches with statistically significant signatures of positive selection, both in Old World monkeys (Fig 1A), and while most were not statistically significant, many branches had a dN/dS >1 (Fig 1A). Furthermore, we performed M8 vs M8a analysis of the hominoid and Old World monkey clades of the tree separately, and found that the Old World monkey clade has a statistically significant signature of positive selection (p<0.05) (Fig 1B). We did not see a statistically significant signature of positive selection in the hominoid-only branch (p = 0.15), although this could be due to a smaller sample size (n = 7).
For antiviral genes, sites under positive selection often correlate with sites of interaction with a viral antagonist [30]. APOBEC3C is antagonized by the lentiviral protein Vif and the interface of Vif binding has been extensively mapped [31]. If APOBEC3C is in fact an anti-lentiviral gene, the Vif binding interface may be evolving under positive selection. Therefore, we performed a site-analysis to determine which amino acids are under positive selection across the tree. Our analysis indicated seven sites under positive selection (posterior probability > 99%) (Fig 1C). Next, we mapped the positively selected sites onto the structure of human APOBEC3C and compared these to the Vif interface of APOBEC3C. Of the seven positively selected sites, two of these, residues 106 and 77, are located within the two helices that are targeted by Vif (Fig 1C). Strikingly, residue 106 has been identified as being a critical amino acid for the interaction of APOBEC3C with Vif [27, 31]. Thus, APOBEC3C has evolved under selection, gene-wide, as well as at the Vif-binding interface. These results suggest that although the common human APOBEC3C variant does not potently block lentivirus replication, primate APOBEC3C may have evolved as an anti-lentiviral protein.
Because the positive selection analyses suggested an ancient or ongoing role of APOBEC3C in lentiviral restriction (Fig 1), we re-evaluated human polymorphisms in APOBEC3C for potential variants with increased activity. There is only one SNP in APOBEC3C above 1% frequency globally, and this is a serine to isoleucine change at position 188 [29]. To evaluate the potential significance of this SNP, we aligned this region of APOBEC3C to other human APOBEC3 genes. Strikingly, we found that in contrast to APOBEC3C, the other ten APOBEC3 deaminase domains all encode a conserved isoleucine at the position homologous to APOBEC3C 188 (Fig 2). Thus, the human I188 polymorphism in APOBEC3C actually encodes an amino acid that is highly conserved at this position across human APOBEC3s, while the more common APOBEC3C in the human population has a different amino acid at position 188.
Since conserved sequences are often important for function and comparative studies indicate that human APOBEC3C (S188) has weak antiviral/anti-retroelement activity compared to the other human APOBEC3s, we posited that the serine change may contribute to the weak restriction activity of the common variant of APOBEC3C. Therefore, we directly compared APOBEC3C S188 and APOBEC3C I188 for their ability to restrict HIV-1. We transfected the two APOBEC3C variants, S188 and I188, along with VSV-G and an env- vif- deleted luciferase-expressing HIV-1 provirus (Δenv, Δvif). Equal amounts of virus, normalized by p24gag, were subsequently used to infect SupT1 cells and infectivity of the viruses was compared by measuring virus-encoded luciferase. Viral infectivity in the presence of no APOBEC3 is set to 100%. APOBEC3G was used as a positive control because it potently inhibits HIV-1 (Δvif). We found that APOBEC3C I188 restricts infectivity of HIV-1(Δvif) to a level approximately ten-fold greater than the common APOBEC3C, S188, (approx. 30% infectivity versus 3%, respectively) (Fig 3A) even though both proteins are expressed at similar levels. Furthermore, infectivity assays were conducted as a dose-response in the presence of decreasing concentrations of APOBEC3, and the 188 isoleucine variant restricts HIV-1(Δvif) more potently for all conditions (Fig 3B) at similar protein expression levels.
To determine if the APOBEC3C I188 variant has increased potency against another lentivirus, we evaluated its activity against SIVagm (simian immunodeficiency virus that infects African green monkeys). As shown by others, the S188 variant of APOBEC3C restricted infectivity of SIVagm to a greater extent than HIV-1 [9]. However, we found that APOBEC3C I188 restricted SIVagm infectivity ten-fold more than the restriction caused by APOBEC3C S188 (10% versus 1% infectivity, respectively, p<0.05) (Fig 3C). Some APOBEC3s also restrict endogenous retroelements, such as LINE-1s [14, 15]. However, the APOBEC3C I188 variant does not confer increased restriction of LINE-1 as we have previously published [29] and have repeated for this study (S1 Fig). Therefore, the human polymorphism in APOBEC3C at position 188 enhances restriction of at least two primate lentiviruses. Thus, we conclude that a SNP in human APOBEC3C has increased anti-lentiviral activity relative to the APOBEC3C encoded by most humans.
We wished to investigate whether or not the more potent antiviral activity of APOBEC3C I188 compared to APOBEC3C S188 could be explained by differences in their inherent enzymatic activity. Thus, each protein was produced by expression in a recombinant baculovirus system, purified as described in the Materials and Methods, and tested for its ability to cause cytidine deamination. We examined APOBEC3C S188 and I188 activity using a ssDNA substrate containing two deamination target motifs (Fig 4, top sketch). 5' TTC deamination motifs were used because APOBEC3C preferentially targets this motif [21]. Reactions were carried out as a time-course over 60 minutes and next the substrates were incubated with uracil DNA glycosylase, which modifies uracil-containing DNA and makes it sensitive to cleavage at high pH. Cytidine to uracil mutations leading to DNA cleavage were detected based on a fluorescein label placed between the two deamination motifs (Fig 4, top). Substrate usage was calculated from integrated gel band intensity of cleaved product at either deamination motif relative to the uncleaved substrate (Fig 4). We found that at all time points substrate usage of APOBEC3 I188 was higher than S188, and by 60 minutes I188 had led to twice as many cleavage events as S188 (Fig 4, middle and bottom left). The specific activity of APOBEC3C was determined by calculating the picomoles of substrate used (or deamination events) per microgram of enzyme per minute on a 118 nt ssDNA. The specific activity values were calculated using initial reaction times where the substrate usage was in the linear range (Fig 4, bottom left). We found that APOBEC3C S188 had a specific activity approximately 10-fold lower than I188 (0.010 pmol/μg/min vs 0.130 pmol/μg/min) (Fig 4, bottom right). Therefore, the I188 APOBEC3C more rapidly deaminated cytosines in vitro than S188.
Since APOBEC3C I188 has greater cytidine deaminase activity in vitro than APOBEC3C S188 (Fig 4), we predicted that it would also have a higher mutational frequency than the APOBEC3C S188. To test this prediction, we used a model in vitro system that reconstitutes reverse transcription of RNA to DNA, and observed the ability of APOBEC3 enzymes to induce mutagenesis. The template includes the gene lacZα, and blue/white screening was performed to identify mutated reverse transcription products. White colonies, representing templates that were mutated, were then sequenced and the number of mutations induced by each APOBEC3 were quantified. We found that addition of APOBEC3C I188 induced two-fold higher clonal mutation frequency compared to APOBEC3C S188 (S2 Fig, 0.33 x 10−2 mutations/bp versus 0.15 x 10−2 mutations/bp, respectively). For reactions containing APOBEC3C S188, 100% of clones had zero to one G→A mutation. In contrast, the presence of APOBEC3C I188 caused a noticeable shift in the number of G→A mutations with 32% of clones having more than one mutation and up to four to five mutations in some individual clones (S2 Fig). Overall, isoleucine at position 188 increased the APOBEC3C-induced mutagenesis of ssDNA in vitro.
Previous studies have reported that the S188 variant of APOBEC3C is a monomeric protein, both in solution [31] and in cells [32]. Indeed, by size exclusion chromatography we also found that baculovirus/Sf9-produced APOBEC3C S188 (the common variant) is monomeric (Fig 5A). However, the baculovirus-produced APOBEC3C I188 was in equilibrium between monomer and dimer forms (Fig 5A, apparent molecular weight 21 kDa and 42 kDa, respectively). We confirmed this result using an alternative method of cross-linking the proteins in solution followed by SDS-PAGE and Western blotting. Baculovirus/Sf9-produced APOBEC3C S188 or I188 were incubated in the absence or presence of 10 μM bis(sulfosuccinimidyl)suberate (BS3), an amine-amine chemical crosslinker, and then visualized through SDS-PAGE and Western blotting (Fig 5B). A3C S188 remained monomeric in the presence of crosslinker, whereas A3C I188 was partially dimeric in the presence of the crosslinker. The observation that the isoleucine residue at position 188 was able to shift the oligomeric profile of APOBEC3C suggests that residue 188 is important for dimerization. Dimerization has been previously correlated with improved APOBEC3 catalytic activity because it enables efficient scanning of ssDNA to find cytosine targets for deamination [33]. This provides a potential explanation for the increased in vitro enzymatic activity of A3C I188.
In order to further test the effects of dimerization of A3C on antiviral activity, we constructed an artificial dimer that consists of two tandem S188 APOBEC3Cs (Fig 6A) and tested the anti-lentiviral activity of this protein. We used the linker that naturally exists between the N- and C-terminal domains of the two double-domain APOBEC3s, APOBEC3D and APOBEC3F, which are the APOBEC3 proteins with the highest sequence identity shared with APOBEC3C. This linker consists of amino acids Arg-Asn-Pro followed by the second APOBEC3 domain starting at Met12 (labeled Met12’ here—see schematic at top of Fig 6A). Western blot analysis shows that this artificial double domain APOBEC3C is expressed in cells and runs at about the same molecular weight as the natural double domain APOBEC3 protein, APOBEC3G (Fig 6A).
We examined the antiviral activity of the synthetic dimer APOBEC3C gene (with S188 in both domains, called S188-S188) compared to APOBEC3C S188 and APOBEC3C I188 (Fig 6A) against HIV-1Δ(vif). Again, APOBEC3G was used as a positive control. While the APOBEC3C I188 restricted 5–10 fold better than APOBEC3C S188 (Fig 6A: 60% infectivity compared to 8% infectivity), strikingly, APOBEC3C S188-S188 dimer restricted infection as efficiently as APOBEC3G (Fig 5B: approximately 1% infectivity for both conditions). Importantly, the APOBEC3C S188-S188 synthetic dimer restricts infection far greater than two-fold more than the APOBEC3C S188 monomer (Fig 6A: 60% infectivity relative to 1% infectivity), suggesting that the increased antiviral activity is not simply the result of having twice as many active sites. Thus, these results indicate that forced dimerization is sufficient to induce anti-HIV activity of APOBEC3C regardless of the isoleucine at position 188.
In a separate series of experiments, we also examined the ability of each of the A3C variants to be packaged into virions. We found that A3C I188 was not packaged to a greater extent than A3C S188 (compare Fig 6B lanes 2 (A3C I188) to Fig 6B lanes 3 (A3C S188). Thus, the greater activity of A3C I188 correlates better with its increased enzymatic activity than with virion packaging. On the other hand, the synthetic dimer of A3C S188-S188 is packaged into virons 10–20 fold better than the single domain versions of A3C (Fig 6B, lanes 4). This increased packaging could additionally explain the enhanced antiviral activity of the synthetic dimer. This suggests that while natural dimers of APOBEC3C have increased enzymatic activity, a synthetic dimer of an APOBEC3 protein can be created with improved antiviral activity due to increased packaging into virions
In the absence of direct clinical or cohort data, we next sought to further evaluate the relevance of APOBEC3C to HIV infection. Previous studies had found that APOBEC3C mRNA is well expressed in primary T cells, which are the target cells of HIV-1 [28]. We further reasoned that if APOBEC3C is indeed a restriction factor relevant to HIV, then one would expect it to be antagonized by the viral Vif protein. To test this, we produced HIV-1 (either lacking vif, or expressing either HIV-1 or HIV-2 vif) in the presence of APOBEC3C. When we express APOBEC3C I188 during HIV production, the infectivity of the virus is reduced by about ten-fold. However, in the presence of APOBEC3 S188 or I188, both HIV-1 Vif and HIV-2 Vif, restored viral infectivity (Fig 7A). We also conducted western blot analysis to probe for APOBEC3C (S188 and I188) expression in the presence of HIV-1 and HIV-2 Vif proteins (Fig 7B). Consistent with other reports, A3C S188 protein levels are significantly decreased in the presence of HIV-1 Vif [27, 31] as well as HIV-2 Vif [34]. Likewise, the expression of the A3C I188 variant also dramatically decreased in the presence of HIV-1 and HIV-2 Vif. Thus, APOBEC3C I188 effectively antagonized by HIV-1 and HIV-2 Vif which suggests that even the more active form of APOBEC3C in its partial dimer form can still be targeted by both human lentiviral pathogens. These results suggest that APOBEC3C is relevant to HIV infections since Vif has evolved to induce its degradation and APOBEC3C is expressed in HIV target cells.
The APOBEC3C I188 variant is present at frequency of 2.4% in the 1000 Genomes Project [29]. Therefore, a relatively small proportion of humans carry a variant of APOBEC3C that is more enzymatically active against lentiviruses. To determine which allele is ancestral at position 188, we constructed a phylogeny of primate APOBEC3C sequences. All old world monkeys (N = 15) analyzed encode an isoleucine at position 188 (Fig 8A). Moreover, orangutans, siamangs, and gibbons also encode isoleucine, but the serine change at amino acid 188 occurred in the lineage leading to gorillas, chimpanzees, and humans (Fig 8A). Thus, isoleucine at position 188 is likely the ancestral state, and changed during the evolution of hominids. There are two possible explanations for the existence of the I188 in humans: 1) a reversion back to isoleucine may have occurred in a subpopulation or 2) a polymorphism has been maintained at this site for millions of years, since humans split from their ancestor with gorillas and chimpanzees.
If a serine to isoleucine reversion mutation occurred in recent human evolution, we would expect it to be present only in a limited subset of humans. The frequency of the allele in the 1000 Genomes Project data is 8.9% in populations of African descent, less than 1% frequency in the Americas, and not present in Asia and Europe [29] (Fig 8B). Humans are dramatically more genetically diverse in Africa than on any other continent, therefore we sought to determine if the APOBEC3C I188 allele is distributed across divergent populations in Africa, or if it is present in only a particular subpopulation. The APOBEC3C I188 allele is present in all six African subpopulations analyzed by the 1000 Genomes project, with a frequency ranging between 5.6% and 13% (Fig 8B). However, many of the sub-populations included in the 1000 Genomes Project live in regions affected by the Bantu Expansion, a migration event when Bantu-speaking tribes swept across the continent approximately 3,000 years ago[35, 36]. To determine if the isoleucine allele is present in more diverse African genomes, we determined the APOBEC3C sequence from individuals from four hunter-gatherer groups (Hadza, Sandawe, Mbuti, and Khoe-San)[37, 38]. We found that one of the four Khoe-San individuals was heterozygous for the I188 allele, and two out of five Sandawe individuals were heterozygous for the I188 allele (Fig 8B). In conclusion, I188 seems to be a widely distributed SNP in African populations suggesting that the more active allele is very ancient, and may have even been circulating in humans since the birth of the species. Presence of the I188 in the ancient human relative Homo neanderthalensis would have provided evidence that the allele has been present in the Homo lineage for at least 600,000 years but we failed to find the I188 SNP in the published Neanderthal genomes.
To determine if other hominoids also possess variation at position 188 we probed the APOBEC3C sequences from the Great Ape Genome project[39], and found that none of the great apes included in the study (n = 79) were polymorphic at position 188 (Fig 8C). Ten orangutans were included in the study, and all encoded isoleucine at position 188. In contrast, all gorillas (n = 31), and chimpanzees and bonobos (n = 38), encoded serine at position 188. Humans, gorillas, and chimpanzees diverged from their most recent common ancestor approximately 10 to 20 million years ago[40, 41], and in this ancestral lineage the more active isoleucine allele was lost. However, since some humans express the I188 allele, it is possible S188 never rose to fixation and I188 was maintained as a minor allele for a long period of the evolutionary history of hominoids. Alternatively, it is possible that serine became fixed in the ancestor to gorillas, chimpanzees and humans, but more recently the serine reverted to isoleucine in a subpopulation of humans. Nonetheless, we find that the APOBEC3C I188 is relatively ancient to humans, but is not present to an appreciable extent in out-of-Africa human populations, nor have we found it in other hominids.
APOBEC3C stood out among the seven human APOBEC3 paralogs as it little antiviral or anti-retroelement activity. We observed that the six APOBEC3s with known functions possess a conserved isoleucine at the residue homologous to APOBEC3C position 188, whereas APOBEC3C encodes a serine at this position. However, human APOBEC3C is, in fact, polymorphic at this site, and some humans encode an isoleucine, the residue that correlates with APOBEC3 antiviral/anti-retroelement function. This led us to hypothesize that APOBEC3C may have an as yet overlooked role as a restriction factor, and that the I188 variant may have enhanced antiviral activity compared to the more common variant, S188. APOBEC3C has evolved under positive selection in primates and within the interface of binding by the viral protein Vif, suggesting that this gene may have played a role in restriction of lentiviruses over primate evolution. Furthermore, we found that APOBEC3C I188 encodes a protein with increased antiviral activity, increased enzymatic activity, and the ability to dimerize in solution. Consistent with this conclusion, an artificial forced dimer of APOBEC3C S188 has vastly increased antiviral activity. We find that the isoleucine at position 188 was lost during hominid evolution but was either reacquired by some humans since humans split with our most recent common ancestor with chimpanzees, or alternatively, has never been lost as an allele and has been maintained as a polymorphism through several million years of hominoid evolution.
Previous studies have shown that APOBEC3C binds to HIV-1 Vif and that E106 is important for Vif binding since mutation to lysine at position 106 completely abrogated HIV-1 Vif binding to APOBEC3C [27, 31]. We found that this residue within the Vif binding interface is evolving under positive selection, and another residue in the Vif-binding region, 77, is also under positive selection. Residue 77 is within the α-2 helix of APOBEC3C, which has also been shown to be important for HIV-1 Vif binding [31]. Additionally, it is possible that Vifs from other lentiviruses target APOBEC3C at different motifs, driving the positive selection in other regions of the protein. For example, APOBEC3C is under positive selection at residues 128 and 130. While these residues are not in the known APOBEC3C:HIV-1 Vif binding interface, the homologous residues of APOBEC3G are involved in HIV-1 Vif binding[42, 43]. Therefore, it is possible that other Vif proteins from other lentiviruses target APOBEC3C at positions 128 and 130, or that ancient lentiviruses have targeted these residues in the past. In summary, rapid evolution of APOBEC3C at the known APOBEC3C:Vif binding interface suggests that APOBEC3C has evolved to block lentiviruses in primates.
Our results indicate that the difference in the anti-HIV activity of the APOBEC3C variants S188 and I188 lies in the enzymatic efficiency of the two APOBEC3C proteins. We found that I188 more rapidly deaminates ssDNA in vitro. Furthermore, in an in vitro RT model system, the presence of APOBEC3C cause a higher mutation frequency than APOBEC3C S188. A previous study correlated multimerization of APOBEC3s with the capacity to restrict lentiviruses [32], and our finding that the monomeric variant (S188) was less antivirally active than the dimer-forming, more active variant (I188), is consistent with this conclusion. Therefore, our model is that isoleucine at position 188 of APOBEC3C enhances lentiviral restriction by improving dimerization and in turn, the enzymatic activity of the protein. One possible reason dimerization is important for APOBEC3C activity, is that it could improve the protein’s ability to scan DNA substrates for cytidine deamination motifs. In fact, I188 lies within α-helix 6, which has been implicated as important for DNA scanning of another APOBEC3, APOBEC3G [44]. The residue at position 188 may not be directly involved in the dimer interface since it is not surface exposed on the crystal structure of A3C I188 (PDB #3VOW)[31]. Nonetheless, our data suggest that the APOBEC3C I188 protein has greater antiviral activity than the more common APOBEC3C protein due to better enzymatic activity that correlates with increased dimerization.
This model that dimerization is a key determinant of APOBEC3C activity is further supported by the fact that a synthetic dimer formed by linking two tandem S188 APOBEC3Cs drastically enhances antiviral activity. In fact, activity is improved even in comparison to I188, the more active variant. I188 only partially dimerizes, and compared to S188 and S188-S188, has an intermediate ability to restrict HIV. Interestingly, the mechanism of increased antiviral activity of A3C S188-S188 is likely due to its increased ability to be packaged into virions. These results suggest that artificial forms of human APOBEC3C proteins can be created that have enhanced antiviral properties that could have therapeutic uses in controlling viral infection.
The isoleucine at position 188 of APOBEC3C is present at approximately 10% frequency across diverse African populations, but almost absent from all other global populations. All human populations outside of Africa are thought to have descended from one or a few migration events out of Africa[45]. As such, humans from non-African populations may lack the APOBEC3C I188 allele because it was excluded in a population bottleneck during the migrations. Or, the allele may have been lost in non-African populations due to drift or a lack of selective pressure. Alternatively, it is possible that loss of the allele was selected for non-African populations. Expression of another APOBEC3, APOBEC3B, has been associated with increased risk of cancer [46, 47]. Therefore, the antiviral function of APOBEC3s may come at an evolutionary trade-off. In fact, this may have driven the maintenance of the less enzymatically active S188 allele for millions of years in humans and ancient human ancestors.
Our phylogenetic analysis shows that APOBEC3C I188 is ancestral in primates, but changed to serine in the clade of apes including gorillas, chimpanzee, and humans. The fact that humans have a polymorphism that corresponds with the ancestral residue could be due to a reversion back to the amino acid present in other primates, but not in gorillas nor chimpanzees. If a reversion occurred it must have happened long ago in human history, since the allele is present in such deeply divergent populations across Africa. However, the allele was likely lost due to a bottleneck in the out-of-Africa populations because it is almost completely missing from non-African populations. Alternatively, it is possible that the isoleucine allele has continued in the human lineage through incomplete lineage sorting (the maintenance of a polymorphism after the divergence of species), since before humans split with their most recent common ancestor with gorillas more than 10 million years ago. Notably, the isloleucine codon, ATT, at position 188 is the same in the human SNP as in all other primates with an Ile at this position in APOBEC3C. While we did not find support for incomplete lineage sorting since we did not find any other hominids that were polymorphic at position 188, the limited number of great ape sequences were included does not allow us to completely rule out this second possibility. Nonetheless, given the increased antiviral activity of APOBEC3C I188 and its fixation in primates other than hominids argues that the gain (or maintenance) of this allele in humans has been driven by a function for protection against pathogens.
We discovered that an APOBEC3C single nucleotide polymorphism (SNP) that is common in Africa enhances anti-lentiviral activity. This polymorphism may impact human susceptibility to cross-species transmissions of lentiviruses because Vifs from other lentiviruses may not antagonize human APOBEC3C. HIV-1 and HIV-2 Vif are able to antagonize both variants of APOBEC3C so the I188 SNP may not block HIV transmission, so Vif may effectively counteract I188 activity during infection. However, the fact that APOBEC3C is antagonized by Vif does suggest that APOBEC3C is an important barrier that must be countered by the virus during natural infections. Alternatively, it is possible that APOBEC3C antagonism by Vif is an unintended consequence due to Vif binding to another APOBEC3 such as APOBEC3F since APOBEC3C has a Vif binding pocket that is nearly identical to the Vif binding pocket of APOBEC3F[27, 31, 48]. Despite the ability of Vif to antagonize APOBEC3C, it is possible that APOBEC3C I188 still influences HIV susceptibility. In infected individuals possessing the whole APOBEC3 repertoire, Vif has to adapt to counteract multiple antiviral proteins and this may constrain Vif and weaken its activity. In fact, viral genomes sequenced from HIV-1-infected patient cells are extensively mutated by APOBEC3s despite the presence of Vif [49, 50] and the extent of APOBEC3-induced mutagenesis negatively correlates with disease progression rate [51]. As such, it is possible that APOBEC3C I188 may provide some level of protection from HIV transmission or pathogenesis.
APOBEC3C was amplified by RT-PCR from total RNA extracted from chimpanzee, gorilla, orangutan, white-cheeked gibbon, siamang, baboon, sooty mangabey, and red-capped mangabey, and proboscis monkey cells (either fibroblast or lymphoid) obtained from Corriell Repository as well as from the vervet monkey cell line Vero, the tantalus monkey cell line CV-1, and the sabeus cell line V038 provided by the Nonhuman Primate Research Resource (NPRR). Primers were designed to amplify from the 3’ and 5’ UTRs of APOBEC3C mRNA transcripts (5’UTR: CTAAGAGGCTGAACATGAATC’3, 3’UTR: 5’GGCTAGAGGAGACAGACCATGA’3). The APOBEC3C amplicons were cloned into pGEM vectors, and then sequenced. The S188-188 forced dimer was designed to mimic the linker between the two domains of the double-domain APOBEC3F. The N-terminal subunit consists of APOBEC3C residues 1–189 (residue 190 is removed), followed by the residues RNP, which serve as a linker. The C-terminal APOBEC3C begins at the second start codon, M12. The dimer S188-S188 APOBEC3C was constructed by overlap extension PCR. Two separate PCRs were performed for the N terminal and C terminal APOBEC3C subunits (1st domain, For: TTCAGGATCCATGAATCCAGAGATC, 1st domain, Rev: GCCTCCATTGGGTCCCGGAGACTCTCCCGTAGCCTTCTTT, 2nd domain, For: TCCAGGATCCATGAATCCACAGATC, 2nd Rev: GCCCTCTAGATTAGGCGTAGTCAGG), and these amplicons were annealed in a third PCR reaction using the 1st domain For and the 2nd domain Rev primers.
APOBEC3C genes were aligned using Geneious software. To test for positive selection, maximum likelihood tests were performed using the PAML statistical software suite [52]. The APOBEC3C genes were subjected to tests that allowed for positive selection (M8 model), or disallowed positive selection (M8a model). The analyses were performed with the F3X4 codon model, and multiple starting omega values were used, ranging between 0.5 and 1.4. Specific residues with signatures of positive selection with a posterior probability of 99% or greater were identified by Bayes Empirical Bayes analysis. Ancestral APOBEC3C sequences were reconstructed by the likelihood/Empirical Bayes approach using the codeml program in PAML. Brach analysis to identify particular primate branches with signatures of positive selection in APOBEC3C were performed in two ways. Overall dN/dS values were calculated with PAML, using the free ratio model. Additionally, a branch-site test to identify statistically significant signatures of episodic selection was performed using the Branch-site REL method in the HyPhy software suite [53].
APOBEC3Cs were cloned into the BamHI and XhoI sites of pCDNA3.1 by PCR addition of restriction sites (BamHI and XhoI) to the N and C termini of APOBEC3C. The human APOBEC3C plasmid we previously obtained from the AIDS Repository contained the SNP rs11551111, which is not common (no reported frequency according to dbSNP). Therefore, we used site-directed mutagenesis to change the asparagine at position 31 to aspartic acid (For: GCCAACGATCGGGACGAAACTTGGC, Rev: GCCAAGTTTCGTCCCGATCGTTGGC). A hemagglutinin tag was inserted into the XhoI and XbaI sites of pCDNA3.1, at the C-terminus of each APOBEC3C sequence. APOBEC3G and APOBEC3A were also in a pCDNA3.1 backbone, with a Kozak sequence, as well as a hemagglutinin tag at the N-terminus. HIVΔenv,Δvif, HIVΔvif + HIV-1 vif, HIVΔvif +HIV-2 vif have been described elsewhere [54]. SIVagm Δenv, Δvif was kindly provided by Nathaniel Landau.
Single round HIV-1 and SIVagm infectivity assays were performed as previously described [55]. 293T cells (American Type Culture Collection) were plated at a density of 5 X 103 cells per well of a 24-well plate. The next day, the cells were transfected with 0.3μg provirus encoding luciferase as a marker gene 0.1μg pL-VSV-G, and 0.3μg pCDNA3.1.APOBEC3.HA or empty pCDNA3.1 plasmid. For the dose response infectivity assay, either 0.1 μg, 0.2μg, or 0.3μg APOBEC3 plasmid was used. For experiments involving Vif expression, 0.2μg of APOBEC3 was used. Forty-eight hours after transfection, virions were harvested. For SIVagm infectivity assays, SupT1 cells were infected with 10μl of each virus and treated with 20μg/ml DEAE/dextran. For HIV infectivity assays, ELISA was performed to quantify p24, and virus equivalent to 2ng p24 was used for infections. For all infectivity assays, 5 X 104 were infected in a 96 well dish. Seventy-two hours later, infected cells were lysed in luciferase lysis reagent (Brightglo, Promega) and luciferase expression was measured on a luminometer (LUMISTAR Omega, BMG). Infectivity of each virus was compared by setting infectivity of the “No APOBEC3” control to 100%. All HIV-1 constructs are based on the LAI strain.
To assay for restriction of LINE-1 retrotransposition 293T cells were transfected with 200ng LINE-1 plasmids pYX016 and pYX015[56], along with 100ng of APOBEC3C S188 or I188, APOBEC3C, 10ng APOBEC3A, or empty pCDNA3.1 plasmid. The next day, the cells were treated with 2.5 ug/ul puromycin to select for transformants. Three days later, expression of renilla and firefly luciferase were assayed using a luminometer. The LINE-1 plasmids encode firefly luciferase disrupted by a splice site, so expression only occurs after retrotransposition, whereas renilla luciferase expression is not dependent upon retrotransposition. Percent retrotransposition is reported by setting retrotransposition (firefly luciferase values divided by renilla luciferase values) in the absence of APOBEC3 to 100%.
Intracellular expression of the APOBEC3 proteins during virion production was evaluated by lysis of the virion-producing 293T cells with Radio Immunoprecipitation Assay buffer (RIPA), with protease inhibitor (50mM Tris, 150mM sodium chloride, 0.1% SDS, 0.5% sodium deoxycholate, 1% NP-40, protease inhibitor cocktail cOmplete by Roche). Lysates were resolved on an SDS-PAGE gel in MES buffer, and transferred to a PVDF membrane for Western blot analysis, using and anti-HA (BioLegend) antibody and anti-tubulin (Sigma-Aldrich) antibody. Endogenous levels of APOBEC3C were measured by Western blotting with antibody purchased from Fisher (product # PA5- 27629). HRP-conjugated secondary antibodies (Santa Cruz) were used to detect primary antibodies.
Packaging of APOBEC3 into virions was evaluated by co-transfection of 100 ng of each APOBEC3 expression plasmid with 500 ng of an HIV proviral clone (LAI) containing a deletion in vif in each well of a 12-well plate. Three days after transfection, 1 ml of supernatant was collected, filtered through a 0.2 micron filter, and concentrated by pelleting in a microcentrifuge at 13K rpm for 60 minutes and resuspended in 80 μl. The amount of p24gag was determined by ELISA (Advanced Bioscience Laboratories). Equal quantities of p24gag were lysed and run on an SDS-PAGE gel. The Western blots were probed with an anti-HA antibody for A3C protein and with a p24gag antibody for virus production and HRP-conjugated secondary antibodies were used to detect primary antibodies. Cells were lysed as described above. The chemiluminescent signals from the Western blots were imaged using a ChemiDoc MP Imaging System (Bio-Rad) and quantified in the linear detection range.
Recombinant baculovirus production for APOBEC3C S188 was carried out in the pACG2T transfer vector (BD Biosciences), as described previously [57]. Recombinant baculovirus production for APOBEC3C I188 was carried out in the pFastbac1-GST-APOBEC3C vector according to the Bac-to-Bac expression system (Life Technologies) and as described previously [58]. Recombinant virus was then used to infect Sf9 cells. Cells were harvested 72 hours after infection, lysed, treated with RNaseA, and clarified cell lysates were incubated with glutathione-sepharose 4B resin (GE Healthcare) at 4°C and subjected to a series of salt washes, as described previously[59]. The APOBEC3C S188, APOBEC3C I188 enzymes were eluted from the glutathione-sepharose resin (GE Healthcare) with the GST tag, as previously described [59]. The samples were then treated with thrombin (GE Healthcare) for 6 hr at 21°C to cleave the GST tag.
The oligomerization states of the APOBEC3C enzymes were determined by loading 10 μg of purified enzyme on a 10 mL Superdex 200 (GE Healthcare) size exclusion column. The column was prepared by pouring the resin bed in a column with 16-cm height and 0.5-cm diameter. The running buffer contained 50 mM Tris pH 8.0, 200 mM NaCl and 1 mM DTT. The Bio-Rad standard set was used to generate a standard curve from which molecular masses and oligomerization states of the enzymes were determined.
A3C S188 and A3C I188 (0.5 μM) were incubated with 10 μM BS3 in 20 mM Hepes (pH 7.5), 150 mM NaCl and 1 mM DTT for 1 hour at 21°C. Crosslinked proteins were resolved on a 12% SDS-PAGE gel, transferred to a nitrocellulose membrane for Western Blot analysis and visualized using primary antibody for native APOBEC3C (GeneTex) and secondary IRdye labeled goat anti-rabbit antibody compatible with the LI-COR/Odyssey system.
All ssDNA substrates were obtained from Tri-Link Biotechnologies as previously published [44]. Reactions were carried out under single-hit conditions (i.e. <15% substrate usage) to ensure that a single enzyme carried out the deaminations on the ssDNA[60]. A ssDNA substrate containing two 5′-TTC motifs (100 nM) was incubated with 350 nM of APOBEC3C I188 or 700 nM of APOBEC3C S188 for 5 to 30 min at 37°C in RT buffer (50 mM Tris, pH 7.5, 40 mM KCl, 10 mM MgCl2, and 1 mM DTT). The reaction time was varied on each ssDNA according to the specific activity of the enzymes to ensure <15% substrate usage. Reactions were started by the addition of the ssDNA substrate. APOBEC3C-catalyzed deaminations were detected by treating the ssDNA with uracil DNA glycosylase (New England Biolabs) and heating under alkaline conditions before resolving the fluorescein-labeled ssDNA on 10 or 20% (v/v) denaturing polyacrylamide gels, depending on the sizes of the ssDNA fragments. Gel photos were obtained using a Typhoon Trio multipurpose scanner (GE Healthcare), and integrated gel band intensities were analyzed using ImageQuant (GE Healthcare). The specific activity was calculated from single-hit condition reactions by determining the picomoles of substrate used per minute for a microgram of enzyme.
Mutagenesis of ssDNA by A3 enzymes during reverse transcription of an RNA template was assessed using an in vitro assay, which models reverse transcription of an RNA template and second-strand synthesis. The method is described in detail in Feng and Chelico 2011 [33]. This system uses an in vitro synthesized RNA, which contains a polypurine tract (PPT), a protease gene (prot) of HIV, and a lacZα reporter for blue/white screening. The RNA is reverse transcribed to (−)DNA by reverse transcriptase (RT) by annealing a DNA primer and after the RNaseH domain of RT removes the RNA, the PPT enables second-strand (+)DNA synthesis by acting as a primer. A 368-nt RNA template (50 nM) is annealed to a DNA primer (24-nt) and incubated with 1.5 μM of nucleocapsid (NC), 1.2 μM of reverse transcriptase (RT) and 500 μM of dNTPs in RT buffer in the presence or absence of 350 nM of each APOBEC3C enzyme. The RNA template contained an HIV-1 PPT, nucleotides (nt 2282–2401) from the HIV-1 clone 93th253.3 (accession number U51189), and lacZα. The resulting dsDNA that is synthesized from this in vitro system was PCR amplified using Pfu Cx Turbo Hotstart (Agilent Technologies) that can use uracils as a template with high fidelity. These amplicons were then cloned into a pET-Blue vector backbone that allows for blue-white screening of the synthesized lacZα. At least twenty-five mutated clones for each condition were tested.
1000 Genomes Project data was mined for the presence of SNPs at position 188 of APOBEC3C (SNP ID rs112120857). To further elucidate the frequency of the APOBEC3C I188 SNP across Africa, we analyzed the genomes reported by Schuster et al.[38] and Lachance et al.[37] for the presence of the I188 allele. To assay for the presence of SNP at position 188 in other hominoids, we mined the Great Ape Genome Project [42] (accession number SRP018689) sequences in the NCBI short read archive.
|
10.1371/journal.pgen.1006862 | Distinguishing functional polymorphism from random variation in the sequences of >10,000 HLA-A, -B and -C alleles | HLA class I glycoproteins contain the functional sites that bind peptide antigens and engage lymphocyte receptors. Recently, clinical application of sequence-based HLA typing has uncovered an unprecedented number of novel HLA class I alleles. Here we define the nature and extent of the variation in 3,489 HLA-A, 4,356 HLA-B and 3,111 HLA-C alleles. This analysis required development of suites of methods, having general applicability, for comparing and analyzing large numbers of homologous sequences. At least three amino-acid substitutions are present at every position in the polymorphic α1 and α2 domains of HLA-A, -B and -C. A minority of positions have an incidence >1% for the ‘second’ most frequent nucleotide, comprising 70 positions in HLA-A, 85 in HLA-B and 54 in HLA-C. The majority of these positions have three or four alternative nucleotides. These positions were subject to positive selection and correspond to binding sites for peptides and receptors. Most alleles of HLA class I (>80%) are very rare, often identified in one person or family, and they differ by point mutation from older, more common alleles. These alleles with single nucleotide polymorphisms reflect the germ-line mutation rate. Their frequency predicts the human population harbors 8–9 million HLA class I variants. The common alleles of human populations comprise 42 core alleles, which represent all selected polymorphism, and recombinants that have assorted this polymorphism.
| The HLA complex is a region of the human genome containing immune system genes. Our study concerns those HLA genes that orchestrate defense against viral infections. Distinguishing HLA genes from other human genes is their extensive variation within individuals, families and populations. One advantage of this genetic variation is to increase the depth and breadth of the weaponry used against viruses; another is to impede the spread of infection within families and communities. A drawback to HLA variation is that bone-marrow transplants between donors and patients of different HLA type trigger immune reactions that attack and can kill the patient. For some patients an HLA identical family member can be the donor, but for others an unrelated HLA identical donor is sought. Facilitating these searches are registries, listing millions of possible donors whose HLA types were determined by gene sequencing. During the last ten years, this effort produced exponential growth in the number of HLA variants sequenced. This gave us the unprecedented opportunity to compare more than 10,000 sequences and distinguish aspects of the variation that are important for immune functions, from those that are not. First, however, we needed to develop software that could handle this mass of data.
| Present in all jawed vertebrates, the Major Histocompatibility Complex (MHC) is a genomic region that encodes fundamental components of the immune system. Hallmarks of the MHC are highly polymorphic genes that encode diverse MHC class I and II antigen-presenting molecules [1, 2]. The human MHC is called the HLA region and is present on the short arm of chromosome 6 [3]. HLA class I and II glycoproteins have homologous structures and complementary functions in binding peptide antigens and presenting them to lymphocyte receptors [4, 5]. HLA class II is dedicated to adaptive immunity and engagement of the αβ antigen receptors of CD4 T cells [6]. In contrast, HLA class I contributes both to innate immunity, by engaging Natural Killer (NK) cell receptors, and to adaptive immunity, through engagement of the αβ antigen receptors of CD8 T cells [7]. Correlating with these functional differences, polymorphism within the antigen-binding site is restricted to one of the two domains that form the site for HLA class II whereas HLA class I polymorphism is spread throughout the two domains [8, 9]. Consequently, the number of alleles and the differences between them are greater for HLA class I, the subject of our investigation, than HLA class II [10].
Within the HLA region, three genes, HLA-A, HLA-B and HLA-C, encode highly polymorphic HLA class I molecules. Sequence variation is concentrated in the α1 and α2 domains that are encoded by exon 2 and 3, respectively. These two domains contain the binding sites for peptide antigens and lymphocyte receptors [11]. The functional effects of the polymorphism are first to increase the breadth of an individual’s immune response to a pathogen, and second to diversify that response within families and populations. One clinical corollary of HLA polymorphism is that numerous diseases are associated with particular HLA alleles and haplotypes, and are frequently the strongest genetic associations [7, 12]. Another clinical corollary is that the success of allogeneic transplantation of tissues and organs improves with the extent of HLA match between donor and recipient [13].
HLA class I typing for clinical transplantation was begun in the 1960s using low-resolution serological methods. Nucleotide sequencing of HLA class I alleles began in the 1980s and by 1988 had led to establishment of the HLA database as the source for accurate, curated HLA sequence data [10, 14–16]. Since that time, improvements in methods [17] have progressively increased the discovery rate of novel alleles. By July 2016 sequences for more than 10,000 HLA-A, -B and -C alleles were deposited in the database. These alleles represent a worldwide sampling of many, but not all, human populations. They provide a unique data set for analysis of HLA class I variation. To analyze this variation, we developed new and general methods for handling and analyzing these large numbers of homologous sequences. Using these tools we examined variation in exons 2 and 3 of HLA-A, -B, and –C, which encode α1 and α2, with the goal of identifying those aspects of HLA class I variation that have most impact on the diversity of human immune function. The methods used here to study exons 2 and 3 of HLA class I are directly applicable to polymorphic HLA class II genes. They can also be applied to other regions of HLA genes, which are known to harbor functionally relevant polymorphism [18–20], when sufficient sequence data become available.
The α1 and α2 domains of HLA class I glycoproteins contain the functional sites that bind peptide antigens and engage lymphocyte receptors. These domains are also the site for the extraordinary polymorphism of HLA class I. Clinical sequence-based typing of HLA-A, -B and -C, targets exons 2 and 3 that encode α1 and α2, respectively. Such typing, of millions of prospective transplant donors, facilitated this analysis of sequence variation in 3,489 HLA-A, 4,356 HLA-B and 3,111 HLA-C alleles (S1 Fig).
A general method of multi-sequence dot-plot analysis was developed (see Materials and methods) and used to compare the exon 2 and 3 sequences of HLA-A, HLA-B and HLA-C individually (Fig 1A–1C), and in combination (Fig 1D). The mean intragenic distances of the three genes differ significantly (p<1 x 10−10, One-Way ANOVA), with HLA-C showing the shortest average distance of 16.60 nucleotide differences (3%) compared to HLA-B, which has the largest with a mean 27.65 differences between alleles (5%). HLA-A is intermediate with 22.82 differences between alleles (4%). The average number of differences between alleles of the same gene is 23.75, whereas the average between alleles of different genes is significantly higher at 51.12 (p<1 x 10−10, One-Way ANOVA).
The HLA-A and HLA-B dot plots show well-defined triangular clusters of closely related alleles (Fig 1A and 1B). These clusters correspond to the HLA-A and HLA-B antigens defined by serological typing, the method first used to define HLA class I polymorphisms [21]. Most pairwise differences are greater than 20 nucleotides, producing an extensive white background on which there are well-defined triangles of color. The dot-plot comparison of HLA-C alleles also has well-defined clusters corresponding to serological HLA-C types (Fig 1C). However, in contrast to the HLA-A and HLA-B dot plots, white areas do not dominate because HLA-C alleles have diverged to lesser extent than HLA-A or HLA-B alleles. One likely cause of this difference is that MHC-A and MHC-B are ten million years older than MHC-C, another is that HLA-C has distinctive functions in reproduction, which are not shared with HLA-A or -B. In particular, HLA-C expressed on fetal trophoblast interacts with KIR on maternal uterine NK cells to facilitate placental development [22]. Fig 1D, shows all pairwise comparisons of HLA-A, –B and -C alleles. The color patterns show how HLA-B and HLA-C are more closely related to each other than either is to HLA-A. The median number of differences between sequences of HLA-B and HLA-C is 42 compared to 55–56 for differences between HLA-A and HLA-B or HLA-C (S9 Fig). These results are consistent with MHC-C having originated with duplication of an MHC-B allele.
Each of the 546 positions in exons 2 and 3 can have five alternative forms, the four different nucleotides and insertion/deletion (indel). The distribution of the variability is shown as histograms in S2 Fig and the numbers per exon for each gene are given in S3 Fig. In summing the data for the three genes, we find only 4.5% of the positions are invariant, whereas 23.2%, 34.3% and 32.2% positions have two, three and four forms, in HLA-A, –B and –C, respectively. All five forms are present at 5.7% of positions. The pattern of variability is similar for HLA-A, -B and -C (S2 Fig). Variation was thus found at almost every position in exons 2 and 3 of these genes.
We performed similar analysis of the amino-acid sequences of the α1 and α2 domains. The results are displayed as histograms in S4 Fig and summarized in Table 1. The striking result is that, for each of the three genes, there are no positions in the sequences of their protein products that exhibit only one or two amino acids. The number of residues at a given position varies from 3 to 14, with 149 of the 181 positions having between 5 and 9 alternative amino acid residues (Table 1).
To distinguish positions having a balanced polymorphism between two or more nucleotides, from positions dominated by one nucleotide, we determined the incidence (in the dataset of allelic sequences) for the second-most common nucleotide at each position in the exon 2 and 3 sequence (Fig 2). Positions where the incidence of the second nucleotide exceeded 1% were considered polymorphic, whereas positions with lower incidence were considered to exhibit rare variation. The second nucleotide occurs in more than 1% of the alleles for 70 positions in HLA-A, 85 in HLA-B and 54 in HLA-C (S5 Fig). These comprise a minority of positions in the 546 bp sequence of exon 2 and 3, demonstrating that the variation observed at most positions in exons 2 and 3 (S2 Fig) is due to the contribution of nucleotide substitutions that are present in one or a few alleles.
Analyzing the incidence of the second most common amino acid residue showed that all 181 positions in the α1 and α2 domains of HLA-A, -B and -C exhibit some variation. Of these positions, however, only 45 in HLA-A, 46 in HLA-B and 32 in HLA-C have a second amino acid incidence of >1% and are thus considered polymorphic (Fig 3, S6 Fig). Twelve of these positions are shared by HLA-A, -B and -C: four in α1 (residues 9, 66, 77 and 80) and eight in α2 (95, 97, 99, 114, 116, 152, 156 and 163). Larger numbers of polymorphic positions are shared by two of the three HLA class I: 26 by HLA-A and -B, 20 by HLA-B and -C, and 14 by HLA-A and -C. On the other hand, 17 polymorphic positions are unique to HLA-A, 12 to HLA-B and 10 to HLA-C. These 39 positions impart considerable gene-specific character to the polymorphism (Fig 4). This reflects functional specialization of the three HLA class I.
For polymorphic positions with a second nucleotide incidence of >1%, the mean number of different nucleotides is 3.8 for HLA-A, 3.7 for HLA-B and 3.6 for HLA-C. The values are higher than the mean differences for all other variable positions: 3.1 for HLA-A and HLA-B and 2.9 for HLA-C. The polymorphic positions have a significantly increased incidence of three or more nucleotides at each position (91%) when compared to the other positions in the dataset (73%) (Chi squared test, p = 2.08 x 10−6). Additionally there are polymorphic positions with three or more nucleotides with an incidence of >1%. There are nine positions in HLA-A, 14 in HLA-B and nine in HLA-C having three nucleotides with an incidence >1%. With four nucleotides at an incidence >1% are position 527 (codon 152) in HLA-A, positions 206 (codon 45), 272 (codon 67) and 362 (codon 97) in HLA-B, and position 368 (codon 99) in HLA-C. These results suggest that variation arising at these sites is more likely to be retained in the population. This is consistent with the sequence variation at such sites serving to diversify the functional interactions of HLA class I with peptide antigens and lymphocyte receptors.
Crystallographic analyses have identified 70 residues in α1 and α2 domains of HLA class I that are involved in binding peptide antigens and/or lymphocyte receptors [11, 23–27]. These functionally defined residues overlap considerably with the set of polymorphic residues defined by the incidence of the second nucleotide. Thus, 35 of 45 polymorphic HLA-A positions, 32 of 46 polymorphic HLA-B positions and 19 of 33 polymorphic HLA-C positions are functionally important sites. This correlation of function with polymorphism is highly significant for HLA-A (p = 6.52 x 10−7) and HLA-B (p = 1.18 x 10−6), but less so for HLA-C (p = 0.0124) (2x2 Fisher’s Exact test). The difference is consistent with highly polymorphic HLA-A and -B molecules interacting mainly with highly diverse αβ CD8 T cell receptors, and less polymorphic HLA-C molecules interacting mainly with the less diverse killer cell immunoglobulin–like receptors (KIR) of NK cells.
The striking correlation between immunological function and genetic polymorphism was further investigated by testing the polymorphic sites for evidence of positive selection. Our null hypothesis was that polymorphic sites are not subject to positive selection. If correct there would be no bias in the rates of synonymous and non-synonymous nucleotide substitutions, as measured by the parameters dS and dN. For each test performed, the probability for rejecting the null hypothesis of neutral variation (dN = dS) is shown in Table 2. Values of P<0.05, following a Bonferroni correction and bootstrapping of 1,000 replicates, were considered significant at the 5% level and are highlighted.
We first compared the 70 codons encoding functionally critical α1 and α2 domain residues (Binding site codons in Table 2), as defined previously [11], to the other 112 codons of exons 2 and 3. For the 70 functional positions, the dN-dS values all point in the direction of positive selection (3.58 for HLA-A, 2.89 for HLA-B and 2.58 for HLA-C) and are statistically significant for HLA-A (p = 0.0031) and HLA-B (p = 0.0275) but not for HLA-C (p = 0.0720) (statistical significance is achieved at p<0.05, after application of Bonferroni correction to the tests on a per gene basis). In contrast, the 112 other positions (Not binding site codons) have negative dN-dS values consistent with the null hypothesis: -1.78 for HLA-A (p = 1.0), -1.73 for HLA-B (p = 1.0) and -1.25 for HLA-C (p = 1.0). These results argue strongly against positive selection at the other positions.
Having validated the selection analysis on functional sites, we compared the polymorphic codons, as defined by having at least one nucleotide position where the incidence of the second nucleotide >1%, with the remaining codons of exons 2 and 3. For the polymorphic codons the dN-dS values pointed clearly in the direction of positive selection and were statistically significant: 4.98 for HLA-A (p = 0.0001), 4.55 for HLA-B (p = 0.0001) but not for HLA-C (2.20, p = 0.1800). In contrast, the values for the codons where the second nucleotide was present at less than 1% were all decidedly negative: -2.66 for HLA-A (p = 1.0), -3.06 for HLA-B (p = 1.0) and -2.78 for HLA-C (p = 1.0). These data strongly support positive selection at the polymorphic positions.
Independent analysis of the α1 and α2 domains (Table 2) shows that dN-dS for HLA-A is higher in α2 for both binding sites and polymorphic positions (3.342, p = 0.0067; 4.517 p = 0.0008) than α1 where selection is detected only for polymorphic positions (1.359 p = 1.0000; 3.135 p = 0.0130) which represent a subset of the functionally important residues. For HLA-B selection was detected for the polymorphic positions in both α1 (3.467, p = 0.0044) and α2 (2.875, p = 0.0286) and for complete set of binding site codons (2.889, p = 0.0275) but not the individual domains. The HLA-C sequences show no significant selection differences between the α1 and α2 domains, with neither the functional nor polymorphic positions showing significant positive selection.
Assessment of selection at gene-specific positions of polymorphism (Fig 5) showed there has been positive selection only for HLA-C specific polymorphisms and those are limited to one of the two domains. The α1 domain has been subject to strong positive selection (dN-dS = 3.65, p = 0.0023), but that is not the case for HLA-C specific sites of α2 (dN-dS = -0.16, p = 1.00). The gene-specific sites of HLA-A and HLA-B show no evidence for significant positive selection.
Previous analysis of HLA class I variation, studied small numbers of alleles and relied on visual inspection to discern the relationships between them [5]. To analyze the current dataset of 10,956 HLA class I sequences, we developed the Sq2 algorithm (see Materials and methods), which provides a quicker, more objective and largely automated approach. In two separate phases of analysis, Sq2 divided the alleles into three categories.
In the first phase, Sq2 identified all SNP alleles, which constitute ~85% of the dataset. These are alleles of more recent origin that differ from an older allele by just one nucleotide substitution. After identifying and removing the SNP alleles, the reduced database of 1,555 alleles was subjected to the second phase of analysis. This identified all alleles that are recombinants of other alleles. To do this, Sq2 identified motifs of several substitutions that are present in multiple allelic backgrounds as a consequence of recombination (Fig 5). The iconic example is the Bw4 motif. Present in codons 76–83 of one third of HLA-B alleles, Bw4 defines the ligand recognized by a major NK cell receptor, KIR3DL1 [28, 29]. As well as being present in 12 of the 33 HLA-B allele families, Bw4 was transferred by a gene conversion from HLA-B to HLA-A, where it spread by recombination to four HLA-A allele families [30]. By comparing the distribution of such motifs among alleles, Sq2 identified pairs of alleles differing only by presence or absence of a particular motif.
In this way 1,171 recombinants were identified. Of these 1,092 were formed by recombination between alleles of the same gene (intragenic recombinants), and 79 were recombinants formed by recombination between alleles of different genes (intergenic recombinants). Of the latter, 16 are products of single recombination (crossover) and 63 (10 HLA-A, 37 HLA-B, and 16 HLA-C) are products of double recombination (conversion). HLA-B is clearly seen as the more frequent beneficiary of recombination (Table 3).
Among intragenic recombinants, double recombinants (N = 735) outnumber single recombinants (N = 357) by a factor of two. It is likely that some alleles assigned as single recombinants are actually double recombinants, for which the second recombination is not in exon 2 or 3 but in a flanking intron, for which we had no sequence. Both forms of recombinant are more prevalent at HLA-B (N = 728) than either HLA-A (N = 226) or HLA-C (N = 138). The frequency of double recombination for HLA-B is similar in exons 2 and 3, whereas it is greater in exon 2 of HLA-A and in exon 3 of HLA-C. A similar hierarchy is observed for the single recombinants.
Removal of SNP and recombinant alleles, reduced the database to <1% of its original size. This left 11 HLA-A, 17 HLA-B and 14 HLA-C alleles (Fig 6A). Because these 42 alleles represent all functionally significant variation (polymorphism) in exons 2 and 3 of HLA-A, -B and -C, we call them ‘core’ alleles (Fig 6B). Although they are older in their origins than the SNP alleles and recombinant alleles, they are unlikely to represent, or reflect, any particular human population, either ancient or modern.
Core alleles vary widely in their contribution to the total set of alleles (Fig 6A), in their geographical distribution (S7 Fig) and in their abundance in the modern human population. A substantial proportion of the core alleles, 5 HLA-A, 8 HLA-B and 6 HLA-C, are likely derived from archaic humans (Fig 6A) [31]. A dot plot analysis of the core alleles (S8 Fig) has similar substructure to that of the complete set of alleles (Fig 1D) and for each gene the mean pairwise differences for core alleles and all alleles is remarkably similar (S9 Fig). Analysis of selection on the polymorphic and functional sites of core HLA-A, -B and -C alleles (Table 4) gives comparable results to those obtained for the full sets of alleles (Table 2) for HLA-A. For HLA-B and -C the results are comparable when looking at the full-length sequence, but some differences are seen for the individual domains. This could, however, be due to the small number of sequences analyzed.
The effects of applying the Sq2 algorithm to the HLA class I data set are seen in histograms constructed from the pairwise differences of nucleotide sequences (Fig 7, top row). For complete sets of HLA-A, -B and -C alleles, the histograms have a characteristic bimodal distribution with one peak at 2 nucleotide differences and a second peak at 20–30 nucleotide differences. The first peak contains the large number of pairwise comparisons between alleles differing by one or two nucleotide substitutions. Pairs differing by one nucleotide substitution usually involve an older, common allele and a rare SNP variant. Pairs differing by two nucleotide substitutions involve two rare SNP variants that differ from the same parental allele by different SNPs.
Taking the SNP alleles out of the analysis, led to loss of the first peak and retention of the second peak (Fig 7, middle row). For HLA-A and -B the loss is complete, but for HLA-C it is not. HLA-A gives a bimodal distribution, which differs from that observed in the complete dataset. This is because HLA-A comprises a small number of large and divergent allele families. Thus the minor distribution, seen as the shoulder at 4–12 nucleotide differences, comprises the differences between members of the same allele family, whereas the major distribution is formed from the larger differences between members of different allele families. In contrast to HLA-A, HLA-B comprises a large number of less divergent allele families than HLA-A, as well as a few highly divergent alleles with no close family ties. This gives HLA-B both a more symmetrical and broader distribution.
Histograms for the pairwise differences between core alleles (Fig 7, bottom row) represent much of the range of difference seen with the larger data sets, with the notable absence of allele pairs differing by small numbers of substitutions. That the HLA-C core allele histogram has a distribution with a more coherent shape, than the HLA-A and -B core histograms, probably reflects the more recent origin of HLA-C [4, 32].
Because we have detected variation at all nucleotide positions in exons 2 and 3 of HLA-A, -B and -C (S2 Fig) the maximum number of possible HLA class I alleles is 5546 (4.3 x 10381). This calculation is based on observing all four nucleotides or an indel at each of the 546 positions in the exon 2 and 3 sequence. This number far exceeds the size of the modern human population, which is estimated to be 7.5 billion (http://www.worldometers.info/world-population/). This difference means that the number of variants present in a population is limited only by the size of that population.
To estimate the total number of HLA-A, -B and -C alleles now present in the human population, we first determined the rate at which novel alleles are being identified. In this context, the rate is simply the ratio between the number of individuals typed and the number of new alleles discovered. For each gene, the product of the rate and the population size (7.5 billion) gives an estimate of the total number of alleles. To provide an internally consistent dataset, we analyzed HLA typing data from donor cohorts recruited by various transplantation registries, but all typed at the same sequencing center (Histogenetics).
Similar rates, of 1.80, 2.13, and 2.18 x 10−4, were observed for the acquisition of novel HLA-A, -B and –C alleles, respectively (Table 5). Using these rates, we estimate there are 2.7 million HLA-A, 3.3 million HLA-B and 3.2 million HLA-C alleles in today’s human population. These estimates are comparable to the 3.5 million alleles per HLA gene predicted by Klitz, et al [33], using estimates of effective population size and mutation rates.
Our method for estimating the total numbers of HLA-A, -B and -C alleles used a constant rate for the discovery of novel alleles. This assumption was based on the results of two recently published studies [34, 35], which both indicated that the rate of discovery of new alleles is not tapering off over time, even for European populations [34, 35], which have been intensively studied compared to the populations of other continents.
Analyses show that the human population has a small number of common HLA class I alleles (68 HLA-A, 125 HLA-B, 44 HLA-C) that are present at appreciable frequency in different populations [36]. In contrast, the overwhelming majority of HLA class I alleles are very rare and highly localized in their distribution. Consistent with these properties, each newly sampled cohort or population is expected to harbor a subset of HLA-A, -B and -C alleles that are novel and present in only one or a few individuals. Because of their rarity and population specificity, the relative frequency of novel alleles will not diminish in time as further cohorts of donor are HLA typed at high resolution.
We studied sequence variation in exons 2 and 3 that encode the highly polymorphic α1 and α2 domains of HLA-A, -B and -C. The analysis was restricted to these exons and genes to enable an in depth study of the maximum number of sequences. The tools developed for this analysis can, and should, be extended to study the remaining exons of these genes, which are known to contain functionally relevant polymorphism [18–20], when sufficient data becomes available. These analyses can also be applied to the study of polymorphism in the HLA class II genes.
Sequence differences in the α1 and α2 domains of HLA-A, -B and -C determine the peptide antigens that are bound by an HLA class I allotype, as well as the lymphocyte receptors that can engage the complex of peptide and HLA class I. HLA-A, -B and -C are candidates for being the most polymorphic of human genes [22]. Moreover, their polymorphisms are associated with numerous clinical factors including infectious diseases, autoimmune and inflammatory diseases, pregnancy syndromes and success in the transplantation of allogeneic organs and tissues [7, 37–42].
Transplantation of bone marrow, and other sources of hematopoietic stem cells, is a successful and widely used therapy for leukemias, lymphomas and other malignancies of hematopoietic cells. The preferred donor is an HLA identical sibling, but in the absence of such a donor, the next best choice is an unrelated individual having the same, or very similar, HLA type as the patient. To identify such donors, there exists an international network of donor registries, which has HLA typed more than 30 million potential HCT donors [43]. During the last ten years, less precise methods of HLA typing have been superseded by nucleotide sequencing exons 2 and 3 of the HLA-A, –B and –C genes.
The set of HLA class I sequences we studied derive from sequence-based typing of >3 million individuals, as well as earlier studies in which typing at lower levels of resolution identified variants, which were followed up with targeted sequence analysis. The prospective donors of hematopoietic stem cells were recruited to registries in varied countries and continents, but demographically and anthropologically they are not, in the main, well characterized. A total of 10,956 different exon 2 and 3 sequences were analyzed: 3,489 HLA-A, 4,356 HLA-B and 3,111 HLA-C alleles. In our analysis of these three sets of alleles, each sequence was given equal weight, irrespective of its abundance or scarcity in any human population.
At the nucleotide level, we found substitutions at >95% of all positions in each of the three genes. As the exceptions are at different positions in each gene, we predict that substitution at these positions will soon be identified. At the amino-acid level, we found substitutions at every position in the α1 and α2 domains of HLA-A, -B and –C. A majority of the substitutions, >84%, are in rare alleles, which in many cases have been detected in only one individual or one family. Most of the alleles differ from a common allele by the single substitution that defines them. The obvious interpretation of these data is that these substitutions reflect the germ line mutation rate of the HLA-A, –B and –C genes. Consistent with this thesis, there is no evidence for positive selection at these sites, many of which are, otherwise, highly conserved. The remaining alleles are formed by intragenic or, rarely, intergenic recombination events. From the rate at which new alleles in exons 2 and 3 have been defined by sequence-based typing we estimate there are 2–3 million each of HLA-A, -B and -C alleles in the human population worldwide.
The majority of the variable nucleotide positions are characterized by one dominant and one or more rare nucleotides. However, variation at a smaller number of nucleotide positions, (70, 85 and 54 in HLA-A, -B and -C, respectively) has a very different character. These positions have two, three or four nucleotides at appreciable frequency. They have also been spread by recombination throughout the population of alleles and are thus found in numerous combinations. There is good evidence for positive selection at these sites, which has over time, given them a balanced polymorphism. Supporting this conclusion, numerous immunological studies have correlated substitution at polymorphic sites with modulation of HLA-A, -B and –C function [7, 37, 40, 41, 44–46]. Thus we can divide the alleles into two distinctive groups. Firstly SNP alleles, defined by substitution that confers no functional benefit, but could be detrimental in the context of transplantation. Secondly, functional alleles, with functional benefit conferred by combinations of substitutions at positions with balanced polymorphism.
We further divided the functional alleles into two subgroups: 1,171 recombinant alleles that were derived by recombination from other alleles and 42 core alleles (11 HLA-A, 17 HLA-B and 14 HLA-C) that cannot be derived by simple events of recombination from other alleles. The core alleles, many of which were passed by introgression from archaic to modern humans [31], contain all elements of HLA-A, -B and –C polymorphism present in the modern human population. Although the core alleles are probably older than the SNP alleles and the recombinant alleles, they are very unlikely to represent the HLA-A, -B and -C alleles carried by any particular ancestral human population.
Because polymorphic MHC class I and II genes have no wild-type, understanding their genetics and biology in any species requires extensive study of populations. For reasons of cost and logistics this has been rarely, if ever, achieved. Many population studies have recruited only small numbers of individuals (therefore, likely missing rare alleles) and until recently have reliably assayed only known alleles. Because the HLA class I and II genes contribute to so many numerous and diverse aspects of human health and disease [7, 37–42], the MHC of the human species is by far the most studied and, by default, provides the model for studies of other placental mammals [4, 32, 47–49]. The capacity to acquire large datasets, of the type we have analyzed and reported here, should enable HLA population genetics and disease associations to be studied to increasingly higher definition, resolution and coverage of the world’s human populations.
The minimum requirement for naming an HLA class I allele and depositing it in the IPD-IMGT/HLA Database, is the nucleotide sequence of exons 2 and 3. Because of this requirement, a majority of deposited sequences (~65% of HLA-A and -B alleles and ~80% of HLA-C alleles) consist of only exons 2 and 3, encoding residues 2-182 of the mature HLA class I protein. Thus to maximize the number of alleles analyzed we limited this study to the sequences of exons 2 and 3. Our analysis used all sequences in the IPD-IMGT/HLA database as of July 2016 (Release 3.25.0).
All analyses used custom written Perl scripts, http://www.perl.org [50], with graphical outputs generated using the Perl::GD modules or R, http://www.r-project.org [51]. Where appropriate, statistical analysis was also completed in R. For F distribution analyses with df1 and df2 exceeding 1,000, R outputs a p-value of 0, these have been reported as p<1 x 10−10. The set of scripts developed constitutes the Sq2 package. Individual scripts perform different steps of the algorithm. The individual algorithms are listed and described below:
The scripts are available from the ANHIG Gitlab repository which can be found at; https://github.com/ANHIG.
Alleles of an HLA class I gene are of three types: core alleles, recombinant alleles and SNP alleles. Core alleles comprise the set of alleles that cannot be related to each other by single events of recombination or point mutation. Recombinant alleles are the products of one or more recombination events between core alleles. SNP alleles differ from another allele by a single nucleotide polymorphism (SNP). The Sq2 algorithm was developed to assign HLA class I alleles to these three categories. In a series of iterative steps, Sq2 first defines the mutant alleles, then the recombinant alleles and lastly the core alleles.
The MEGA software package, version 5.1, [61–63] was used to assess positive selection in HLA-A -B, and -C sequences using the codon-based Z-test of selection. The analysis used the Kumar method [64], with variance of the difference being computed by the bootstrap method (1,000 replicates, to allow for multiple testing). For each gene dN and dS analysis examined the exon 2 and 3 sequences, both separately and combined. Further to this, analysis was performed on specific codons. In these experiments, the nucleotide positions of interest were extracted from each allele and an artificial sequence created. All allele sequences were then compiled into a single data file that could be run through the dN-dS analysis. Each of these data sets was analysed for the positions across the combined exon 2 and 3 sequence, as well as for the positions within the individual exons. In all cases, the probability of rejecting the null hypothesis of strict-neutrality (dN = dS) in favor of the alternative hypothesis (dN>dS) is shown in Tables 2 and 4. P-values were subject to Bonferroni correction for multiple comparison. P-values of less than 0.05 are considered significant at the 5% level and are highlighted.
Full-length coding sequences of the core alleles were obtained from the IPD-IMGT/HLA Database. Sequences were aligned in Geneious 7 [65] using the MAFFT algorithm [66]. The alignment was input into MEGA 7 and the tree was constructed using the Neighbor Joining method with pairwise deletion, the Tamura-Nei model, and 1000 bootstrap replicates. It is displayed as an unrooted tree in Fig 6B [67, 68] and bootstrap values of >50 are shown on the tree.
For the core alleles, maps of their frequency distribution in human populations were generated using ArcGis 10 [69]. Population allele frequencies and location coordinates were downloaded from allelefrequencies.net [70]. Only anthropologically well-characterized populations of >50 individuals were included. Specifically excluded were admixed populations, populations of recent migrants, bone marrow registry populations and the subjects of disease association studies. Populations with low resolution HLA class I typing, (less than two field, four digit resolution) were not included in the final dataset.
|
10.1371/journal.ppat.1002912 | Molecular Basis for Nucleotide Conservation at the Ends of the Dengue Virus Genome | The dengue virus (DV) is an important human pathogen from the Flavivirus genus, whose genome- and antigenome RNAs start with the strictly conserved sequence pppAG. The RNA-dependent RNA polymerase (RdRp), a product of the NS5 gene, initiates RNA synthesis de novo, i.e., without the use of a pre-existing primer. Very little is known about the mechanism of this de novo initiation and how conservation of the starting adenosine is achieved. The polymerase domain NS5PolDV of NS5, upon initiation on viral RNA templates, synthesizes mainly dinucleotide primers that are then elongated in a processive manner. We show here that NS5PolDV contains a specific priming site for adenosine 5′-triphosphate as the first transcribed nucleotide. Remarkably, in the absence of any RNA template the enzyme is able to selectively synthesize the dinucleotide pppAG when Mn2+ is present as catalytic ion. The T794 to A799 priming loop is essential for initiation and provides at least part of the ATP-specific priming site. The H798 loop residue is of central importance for the ATP-specific initiation step. In addition to ATP selection, NS5PolDV ensures the conservation of the 5′-adenosine by strongly discriminating against viral templates containing an erroneous 3′-end nucleotide in the presence of Mg2+. In the presence of Mn2+, NS5PolDV is remarkably able to generate and elongate the correct pppAG primer on these erroneous templates. This can be regarded as a genomic/antigenomic RNA end repair mechanism. These conservational mechanisms, mediated by the polymerase alone, may extend to other RNA virus families having RdRps initiating RNA synthesis de novo.
| The 5′- and 3′-ends of RNA virus genomes have evolved towards efficient replication, translation, and escape from defense mechanisms of the host cell. Little is known about how RNA viruses conserve or restore the correct ends of their genomes. The Flavivirus genus of positive-strand RNA viruses contains important human pathogens such as yellow fever virus, West Nile virus, Japanese encephalitis virus and dengue virus (DV). The Flavivirus genome ends are strictly conserved as 5′-AG…CU-3′. We demonstrate here the primary role of the DV polymerase in the conservation of the first and last genomic residue. We show that DV polymerase contains an ATP-specific priming site, which imposes a strong preference for the de novo synthesis of a dinucleotide primer starting with an ATP. Furthermore, the polymerase is able to indirectly correct erroneous sequences by producing the correct primer in the absence of template and on templates containing incorrect nucleotides at the 3′-end. The correct primer is productively elongated on either correct or incorrect templates. Our findings provide a direct demonstration of the implication of a viral RNA polymerase in the conservation and repair of genome ends. Other polymerases from other RNA virus families are likely to employ similar mechanisms.
| Most RNA viruses maintain the specific sequences present at the ends of their genomes. The 5′ genome end may carry a cap structure to ensure both genome stability and efficient translation [1]. The 3′-end may carry a poly(A) tail or adopt specific 3′-end sequences required for viral replication [2], [3]. They are generally copied exactly to avoid loss of genetic information, and have supposedly evolved towards optimum replication efficiency. Terminal genome damage can be caused by errors introduced by the viral polymerase during initiation and termination, or by cellular ribonucleases [4]. In addition to special mechanisms to ensure efficient initiation of RNA synthesis, viruses have evolved mechanisms to repair or correct damaged extremities such as the use of abortive transcripts as primers, the generation and use of non-templated primers, and the addition of one or few non-templated nucleotides to the 3′-end by a terminal transferase activity [4]. However, our knowledge about these mechanisms is still very limited. Many RNA virus polymerases, which do not use a primer and thus initiate RNA synthesis de novo, generate abortive transcripts during the initiation phase of RNA synthesis [5], [6], [7]. Primer-mediated repair of template extremities was so far only demonstrated for the positive-strand RNA (+RNA) turnip crinkle virus (TCV) [8]. Non-templated primer synthesis by the viral polymerase might be involved in the repair mechanism of TCV [9]. Such mechanism was also proposed as the molecular basis of the reconstitution of 5′-ends of negative-strand RNA (-RNA) respiratory syncytial virus (RSV) replicons [10]. In this study we demonstrate how the dengue virus (DV) RNA-dependent RNA polymerase (RdRp), which starts RNA synthesis de novo, plays a decisive role in the nucleotide conservation of viral RNA ends.
DV belongs to the Flavivirus genus within the +RNA virus family of Flaviviridae together with viruses of the genera Hepacivirus and Pestivirus [11]. The Flavivirus genus comprises around 50 virus species [12] including major human pathogens such as DV, yellow fever virus (YFV), West Nile virus (WNV) and Japanese encephalitis virus (JEV). Flaviviruses harbour the RdRp activity in the C-terminal domain (amino acids 272–900) of non-structural protein NS5 [13], [14], [15], [16], [17]. The N-terminal domain contains methyltransferase activities involved in RNA capping [18], [19]. Evidence has been presented that the N-terminal domain of NS5 also harbours the central RNA capping guanylyltransferase activity [20]. The structure of full-length NS5 is not known but several structures of methyltransferase domains have been determined (for review see [21]). Likewise, crystal structures of Flavivirus NS5 RdRp domains have been determined for DV [16] and WNV [22]. All structurally characterized viral RdRps so far adopt the basic fold of the SCOP superfamily of DNA/RNA polymerases. As the other subgroups of this superfamily, DNA-dependent DNA polymerases (DdDp, prototype Klenow fragment of the E.coli DdDp I), RNA-dependent DNA polymerase (prototype HIV reverse transcriptase) and DNA-dependent RNA polymerases (DdRp, prototype bacteriophage T7 DdRp), their apo-structure is usually likened to a right hand comprising fingers, palm and thumb subdomains. Viral RdRps contain an encircled active site having connecting elements between the fingers and thumb subdomains. Active sites of viral RdRps performing de novo RNA synthesis are additionally closed in their initiation conformation due to the existence of structural elements allowing the stable positioning of the first NTP into a priming site [23], [24]. All Flaviviridae RdRps studied so far initiate RNA synthesis de novo. Accordingly, Flavivirus RdRp domain structures contain a “priming loop” in the thumb subdomain closing the catalytic site [16], [22]. The putative priming loop of DV RdRp was defined as comprising residues 792 to 804. Of particular interest are two aromatic residues near the tip of the loop, W795 and H798, which are conserved in all Flavivirus RdRps. They might play the role of an initiation platform to which the base of the priming NTP stacks as it was shown for bacteriophage φ6 [23] and proposed for HCV and BVDV RdRps [25], [26]. Structures of DV RdRp in complex with 3′dGTP as well as two models of de novo initiation complexes of DV and WNV RdRps favor Trp795 in the role of the initiation platform [16], [22].
Genomes of Flaviviridae lack a poly(A) tail at the 3′-end. A remarkable trait of Flavivirus genomes is the strict conservation of the 5′- and 3′-end dinucleotides as 5′ AG…CU 3′. The molecular basis for this strict conservation of the 5′- and 3′-end dinucleotides and/or the use of the same starting nucleotide for +RNA and -RNA strand synthesis by the viral polymerases is not known. Its Hepacivirus and Pestivirus counterparts have to display higher nucleotide tolerance. They are able to initiate with (A/G)C and G(G/U), respectively, since the 5′- and 3′-ends of Hepacivirus genomes of different genotypes correspond to 5′ (A/G)C…GU 3′ and the genomes of pestiviruses to 5′ GU…CC 3′. Interestingly, genomes and antigenomes of non-segmented -RNA (ns-RNA) paramyxoviruses, whose RdRps perform de novo RNA synthesis, start with a conserved 5′-AC [10].
Here we show that the strict sequence conservation of Flavivirus genome ends is entirely polymerase-encoded. We demonstrate ATP-specific de novo initiation using the RdRp domain of DV protein NS5 (NS5PolDV) and specific 10-mer oligonucleotidic RNA templates corresponding to the 3′-end of genomic +RNA and -RNA. We document the existence of a built-in ATP-specific priming site of NS5PolDV. This specific site is one of the means by which NS5PolDV ensures that the DV genome and antigenome start with an A, the others being several correction mechanisms including the generation of non-templated pppAG primers as well as the preferential formation and elongation of pppAG even on templates with non-cognate 3′-ends. Finally, we show that the ATP-specific priming site is part of the putative priming loop coming from the thumb subdomain. There, residue H798, and not W795, is essential for de novo initiation and may act as a priming platform stabilizing the ATP priming nucleotide. DV RdRp is actively involved in the conservation of the correct ends of the genome proving thus a direct example of how RNA viruses maintain the integrity of their genomes. The mechanisms described here may more broadly apply to other RNA viruses having viral RdRps able to initiate RNA synthesis de novo.
We set out to study primer synthesis by the RdRp domain of dengue virus protein NS5 (NS5PolDV) using small specific templates corresponding to the 3′-ends of the genome (+RNA) and the antigenome (-RNA). Templates are comprised of 10 nucleotides and are predicted to be devoid of stable secondary structure (see Materials and Methods). Both templates end with the dinucleotide 5′-CU-3′. Product formation over time was followed using either ATP and GTP, or all NTPs needed to form a full-length product when synthesis is precisely started at the 3′-end of the template. Figure 1 shows reaction kinetics of RNA synthesis on DV103′+ corresponding to the 3′-end of the RNA genome 5′-AACAGGUUCU-3′ (left) and on DV103′- corresponding to that of the antigenome 5′-ACUAACAACU-3′ (right). We used either [α-32P]-GTP (αGTP, panel A) or [γ-32P]-ATP (γATP, panel B) as the radioactive nucleotide. For the catalytic ion, either Mg2+ (panel A) or Mg2+ supplemented with Mn2+ (panel B) were used at their optimum concentrations 5 mM for Mg2+ and 2 mM for Mn2+ [14]. Reactions with ATP and GTP render time-dependent accumulation of a short product migrating below the marker G2 (see panel B). Comparison with authentic unlabeled pppAG (see Materials and Methods) visualized using UV-shadowing indicated that it indeed corresponds to pppAG (not shown), the expected product of the first step of de novo RNA synthesis. When DV103′+ is used as a template, pppAG is formed as well as pppAGA and pppAGAA. When all NTPs are used, pppAG accumulates with time as does pppAGA in the case of DV103′+ and pppAGU in the case of DV103′-. After the synthesis of trinucleotides NSPolDV adopts a processive RNA synthesis elongation mode to continue synthesis up to full-length products (labeled by asterisks in Figure 1). As we had observed before [14], when using Mn2+ the reaction is much more efficient and allows for the use of [γ-32P]-ATP (γATP) as radiolabeled nucleotide in order to visualize exclusively de novo RNA synthesis products starting with ATP. The pattern observed with Mg2+ is reproduced when Mn2+ is present (Figure 1B). One difference is that the use of Mn2+ results in longer full-length products, which might be caused by an alteration of the terminal nucleotide transferase activity of NS5PolDV [14], [27], [28].
In conclusion, using RNA templates mimicking viral sequences, dinucleotide and trinucleotide products are formed during initiation and before processive RNA elongation, the most abundant being the dinucleotide pppAG.
The first nucleotide of Flavivirus genomes is an adenosine, followed by a guanosine. This 5′-pppAG sequence is strictly conserved along the Flavivirus genus. In order to answer the question whether the polymerase (and/or the correct template) is at the origin of the conservation of the first nucleotide, we tested a set of DV103′- variants with different 3′-ends. In addition to the correct DV103′- CU, we used DV103′- CC, DV103′- CA and DV103′- CG in the presence of the corresponding priming NTP and GTP. The expected primer products are pppAG, pppGG, pppUG and pppCG, respectively. Figure 2A compares end points of reactions performed in the presence of αGTP and Mg2+ as the catalytic ion. Remarkably, the CU template only is proficient for product synthesis (pppAG). RNA primer synthesis on other templates is almost undetectable. We conclude that in the presence of Mg2+ as a catalytic ion the DV RdRp priming-site accommodates exclusively ATP.
To our surprise, when Mn2+ was used instead of Mg2+, the pppAG primer was generated even in the absence of the template, albeit to a lower extent (Figure 2B). This is not the case in the presence of Mg2+ even at ten-fold higher enzyme concentration (see below Figure 3B). When using Mn2+ and the DV103′- template variants, we therefore included control reactions in the absence of corresponding templates and in the presence of γGTP, which allows exclusive detection of dinucleotides starting with pppG. Figure 2C shows corresponding reaction kinetics with Mn2+ as the catalytic ion in the absence or the presence of templates using αGTP or γGTP as the radioactive nucleotide. Again, using DV103′-CU and ATP/GTP, NS5PolDV generates pppAG to a higher extent than without template. Note that no pppGA product is generated. When DV103′-CC and GTP is used, NS5PolDV synthesizes pppGG in the presence of the template only. DV103′-CA, UTP, and GTP lead to the formation of pppUG and pppGU (see γGTP control reaction), the latter by initiation internal to the template. No product is formed in the absence of the template. Finally, DV103′-CG allows formation of pppCG which is not formed in the absence of the template. In conclusion, NS5PolDV keeps the strict preference for an ATP as the priming nucleotide in the presence of Mn2+ when no template is present. Nevertheless, the use of templates with an altered 3′-nucleotide can force NS5PolDV to start the de novo RNA synthesis with the corresponding base-paired priming nucleotide, and also allows internal initiation.
Collectively, these observations confirm that the priming site of NS5PolDV has a marked specificity for ATP. This preference is strict in the presence of Mg2+. It is equally strict for dinucleotide synthesis in the presence of Mn2+ and in the absence of template. The specificity for ATP as the starting nucleotide is lost when Mn2+ is used in the presence of templates with incorrect 3′-ends; only then NS5PolDV is able to form pppNG products as efficiently as pppAG.
In the presence of Mg2+ and/or Mn2+ the built-in ATP-specific priming site drives NS5PolDV-mediated RNA synthesis starting with pppA. The dinucleotide pppAG is accumulated during RNA synthesis on templates with the correct 3′-end (see Figure 1). Using Mn2+ this pppAG primer is also formed in the absence of an RNA template. We asked the question whether NS5PolDV forms and/or elongates pppAG even on templates with incorrect 3′-nucleotides thus enabling to repair incorrect 3′-ends.
First, pppAG formation was tested on the four DV103′- variants in the presence of only ATP and GTP. Figure 3A shows that NS5PolDV is indeed able to form pppAG in the presence of templates with any 3′-nucleotide and Mn2+. In contrast, in the presence of Mg2+ only the natural DV103′- CU template supports pppAG formation even in the presence of an increased concentration of NS5PolDV (Figure 3B). We then tested pppAG formation exclusively in the presence of Mn2+ on all DV103′- variants in the presence of all nucleotides, a scenario putatively mimicking the situation within the replication complex. Figure 3C shows that pppAG is always formed in parallel to the dinucleotide, which corresponds to the template. In the case of the template variant with a -CG 3′-end, pppAG is produced with even higher efficiency than the base-paired dinucleotide. Note that the dinucleotide pppGU is also produced on all templates by internal initiation. For the reaction in the presence of all templates and all nucleotides, we quantified all products, which were initiated de novo over the very 3′-end, and found that pppAG is formed as the prominent product (32.3±1.5%, three independent reactions). Note that all templates are present at the same concentration, which should not correspond to the situation in vivo. We conclude that in the presence of incorrect templates and Mg2+, NS5PolDV discriminates against these templates and forms pppAG only on the correct template (see also Figure 2A). In contrast, Mn2+ ions enable NS5PolDV to preferentially generate pppAG even in the presence of incorrect templates, which could represent an indirect way of 3′-end repair.
We then considered the elongation of the correct pppAG primer over templates with incorrect 3′-ends. We thus tested the elongation of a chemically synthesized pppAG primer (see Materials and Methods) either without template or in the presence of the four DV103′- variants (Figure 4). The most prominent result is that NS5PolDV is able to productively elongate pppAG on the correct template in the presence of Mn2+ (Figure 4A) and Mg2+ ions (Figure 4B). We also observe that NS5PolDV in the presence of Mn2+ is able to productively elongate pppAG on incorrect templates (Figure 4A), thus demonstrating that the enzyme is able to indirectly correct the error in the template and conserve the 5′-end of the DV genome. Note that as expected there is no primer elongation detectable in the absence of a template.
NS5PolDV harbors an ATP-specific priming site, which is essential for the formation, accumulation, and elongation of the correct primer pppAG. Which elements of NS5PolDV form this site? The crystal structure of NS5PolDV (Figure 5A) allowed the prediction of a priming loop comprising residues 792 to 804 [16], which is expected to provide the priming site during de novo RNA synthesis initiation. We generated a deletion mutant (NS5PolDV TGGK) by replacing residues T794-A799 between T793 and K800 by two glycines (see close-up in Figure 5A). The overall correct folding of the purified, recombinant mutant protein was verified by a fluorescent thermal shift assay giving identical temperatures of denaturation (melting temperature Tm) for both proteins (wild type (wt) NS5PolDV Tm 49.0°C ± 0.5°C, NS5PolDV TGGK Tm 48.4°C ± 0.05°C).
The TGGK mutant is expected to have an open active site, which impedes correct ATP-specific de novo initiation over the 3′-end of a single-stranded RNA template but may favor the accommodation of double-stranded RNA. Its RNA synthesis initiation and elongation activity was first tested using a “minigenomic” RNA template consisting of 224 nucleotides of the 5′-end of the DV genome fused to 492 nucleotides of the 3′-end [14]. It has been shown before using this template and analyzing the products on a denaturing agarose-formaldehyde gel [29] that two types of product are formed (see wt reaction kinetics in the center panel of Figure 5B). Firstly, the de novo RNA synthesis product is generated corresponding to the size of the template. Secondly, an elongation product is generated by back-primed RNA synthesis. There, the 3′-end (…AACAGGUUCU-3′) forms a short hairpin annealing the last di-nucleotide to nucleotides -6 and -7 (underlined in the sequence) and is then elongated [29]. The length of the product is thus ∼twice the size of the template. Reactions were carried out using either Mg2+ or Mn2+ as catalytic ions. The left and right panels of Figure 5B show that in both cases the mutant TGGK shows an increased overall activity on this template compared to wt activity. The center panel shows that this is mainly caused by increased back-priming. Interestingly, instead of one product species of twice the template size NS5PolDV TGGK produces a range of elongated products of different lengths. This might be due to the accommodation of long hairpins, which then create longer products than the template but shorter than the elongation product of wt NS5PolDV.
De novo RNA synthesis initiation by wt NS5PolDV and the TGGK mutant were then tested on DV103′-, in the absence of a template and on DV103′+ using Mn2+ as the catalytic ion, ATP and GTP containing αGTP. Figure 5C (panel 1) shows that in contrast to wt NS5PolDV, NS5PolDV TGGK is not able to catalyze de novo initiation on DV103′-. Secondly, NS5PolDV TGGK does not catalyze pppAG formation without template (panel 2). In contrast, it is able to catalyze de novo initiation on DV103′+ presenting ca. 32% of wt activity (panel 3). In order to understand this apparent contradiction, we used γATP instead of αGTP as radioactive NTP. It became clear that NS5PolDV TGGK was unable to generate the pppAG primer product (panel 4). We conclude that the product observed with αGTP corresponds to pppGA formed by internal de novo initiation being only possible on DV103′+. When using Mg2+ as catalytic ion again we did not observe formation of the de novo RNA synthesis initiation product pppAG on either template (for DV103′- see below Figure 6B).
We conclude that NS5PolDV TGGK is unable to pre-form the ATP-specific priming site necessary for de novo RNA synthesis initiation at the very 3′-end. The predicted priming loop plays indeed an essential role in providing the correct priming site. We explain the increased activity of NS5PolDV TGGK on minigenomic RNA templates by its increased propensity to catalyze back-priming due its more accessible catalytic site, i.e. to harbor the minigenome in different hairpin conformations allowing 3′ elongation.
Two aromatic residues, W795 and H798, within the priming loop were proposed to play a particular role in providing an initiation platform to which the base of the priming ATP could establish a stacking interaction [16]. Residue W795 was given special attention because it was found near the triphosphate moiety of a 3′-dGTP bound to NS5PolDV [16]. In addition, this tryptophan was better placed than the histidine for stacking a priming ATP in two models of de novo RNA synthesis initiation complexes of NS5PolDV and NS5PolWNV [16], [22]. We generated two mutants of NS5PolDV, W795A and H798A. Overall correct folding of the purified recombinant mutants was equally verified by a fluorescent thermal shift assay giving Tm values corresponding to the wt protein (wt NS5PolDV Tm 49.0°C ± 0.5°C, W795A mutant Tm 48.6°C ± 0.6°C, H798A mutant Tm 48.1°C ± 0.04°C).
The RNA initiation and elongation activities of wt NS5PolDV and the W795A and H798A mutants were tested using the minigenomic RNA template and either Mg2+ or Mn2+ as catalytic ions (Figure 6A). In both cases the H798A mutant shows an increased activity on this template whereas W795A shows a similar overall activity compared to wt NS5PolDV. Figure 6B shows the analysis of the reaction products on a denaturing agarose-formaldehyde gel. The W795A mutant behaves indeed like wt NS5PolDV, the percentage of the de novo RNA synthesis initiation product of template size is unchanged. In contrast the H798A mutant generates considerably less de novo RNA synthesis product whereas the yield of RNA elongation products is higher.
We then compared the capacities of wt and all mutant NS5PolDV proteins to catalyze de novo RNA synthesis initiation on DV103′-, without template and on DV103′+ using Mn2+ as catalytic ion (Figure 6C panels 1, 3 and 4). Indeed, the H798A mutant is considerably less capable of correct de novo RNA synthesis initiation than wt NS5PolDV whereas W795A behaves as wt NS5PolDV. Note that the product formed by NS5PolDV TGGK on DV103′+ (panel 4) corresponds to pppGA generated by internal RNA synthesis initiation (see also Figure 5C); and therefore part of the product formed by the H798A mutant may correspond to pppGA. When Mg2+ is used on both templates, the same results are obtained (Figure 6C panel 2 for template DV103′-). We thus conclude that residue H798 is essential for the formation of the correct ATP-specific priming site and may act as a priming platform.
In this study, we present evidence that the dengue virus NS5 polymerase domain (NS5PolDV) alone is responsible for maintenance of A and U as first and last nucleotides of the DV genome, respectively. NS5PolDV was used instead of full-length NS5 in the frame of this study in order to avoid any interference of the RNA-binding, NTP-binding, or enzymatic activities of the N-terminal domain of NS5. We report that NS5PolDV is endowed with several structural and mechanistic features converging to the specific de novo synthesis and elongation of the correct ATP-initiated primer even on templates that lack the correct corresponding U at the 3′-end. The first and last nucleotides of the genome are strictly conserved in the genus Flavivirus thus the results presented here may apply to the entire genus.
We demonstrate the generation of a dinucleotide primer pppAG on both genomic and antigenomic RNA templates. We have previously observed the production of such dinucleotide primer on homopolymeric templates [14]. In the following step pppAG(A/U) trinucleotides are formed before processive RNA elongation occurs. During the latter, NS5PolDV continues RNA synthesis to the very end of the template. We do not know if di- and tri-nucleotide primers as detected in the reaction, originate from a slow but processive RNA synthesis reaction, or are actually released from the complex and re-used by the polymerase acting in a distributive RNA synthesis mode. We also show that the pppAG primer is effectively elongated in the presence of Mg2+ or Mn2+ and the correct template. Thus, after initial phosphodiester bond synthesis, the pppAG primer is aligned at the correct position in order to be elongated. The efficient use of the short primer pppAG reported here is in apparent contrast to the inefficient use of 5′-OH-AG dinucleotide previously reported [13], [30]. The 5′-triphosphate moiety of the chemically synthesized pppAG primer is most probably an important binding determinant allowing efficient elongation (see discussion of the proposed de novo initiation complex Figure 7).
We then demonstrate that in its de novo RNA synthesis initiation state NS5PolDV contains a built-in ATP-specific priming site. Major structural elements of NS5PolDV contributing to this site reside within residues T794 to A799. Their deletion forces NS5PolDV to initiate de novo RNA synthesis internal to the template using GTP as the first nucleotide (Figure 5C panel 1) and to perform primer-dependent RNA synthesis (Figure 5B). In analogy to the structure of HCV NS5B in complex with a nucleotide in its priming site [31] and because of the amino acid conservation observed within a larger group of de novo RdRps [25], we expect that NS5PolDV residues R472 (RdRp catalytic motif F3, see [14]) as well as S710 and R729 (motif E) are involved in triphosphate binding. This might explain why de novo RNA synthesis initiation by the loop-deleted mutant is still possible, albeit internal to the template. We conclude that indeed the T794-A799 loop plays a major role both in correct de novo initiation and in shaping the priming site. Within the priming loop, residue H798 is essential for primer synthesis (Figure 6). We propose that H798 provides the initiation platform against which the priming nucleotide ATP is stacked. Using the structure of the de novo initiation complex of the RdRp of bacteriophage φ6 [23] as a starting point, we generated a model of the initiation complex of DV serotype 2 RdRp in complex with the 3′- end of the genome UUCU and both ATP and GTP as first and second nucleotide, respectively (Figure 7). In this model, the triphosphate moiety of ATP indeed interacts with residues S710, R729 and R737 of the thumb subdomain of NS5PolDV. The aromatic ring of H798 stacks the adenine nucleobase of ATP in a similar position to a φ6 RdRp tyrosine residue against which the guanine nucleobase of its priming GTP is stacked. In several protein complex structures histidine has been shown to bind an adenine nucleobase by stacking interactions [32]. Nevertheless, histidine does not seem to provide any specificity towards adenine versus guanine [33]. Our model does not propose any obvious specific interaction with the adenine base. This might be due to the fact that the structure of NS5PolDV has been captured in a pre-initiation state. In this state, motif F, which provides the upper part of the NTP entry tunnel in the active initiation and elongation conformation of viral RdRps, is not yet correctly positioned [34]. The fine characterization of the ATP-specific built-in priming site of NS5PolDV awaits the crystal structure of a de novo RNA synthesis initiation complex.
We provide a mechanistic basis for the conservation of nucleotides A and U as the first and last nucleotides of the DV genome, respectively. Figure 8 summarizes the different levels of control that ensure ATP-specific de novo RNA synthesis initiation. Firstly, it generates and elongates the bona fide pppAG primer (red arrows and green arrows on the right). Even in the absence of any template and in the presence of Mn2+ (Figure 8 left red arrow) NS5PolDV is able to exclusively synthesize the pppAG primer (Figure 2B and C, Figure 3A and C). Note that we have also observed pppAG synthesis by full-length NS5 in the absence of a template (not shown). Since a sufficiently high Mn2+ concentration is present in the cell (0.1 µM to 40 µM Mn2+ in blood, brain, and other tissues [35]), NS5 in the replication complex might already be loaded with pppAG and thus be ready to elongate pppAG on the viral template. The same pppAG primer is preferentially synthesized in the presence of the correct template irrespective of the metal ion present at the polymerase active site (Figure 8 right red arrows, Figure 2A and B, Figure 3). In the presence of Mg2+, NS5PolDV supports neither formation nor elongation of pppAG on incorrect templates (Figure 8 blue blocked arrow, Figure 4B). In the presence of Mn2+, NS5PolDV is able to synthesize cognate dinucleotides on incorrect templates (Figure 2C), but in the presence of all nucleotides and all templates (a probably biased and more unfavorable set-up compared to the situation in the replication complex in vivo), pppAG is still a major product (Figure 3C). Remarkably, the pppAG/Mn2+-loaded polymerase is able to mismatch and extend pppAG in order to restore the correct 5′-end (Figure 8 blue arrows, Figure 4). The selective extension reaction thus refrains synthesis of incorrect RNAs that could occur in the presence of incorrect templates. All these reactions converge to the formation of pppAG and the conservation of A as the starting nucleotide at the 5′-end of viral genomic and antigenomic RNAs. Note that the mechanistic basis of the conservation of the second nucleotide G is beyond the scope of this study. Preliminary results generated in our laboratory indicate that both template and polymerase are important to ensure the specific incorporation of a G as the second nucleotide (not shown).
Several ways of viral RNA genome maintenance and repair concerning terminal damage have been discussed [4], among others the generation of “non-templated” primers and the use of abortive transcripts as primers. Here we demonstrate that NS5PolDV uses these two mechanisms. Non-templated primers are generated only in the presence of Mn2+. Abortive transcripts are used as primers in the presence of either Mg2+ or Mn2+. A third mechanism observed here is the discrimination against an incorrect template in the presence of Mg2+. In addition, in the case that a 3′-end might be shortened, the correction upon de novo initiation should be preceded by the addition of (a) nucleotide(s) by the terminal transferase activity of NS5. This activity has also been listed as another way of repairing terminal damage of viral RNA genomes [4]. For NS5PolDV we have observed this activity before [14] and now again in the presence of Mn2+ (Figure 1B).
The DV polymerase endows several of the proposed mechanisms to maintain the correct 5′ and 3′-ends of the DV genome and antigenome. The ability of DV and WNV to restore a U at the very 3′-end of genomes with 3′-end deletions has been demonstrated [2], [36]. This observation is in accordance with the existence of an ATP-specific priming site in NS5PolDV. Tilgner et al. [2], [36] reported the complete reversion of WNV replicon CA and CG 3′-ends to CU whereas CC was only partially reverted. Since we have not seen preferential de novo RNA synthesis initiation starting with GG in comparison to UG or CG (all three are possible in presence of Mn2+, Figure 2), this might be due to an intrinsic difference between DV and WNV RdRp or caused by different propensities of the erroneous templates to allow pppAG elongation. Indeed CA and CG 3′-ends allow pppAG elongation more readily than the CC 3′-end (Figure 4, two independent reactions were performed). Thus the CC 3′-end might therefore take longer to revert. Furthermore, Teramoto et al. [2], [36] observed the correction of the 5′-end from pppGAG to pppAG. Our work provides a mechanistic explanation for their observation.
The observation of non-templated pppAG formation in the presence of Mn2+ by a viral RdRp has not been reported before using recombinant RdRp assays. However, previous reports convey the occurrence of non-templated dinucleotide formation. RSV, a member of the ns-RNA virus family Paramyxoviridae restores the correct 5′-pppA although minireplicons did not encode the correct 3′-U [10]. The authors propose that RSV RdRp contains a built-in ATP-specific priming site and cite the observation that the RdRp of the related ns-RNA vesicular stomatitis virus (VSV, Rhabdoviridae) contains a specific ATP-binding site [37] as an argument in favor of their proposition. When VSV RdRp assays were carried out using recombinant RdRp in the presence of Mg2+, non-templated 5′-initiation was not observed [6]. There is either the possibility that RSV and VSV belong to two different ns -RNA viral families and thus developed different strategies or, in analogy to our results that their RdRps use Mn2+ to correctly initiate RNA synthesis on erroneous templates as observed for NS5PolDV here. It is generally believed that Mg2+ is the activating cofactor of polymerases in vivo because viral RdRp properties observed with Mg2+ in vitro are more consistent with properties observed biologically. A second reason for giving the preference to Mg2+ is its cellular abundance in comparison to Mn2+ (i.e., 0.5 mM free Mg2+ versus 0.7 µM free Mn2+ in rat hepatocytes [38], [39] and 0.1 µM–40 µM Mn2+ in blood, brain and other tissues [35] versus 0.2 to 0.7 mM Mg2+ in human blood [40]. Nevertheless, some events especially involved in correct and efficient de novo RNA synthesis initiation may require the specific use of Mn2+ by viral RdRps under physiological conditions (our study and [10], [36], [41], [42]).
The pppAG primer synthesis by the DV RdRp can be considered as the first line of control of the conservation of Flavivirus genome and antigenome ends. However, there might be other mechanisms to tighten the selection. The first one could be the base pairing of the genome ends maintaining specific RNA secondary structures, which are necessary to recruit the replication machinery. Computer simulations of such structures [43] indicate that the last U of the 3′-end of the genome may be unpaired or paired (structure I or II, respectively in [43]). Thus, requested base pairing may exert selective pressure to keep a U at the end of the Flavivirus genome. Another selection level concerns only the 5′-end of the genome and is due to the counterselection of incorrect 5′-ends through the NS5 RNA-cap methyltransferase. Indeed, several crystal structures of the cap-dependent bi-functional methyltransferase domain of NS5 show that specific binding of the 5′-cap involves specific recognition of the first transcribed 5′-adenosine through its N1 position and residue Asn18 [44], [45]. Therefore, for the genomic strand, methylation at the cap N7-guanine and the subsequent 2′-O position of the first transcribed adenosine should be efficiently achieved only when ATP is the starting 5′-nucleotide. Finally, cap addition seems to involve 5′-ATP selectivity as well [20]. Collectively, we propose that the RdRp of flaviviruses is the first actor responsible for the conservation of the correct ends of their genome, and that other mechanisms such as genome cyclization and the specificity of guanylyltransferase and methyltransferase activites add to the selective pressure. These mechanisms of maintenance might also apply to other RNA virus genera with conserved genome ends and viral RdRps initiating RNA synthesis de novo.
DV103′+ (5′-AACAGGUUCU-3′) and DV103′- (5′-ACUAACAACU-3′) were synthesized at the RIBOXX GmbH Dresden and by Dharmacon. The templates are devoid of stable secondary structure when submitted to the Mfold server [46] (ΔG = 3.60 kcal/mol for DV103′+ and no folding for DV103′-).
The gene coding for N-terminal His6-tagged NS5PolDV (serotype 2, New Guinea C) as defined in [14] cloned in a pQE30 plasmid was expressed in E.coli (Tuner (Novagen) or NEB Express (New England Biolabs)) cells carrying helper plasmid pRare2LacI (Novagen). Expression was carried out in Luria broth overnight at 17°C after induction with 50 µM IPTG, addition of 2% EtOH and a cold shock (2 h at 4°C). Sonication was done in 50 mM sodium phosphate lysis buffer, pH 7.5, 500 mM NaCl, 20% glycerol, 0.8% Igepal (10 ml of this lysis buffer for around 2 g cell pellet from 1l culture) in the presence of DNase I (22 µg/ml), 0.2 mM benzamidine, protease inhibitor cocktail (SIGMA), 5 mM β-mercaptoethanol and 1 mg/ml lysozyme after 30 min incubation at 4°C. After centrifugation the soluble fraction was incubated in batch with 2 ml TALON metal-affinity resin slurry (Clontech) for 40 min at 4°C. Protein bound to the beads was washed once with 10 volumes of sonication buffer containing 1 M NaCl and 10 mM imidazole and once with the former buffer without Igepal. Protein fractions were then eluted with sonication buffer containing 250 mM imidazole, no Igepal and 250 mM glycine. After dialysis into 10 mM Tris buffer, pH 7.5 containing 300 mM NaCl, 20% glycerol, 250 mM glycine and 1 mM DTT the protein was diluted with the same volume of this buffer without NaCl and loaded onto a HiTrap heparin column (GE Healthcare). Pure NS5PolDV was then eluted in a single peak applying a gradient from 150 mM to 1 M NaCl. Alternatively, gel filtration was used as a second purification step using a Superdex 75 HR 16/60 column (GE Healthcare) and the dialysis buffer. NS5PolDV was stored at −20°C at a concentration of 40 to 60 µM after a final extensive dialysis into 10 mM Tris buffer, pH 7.5 containing 300 mM NaCl, 40% glycerol and 1 mM DTT. Purity was higher than 98% as judged by SDS-PAGE.
Mutant TGGK, W795A and H798A NS5PolDV expression plasmids were generated using the kit QuikChange (Stratagene). Protein expression and purification was done as for the wt protein. Analysis by gel filtration showed a single peak eluting at the same volume as wt NS5PolDV.
Melting temperature (Tm) values of wt and mutant NS5PolDV were determined using a thermofluor-based assay [49]. In 96-well thin-wall PCR plates 3.5 µl of a fluorescent dye (Sypro Orange, Molecular Probes, 714-fold diluted in H2O) was added to 21.5 µl protein solutions at a concentration of 0.5 or 1 mg/ml (6.7 or 13.4 µM) in storage buffer. Thermal denaturation of the proteins was followed by measuring fluorescence emission at 575 nm (excitation 490 nm). Tm values were calculated using GraphPad Prism software and the Boltzmann equation as in [49].
Reactions were done in 50 mM HEPES buffer, pH 8.0 containing 10 mM KCl, 10 mM DTT and template, NS5PolDV, non-labeled NTPs, and catalytic ions at final concentration as given in the figure legends. Radiolabeled [γ-32P]-ATP, [α-32P]-GTP, or [α-32P]-UTP was used at 0.4 µCi per µl reaction volume (3000 Ci/mmol, Perkin-Elmer). Reactions were started by addition of a mixture of HEPES buffer, KCl, catalytic ions and UTP and CTP when used (given in Figures). After given time points samples were taken and reactions stopped by adding an equal volume of formamide/EDTA gel-loading buffer. Reaction products were separated using sequencing gels of 20% acrylamide-bisacrylamide (19∶1), 7 M Urea with TTE buffer (89 mM Tris pH 8.0, 28 mM taurine (2-aminoethanesulfonic acid), 0.5 mM EDTA). RNA product bands were visualized using photo-stimulated plates and the Fluorescent Image Analyzer FLA3000 (Fuji) and quantified using Image Gauge (Fuji). The oligoG marker was produced as explained in [14].
The minigenomic template was produced by in vitro transcription and tests carried out as described in [14].
Reactions analyzed by filter-binding and liquid scintillation counting contained 50 mM HEPES buffer, pH 8.0, 10 mM KCl, 10 mM DTT, 100 nM RNA template, 200 nM NS5PolDV, 500 µM NTP except for UTP (4 µM), [3H]-UTP at 0.2 µCi/µl and either 5 mM MgCl2 or 2 mM MnCl2. Reactions were started by the addition of a mixture of HEPES, KCl, catalytic ions, CTP, and UTP. After 30, 60, 90, and 120 min 10-µl samples were taken and diluted into 50 µl of 100 mM EDTA, pH 8.0 to quench the reaction. Samples were then transferred onto a DEAE filter mat. Non-incorporated [3H]-UTP was removed by washing with 300 mM ammonium formate and the radioactively labeled product quantified in counts per minute (cpm) using liquid scintillation counting. Product formation was then plotted against time and initial velocities calculated in cpm/min.
Reactions analyzed on formaldehyde-agarose gels contained 50 mM HEPES buffer, pH 8.0, 10 mM KCl, 10 mM DTT, 100 nM RNA template, 200 nM NS5PolDV, 500 µM NTP except for UTP (4 µM), [α-32P]-UTP at 0.4 µCi/µl, and 5 mM MgCl2. Reactions were started by a mixture of HEPES, KCl, MgCl2, CTP and UTP and stopped after 60 and 120 min by adding an equal volume of sample buffer (40 mM MOPS pH 7.0, 83.3% formamide, 2 M formaldehyde, 10 mM sodium acetate, 85 mM EDTA). Samples were denatured for 10 min at 70°C and 1/10 of loading buffer (50% glycerol, 10 mM EDTA, xylene cyanol and bromphenol) added. Samples were then analyzed on a 1.2% agarose-formaldehyde gel in 20 mM MOPS buffer pH 7.0, 5 mM sodium acetate, 1 mM EDTA. Gels were dried and RNA product bands visualized using photo-stimulated plates and the Fluorescent Image Analyzer FLA3000 (Fuji) and quantified using Image Gauge (Fuji).
A homology model of NS5PolDV serotype 2 strain New Guinea C was generated using the Swiss-model server [50] and the X-ray structure of NS5PolDV serotype 3 (PDB code 2J7W [16]). NS5PolDV and the RdRp of bacteriophage φ6 in complex with a template RNA strand and initiating NTPs (PDB code 1HI0) were then superimposed using the three catalytic aspartate residues of both proteins. The structural model of the initiation complex of NS5PolDV serotype 2 was then generated by changing the RNA template to UUCU (3′-end of the DV genome) and the initiating NTP to ATP, and by manually adapting the conformation of the priming loop using the UCSF Chimera software [51]. Subsequently using the same program the computed free energy of the model was minimized.
|
10.1371/journal.ppat.1002547 | PK-sensitive PrPSc Is Infectious and Shares Basic Structural Features with PK-resistant PrPSc | One of the main characteristics of the transmissible isoform of the prion protein (PrPSc) is its partial resistance to proteinase K (PK) digestion. Diagnosis of prion disease typically relies upon immunodetection of PK-digested PrPSc following Western blot or ELISA. More recently, researchers determined that there is a sizeable fraction of PrPSc that is sensitive to PK hydrolysis (sPrPSc). Our group has previously reported a method to isolate this fraction by centrifugation and showed that it has protein misfolding cyclic amplification (PMCA) converting activity. We compared the infectivity of the sPrPSc versus the PK-resistant (rPrPSc) fractions of PrPSc and analyzed the biochemical characteristics of these fractions under conditions of limited proteolysis. Our results show that sPrPSc and rPrPSc fractions have comparable degrees of infectivity and that although they contain different sized multimers, these multimers share similar structural properties. Furthermore, the PK-sensitive fractions of two hamster strains, 263K and Drowsy (Dy), showed strain-dependent differences in the ratios of the sPrPSc to the rPrPSc forms of PrPSc. Although the sPrPSc and rPrPSc fractions have different resistance to PK-digestion, and have previously been shown to sediment differently, and have a different distribution of multimers, they share a common structure and phenotype.
| Prion diseases are protein misfolding disorders. Different strains of prions are known to have variable resistance to proteinase K (PK) digestion. Furthermore, the same strain possesses both a PK sensitive (sPrPSc) and PK resistant (rPrPSc) aggregate of PrP. We developed methods to isolate the sPrPSc from rPrPSc fraction of the 263K strain of hamster-adapted scrapie. Both fractions were infectious, but have different physico-chemical properties. When we analyzed the lesion targets in the brain produced by each fraction they were essentially identical, suggesting that they were the same strain. The biochemical differences in the phenotypes of these two fractions are due to different sized multimers that share common structural properties. Furthermore, the comparison of the sensitive fractions of two hamster strains, 263K and Drowsy (Dy), showed strain-dependent differences in the ratios of the PK-sensitive to the PK-resistant forms of PrPSc.
| The prion (PrPSc) is the infectious agent responsible for a suite of different rare animal and human diseases known as transmissible spongiform encephalopathies (TSEs) [1], [2], [3], [4], [5]. PrPSc is able to convert a normal cellular prion protein (PrPC) into PrPSc when both isoforms make contact, and thereby propagate an infection. The conversion of PrPC into PrPSc involves a conformational change of the protein in which the total amount of β-sheet increases and that of α-helical secondary structure decreases or perhaps disappears [6], [7]. PrPSc is a multimer, while PrPC is monomeric [8], [9]. These conformational differences are the only demonstrated structural differences between PrPSc and PrPC [10]. Detailed mass spectrometric analysis showed they have identical amino acid sequences [11]. No post-translational differences have been found between PrPSc and PrPC: both share one disulfide bond, two or less sugar antennae and a single glycophosphatidylinositol (GPI) anchor [12]. The composition of the sugar antennae and the GPI anchor vary similarly in both PrPSc and PrPC [13]. On the other hand, the conformational change and consequent aggregation makes PrPSc insoluble in non-denaturing detergents and partially resistant to PK digestion [1]. Thus, treatment of a sample with 50 µg/ml of PK for 1 hour at 37°C completely destroys PrPC, while, typically, PrPSc is partially cleaved at the amino terminal portion, leaving a PK-resistant core termed PrP 27–30.
In 1998 Safar et al. reported the existence of a subset of PrPSc molecules that are completely degraded by PK, which hence were termed, PK-sensitive PrPSc (sPrPSc) [14]. Tzaban et al. later demonstrated for the first time that prion-infected tissues contain sPrPSc molecules that form low molecular weight aggregates [15]. These authors subjected brain homogenates from scrapie-infected animals to sucrose gradients, and found that PrPSc was distributed in a continuum of aggregation sizes. The more dense fractions, corresponding to larger multimers, were PK-resistant, whereas the intermediate fractions, corresponding to smaller multimers, were not. It has also been described that as much as 90% of total PrPSc in the brains of individuals who had died as a consequence of Creutzfeldt-Jakob disease (CJD) was estimated to be sPrPSc [16].
Different studies on other protein misfolding diseases, such as Alzheimer's disease, suggest that large amyloid fibrils may be a means of protecting the host by sequestering the smaller and more toxic multimers as larger less toxic fibrils [17]. In the case of prion diseases, infectivity studies of the different sized fractions of hamster PrPSc revealed a several-fold increase in infectivity for non-fibrillar particles with masses equivalent to 14–28 PrP molecules [18]. These PrPSc particle sizes correspond to the sPrPSc fraction according to Tzaban and our own group's characterization using sucrose gradient ultracentrifugation [15], [19]. Although the definition of sPrPSc is operational, a question arises: are sPrPSc and rPrPSc two populations with different conformations or simply different sized multimers with the same conformation?
To address this question, we investigated first the infectivity of the sPrPSc vs. the PK-resistant (rPrPSc) fraction of hamster PrPSc (263K strain). sPrPSc was further characterized by limited proteolysis and mass spectrometry. Then, PrPSc infectivity and strain characteristics were assessed by inoculation into Syrian hamsters and comparison of the resulting incubation period and lesion distribution with that obtained after inoculation with either total PrPSc or the PK-resistant fraction of PrPSc. The obtained results are reported here.
We obtained sPrPSc and rPrPSc (263K ) fractions from our starting total PrPSc by our previously described ultracentrifugation-based method [19]. When these fractions were subjected to PK digestion for 1 h at 37°C, virtually complete degradation of sPrPSc occurred after treatment with 50 µg/ml of PK whereas partial resistance occurred with lower concentrations of the enzyme (Figure 1). In contrast, rPrPSc was much more resistant to PK under these conditions (Figure 1). These results fully agree with our previously published work [19]. The sPrPSc, rPrPSc, purified PrPSc, and the unpurified PrPSc material was used in our subsequent experiments.
We performed a bioassay of normalized preparations of four prion isolates, purified PrPSc, sPrPSc, rPrPSc and the unpurified PrPSc-containing 0.1% brain homogenate. The amount of PrP present in each sample was determined by a mass spectrometry-based quantitative method [20]. The amount of residual PrPC present in the purified samples represents a negligible contribution [21]. Each of the preparations was then diluted, so that each contained approximately similar amounts of PrP. In addition, three dilutions of each of the four normalized samples were prepared: 1/10, 1/100, and 1/1,000 for the 0.1% brain homogenate sample and 1/100, 1/1,000, and 1/10,000 for the other three ones. Each of these dilution series (16 in total) was inoculated into a set of eight Syrian hamsters (LVG) (128 total). The dates of the appearance of clinical signs and the date of euthanization at terminal disease were recorded [22]. These data were plotted as a Kaplan-Meier estimate graph (Figure 2) for each dilution of each preparation. For further clarity, they are also presented in table form (Table S1).
Animals inoculated with the undiluted sPrPSc preparation got sick sooner than those inoculated with any of the other three undiluted preparations (P<0.01). By the 1/1000 dilution all of the hamsters inoculated with the purified PrPSc (sPrPSc, rPrPSc, and total purified PrPSc) had a similar incubation time, which indicates a roughly similar infectivity. When the rPrPSc and purified PrPSc preparations were diluted 1/10,000 all of the animals became sick. In contrast, of the eight animals inoculated with the 1/10,000 dilution of sPrPSc, only seven became infected. Even after 240 days, the eighth animal showed no clinical signs and there was no detectable PrPSc in its brain by mass spectrometry-based analysis [20]. Paradoxically the sPrPSc shows a shorter incubation time in the undiluted preparation (Table S2). Upon dilution of the sPrPSc preparation the incubation time disproportionately increases compared to the other purified preparations (purified PrPSc and rPrPSc). At this point, it was not clear if these differences were due to the possibility that our isolation procedure facilitates the isolation of different prion strains, as has been described by Bessen and Marsh [23], or if these differences were caused by kinetic factors, due to the smaller size of the PrPSc multimers present in the sPrPSc fraction [15], [19].
These results show that sPrPSc is infectious, and that its infectivity is comparable to that of rPrPSc. We wanted to see how sPrPSc behaved after PK treatment. We took a sample of sPrPSc and divided it into two portions. One was untreated and the other was digested with PK (50 µg/ml of PK; 37°C; 1 h). These samples were inoculated (ic) into two sets of six hamsters. The disease course in both sets was observed. In the group inoculated with PK-treated sPrPSc, animals succumbed 15 days later than in the untreated group (Figure S1). This time interval corresponds to a 100 fold dilution of the sPrPSc fraction (Figure 2). Granting that only an approximate correlation between incubation times and titers can be surmised in our experimental conditions, given that we did not perform a full calibration [20], this result suggests that PK destroyed approximately 99% of the infectivity present in the sPrPSc fraction. Unlike the unpurified 0.1% brain homogenate, where the loss of infectivity upon treatment with PK is much greater (∼99%) than the loss of WB signal (∼70%) [24], the loss of signal and loss of infectivity are proportionately large (∼99%) with the sPrPSc fraction (Figure S1).
To assess whether sPrPSc is also infectious via the ip route, we inoculated intraperitoneally two groups of animals with equal amounts of sPrPSc and rPrPSc, respectively. Both groups succumbed to infection, and although not statistically significant, animals inoculated with the sPrPSc fraction showed an incubation time slightly shorter (116±9 dpi) than the ones inoculated with the rPrPSc fraction (123±14 dpi) (Figure S2). As expected, animals inoculated ip survived longer than those inoculated ic (intracerebral) [25], [26].
After determining that there were phenotypic differences (incubation times) between sPrPSc and a complete mixture of PrPSc, we explored the possibility that there were structural differences between these two PrPSc fractions. We have previously reported that, in addition to the well-known N-terminal PK cleavage sites, around position G-90, there are other additional cleavage sites [27]. In that study, we prepared a cleavage map for total PrPSc after treatment with PK and then treatment with NTCB. The N-terminal cleavage sites include: residues 86 (G-86→D-178), 90 (G-90→D-178), 92 (G-92→D-178), 98 (Q-98→D-178), and 101 (K-101→D-178). Additional cleavage sites comprise residues 117 (A-117→D-178), 119 (G-119→D-178), 135 (S-135→D-178), 139 (M-139→D-178), 142 (G-142→D-178), and 154 (M-154→D-178) ([27] and Figure S3). We used an analogous approach to prepare a cleavage map of sPrPSc. Briefly, sPrPSc was digested with 1 µg/ml of PK for 30 minutes and subsequently treated with NTCB. This reagent cyanylates free cysteines and NaOH is subsequently used to cleave amide bonds at the N-terminal side of the modified cysteine residues [27], [28]. Given the difficulty in identifying the GPI anchor or sugar-containing peptides by mass spectrometry, only those free of sugar moieties were analyzed. The NTCB reagent cleaves at the cysteine residue at position 179, thereby separating the GPI-anchor and glycosylated portion of the protein from the set of amino-terminal truncated peptides. These peptides contain no post-translational modifications and were analyzed by MALDI-TOF.
The MALDI-TOF analysis of the NTCB treated sPrPSc fraction showed cleavage sites identical to the ones previously described for total PrPSc, namely, 117, 119, 135, and 139 (Figure 3a). Curiously, the more abundant peptides from the classic N-terminal PK-cleavage sites (86, 90, and 92) are present in the total PrPSc samples [27], but not in the sPrPSc fraction (Figure 3). This suggests that the N-terminal portion of the PrP molecules present in the sPrPSc fraction might be more exposed and therefore more susceptible to PK digestion than is the N-terminal portion present in the complete mixture of PrPSc. However, it should be remembered that MALDI-TOF is not a quantitative technique and that the signal response of larger peptides, such as (G-86→D-178), or (G-90→D-178), is poorer than that of smaller ones. Indeed, cleavages at around position 90, and perhaps slightly more amino-terminal sites, are prominent in sPrPSc as shown by WB analysis (vide infra).
The similarity of the PK-cleavage pattern between sensitive and total PrPSc is also evident when comparing the digested samples by Western blot assay using a C-terminal antibody after PK-treatment and deglycosylation (Figure 3b). Besides the intense band corresponding to cleavages around position 90, 6–7 additional lower molecular weight bands are seen in both cases, as we have previously described for total PrPSc [27]. Although we do not know the identity of each band, the apparent molecular weight of some of them could correspond with cleavages seen in MALDI-TOF. While the relative intensities of some of the bands vary between sPrPSc and total PrPSc, (the main band centered at 19–20 kDa smears a bit more into the 20–21 kDa region in the sPrPSc sample, the ∼17 kDa band is more intense in total PrPSc, bands around 15, and finally, bands at ∼10 and ∼6 kDa are more intense in sPrPSc), the overall pattern is very similar.
We wanted to ensure that in fractionating PrPSc, we had not inadvertently isolated a new prion strain. We performed comparative histopathological and immunohistochemical studies on the left hemisphere of brains from hamsters infected with each of the four PrPSc preparations. Spongiform change and PrP deposition were detected in all cases with no detectable differences in the PrPSc distribution or the morphology of the PrPSc aggregates among all four groups (Figure 4). The PrPSc appeared as diffuse as well as small, punctate aggregates. To assess whether there were differences in the brain regions targeted, lesion profiles of the hamsters inoculated with four different prion preparations was also performed. We compared the spongiform degeneration and PrPSc deposition in 6 brain regions. The results of this lesion profile analysis suggest that the isolation procedure did not yield different strains (Figure 5).
Western blot analysis of brain homogenates from the four inoculated groups showed the presence of PK resistant PrP in all groups. No differences in band patterns were detected (Figure S4). The relative amount of sPrPSc in each group was assessed by subtracting the amount PrPSc present in an aliquot digested with PK from the amount of PrPSc present in an untreated aliquot. The amount of PrPSc was determined by a previously reported mass spectrometry-based method [20]. The proportion of sPrPSc present in the brains of animals inoculated with either unpurified brain homogenate, purified prions, rPrPSc or sPrPSc was 0.3, 0.3, 0.6 or 0.5, respectively (means of duplicate analyses). In our previous work, we determined that sPrPSc accounted for between 35 and 60% of the total PrPSc present in the sample [19]. Thus the relative amount of sPrPSc in all groups is similar to that present in an unpurified brain homogenate.
We chose to use the drowsy (Dy) strain of hamster-adapted prion disease in order to compare the results with those obtained from 263K-derived prions. The phenotype of the Dy strain is very different from that of 263K [29]. It has a much longer incubation period and the typical symptoms are progressive lethargy and kyphosis [23]. The Dy strain is highly susceptible to PK digestion [30]. We wondered whether such susceptibility correlated with a higher ratio of sensitive to resistant fraction and if this would provide some insight into the nature of sPrPSc isolated from the 263K strain.
We isolated Dy sPrPSc using our method of isolating PK-sensitive prions (Figure S5) [19]. The sensitive fraction of Dy was shown to constitute approximately half of the total PrPSc, whereas in 263K, the analogous PK-sensitive fraction represents a considerably lower proportion of the total PrPSc (Figure S6). This could partially explain why Dy is more PK-susceptible than 263K. These results agree with those of Safar et al. who reported that different strains of hamster scrapie prions contain different ratios of sensitive to resistant PrPSc [14]. It should be noted that, although our PrPSc purification protocol involves sonication and therefore fragmentation of large aggregates, the sonication conditions used in 263K and Dy PrPSc isolation were the same.
The MALDI-TOF spectrum of the Dy sensitive fraction after digestion with 0.1 µg/ml of PK and NTCB cleavage showed the same internal cleavage sites as in 263K (117, 119, 135, and 139) plus others (101 (K-101→D-178) and 92 (G-92→D178)) (Figure 6) also previously described for total PrPSc (Dy strain) (Figure S3 and [27]).
Most prions have a degree of resistance to proteinase K digestion. Digestion of prions with proteinase K yields a characteristic prion core referred to as PrP 27–30 and a loss of infectivity that is disproportionate to the loss of protein. As researchers examined hamster strains they realized that not all strains of prions were equally resistant to proteinase K digestion, as is the case with the hyper (Hy) and drowsy (Dy) strains of hamster-adapted transmissible mink encephalopathy [23], [30]. Other researchers observed a similar phenomenon with other prions [31], [32], [33], [34]. Later Safar showed that PrPSc contained both a proteinase K sensitive and proteinase K resistant fraction [14]. These results indicated that PrPSc can be both infectious and proteinase K sensitive and that these phenotypes are a general characteristic of prions.
Since Safar et al. reported the existence of a PK-sensitive fraction in PrPSc, little has been published about its structural characteristics and infectious properties [14]. Although its infectivity has been suggested [14], [35], it had not been proven. In this work we demonstrate that the sensitive fraction of PrPSc is roughly as infectious as the resistant one. Highly significant is the fact that both fractions presented similar incubation times even when the respective inocula were diluted 1000 times. It has been shown that when PrPSc was fragmented by sonication, infectious titer was reduced because the rate of PrPSc clearance from the brain increased [36]. PrPSc is rapidly cleared from the brain once inoculated and the rate of clearance is influenced by the particle size [37]. Our results indicate that both rPrPSc and sPrPSc fractions have the same degree of infectivity indicating that the aggregation state of PrPSc does not affect its infectivity. This is in contrast with a previous report where the majority of sPrPSc was found to be not infectious when measured by the scrapie cell assay [38]. Also, Weber et al. found that extended sonication of PrPSc was associated with a loss of infectivity as measured by the prolongation of the incubation time [39]. In this case some PrPSc may be degraded during sonication. In our case, total PrPSc is isolated first, when sonication occurred, and then both fractions are separated at an intermediate centrifugation speed with no further sonication. On the other hand, hamster 263K ( = Sc237) prions did not show altered incubation times when the prion rods were fragmented by sonication into spherical particles [40].
Several explanations may account for the similarity in infectivity values in sPrPSc and rPrPSc. Namely, common and highly infectious species are present in both fractions and thus are subject to the same rate of clearance. Another possibility is the presence of one or more cofactors that copurify with both fractions of PrPSc. PrP 27–30 rods that dissociated into much smaller detergent-lipid-protein complexes and liposomes led to a 100-fold increase in infectivity as measured by endpoint titration [41]. Perhaps the binding of PrPSc to the cell membrane increases the efficiency of disease transmission from cell to cell.
Limited proteolysis studies on the sPrPSc fraction of 263K and Dy indicate that, as in total PrPSc, the sensitive fraction is also composed of alternating PK-resistant stretches interspersed with PK-susceptible stretches. The relatively small variations in intensity of some of the PK-resistant bands detected by SDS-PAGE (Fig. 3b), could be interpreted as corresponding to differences in accessibility of the PK-resistant stretches involved, but the global emerging picture is one of a shared basic structural organization.
This result emphasizes the fact that the difference between sensitive and resistant PrPSc seems to be the previously described different size of the aggregates and not the conformation of their respective aggregates [15], [19]. Our data provide information on the N-terminal half of the protein only, therefore we cannot rule out differences in the C-terminal portion.
As we mentioned in the introduction, the sensitive fraction of 263K varies between 20–40% of the total PrPSc. Despite this variability, the amount of sensitive PrPSc obtained for 263K was always less than that for Dy. This result, previously reported by Safar et al. [14], would indicate that a specific ratio of sensitive to resistant exists for each strain. This author also reported that the incubation time of a given prion strain is directly proportional to the level of protease-sensitive PrPSc [14]. According to Deleault et al., variations in the sPrPSc∶rPrPSc ratio between different prion strains appear to be an incidental product rather than a strain property [42]. Nevertheless, we want to point out that for hamster-adapted prion diseases the PK-susceptibility seems to directly correlate with the proportion of sPrPSc present in the strain.
In summary, we have shown that sPrPSc is fully infectious, and that its infectivity can be reduced by approximately 99% by treatment with PK (50 µg/mL; 1 hr). sPrPSc appears to share the same basic architecture with rPrPSc, as judged by the generation of similar fragmentation patterns by limited proteolysis with PK as seen in the WB and MALDI analysis. This, together with the fact that sPrPSc induces a disease with signs and histopathological and lesion profiles that are essentially identical to those induced by total or rPrPSc, strongly suggests that sPrPSc differs from rPrPSc, as previously shown, just in its smaller size, i.e., being made up of fewer PrP units [15], [19]. On the other hand, the fact that different prion strains are characterized by different relative ratios of sPrPSc∶rPrPSc suggests that sPrPSc might play an important role in key properties of strains. Finally, sPrPSc, which is soluble in detergent-containing solutions, might prove to be useful in studies aimed at elucidating the structure of PrPSc.
Animal experiments were carried out in accordance with the recommendations contained in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. The procedures were governed by a protocol that was approved by the Institutional Animal Care and Use Committee of the United States Department of Agriculture, Agricultural Research Service, Albany, CA (Protocol Number: P-10-3). All surgery was performed under isoflurane anaesthesia, and all efforts were made to minimize the suffering of the animals. The small numbers of experiments carried out at the University of Santiago de Compostela were approved by the University Ethics Committee, in accordance with the European Union Council Directive 86/609/EEC.
PrPSc was isolated from brains of terminally ill Syrian hamsters infected intracranially (ic) with the 263K strain of scrapie using a slightly modified version of the procedure of Diringer et al. [43]. A cocktail of protease inhibitors (Complete, Roche Diagnostics, Penzberg, Germany) at a final concentration of 1× was used in all buffers throughout the procedure up to the penultimate pellet, as defined in the mentioned study [43]. The final pellet was resuspended in 20 mM Tris (pH 8.5) containing 1% sarkosyl and no protease inhibitors, at concentrations between 1–2 µg/µl. The stock suspension thus prepared was aliquoted and frozen until further use; its purity was assessed by SDS-PAGE with Coomassie blue staining, and estimated to be approximately 80%. sPrPSc was isolated from total PrPSc by ultracentrifugation at an intermediate speed [19]. A 50–150 µl portion of the PrPSc stock was homogenized by application of 3 sonication pulses of 1 s each, and spun in a Beckman TLX ultracentrifuge (Beckman, Fullerton, CA) using a TLA-120-1 rotor at 40,000 rpm (56,806 g) for 2 hours at 20°C. The supernatant was collected and the pellet was resuspended by sonication in a volume of 20 mM Tris (pH 8.5) containing 1% sarkosyl equivalent to that of the supernatant. Fractions of supernatant and pellet were treated with 50 µg/ml of PK at 37°C during 1 hour; the reaction was terminated with 2 mM Pefabloc (Fluka, St. Louis, MO), and a fraction subjected to SDS-PAGE and either stained with Coomassie brilliant blue or transferred to a PVDF membrane (Immobilon-P, Millipore, Billerica, MA, USA) and analyzed by Western blotting using mAb 3F4 (Dako, Glostrup, Denmark). The supernatant fraction is sPrPSc, and the pellet fraction, PK-resistant PrPSc (rPrPSc). The concentration of PrPSc was estimated in each of these samples and in brain homogenates. PrPSc was isolated by ultracentrifugation using the method of Bolton et al. [44], with slight modifications [20]. The resulting pellets were dissolved in 200 µL of 6 M guanidinium chloride and allowed to stand at RT for 24 hours to inactivate the PrPSc [45]. The inactivated prion solutions were precipitated with methanol and subjected to analysis by mass spectrometry using previously described methods [20], [21]. Briefly, pellets were dissolved in 0.01% aqueous beta-octylglucopyranoside (BOG) and [13C5,15N]-VVEQMCTTQYQK added as internal standard. The samples were then sequentially reduced, alkylated and digested with trypsin. NanoLC/MS/MS was carried out with an Applied Biosystems (ABI/MDS Sciex, Toronto, Canada) model 4000 Q-Trap instrument equipped with a nano-electrospray ionization source. The mass spectrometer was operated in multiple reaction monitoring (MRM) mode, alternating between detection of VVEQMCTTQYQK (precursor ion m/z of 757.8, product ion of m/z 171.1) and [13C5,15N]-VVEQMCTTQYQK (precursor ion m/z of 760.8, product ion of m/z 177.1). Quantitation was done with the Intelliquan quantitation algorithm of Analyst 1.4.1 software (Applied Biosystems) using default parameters [20], [21].
Four prion isolates, 0.1% brain homogenate, purified PrPSc, sPrPSc, and rPrPSc, were diluted to yield four corresponding solutions containing 1.5±0.4 ng of PrP per 50 µL of sample, except in the case of the 0.1% brain homogenate, which contained 0.7 ng of PrP per 50 µL of sample (vide supra). A portion of each of these normalized solutions was diluted 100-fold, 1,000-fold and 10,000-fold into sterile phosphate buffered saline (PBS) to yield a set of four dilutions (including the undiluted material) for each of the prion isolates; in the case of the 0.1% brain homogenate sample, dilutions were 10-fold, 100-fold, and 1,000-fold. Fifty microliters of each dilution were ic inoculated into the right cerebral hemisphere of a four week old anesthetized female hamster. Eight animals were used per isolate/dilution.
The 128 animals were placed in cages (2 per cage) and observed for clinical signs. The first clinical sign to be observed and its date of occurrence recorded was an exaggerated startle response. A progression of clinical signs followed, including an exaggerated startle response that became more pronounced over time, ataxic gait whose severity increased with time, limb rigidity, characteristic head bobbing, and progressive lethargy. When the animals could no longer be roused from their recumbency to feed and drink water, they were humanely euthanized. The incubation time was measured (in days) as the time between inoculation and euthanization.
After 240 days one of the inoculated animals (sPrPSc; 10,000-fold dilution) animals remained healthy. The animal was euthanized and its brain removed. The brain was analyzed for the presence of PrP 27–30 by mass spectrometry [20]. There was no evidence of PrP 27–30 present in the brain.
In a separate experiment, two groups of 6 animals were inoculated ic with either 50 µl of a sample containing ∼40 ng of sPrPSc or the same volume of sPrPSc treated with 50 µg/ml of PK for 1 h at 37°C, followed by quenching of PK with 2 mM Pefabloc. Infection progression was monitored as described above, and approximate infectivity titers of the PK-treated and untreated sPrPSc samples were estimated by incubation period assay according to the procedure of Prusiner et al. [22].
Finally, two groups of 6 animals were intraperitoneally (ip) inoculated with 150 µl volumes of sPrPSc and rPrPSc prion isolates containing ∼125 ng of PrP each, and the course of infection observed as described above. The amount of PrPSc in these two samples was calculated approximately by comparison with a recombinant Syrian hamster (rSha) PrP (90–231) standard (a generous gift of Giuseppe Legname) after deglycosylation with PNGase F (New England Biolabs, Ipswich, MA, USA) and Western blotting with mAb 3F4.
Four µm thick sections were cut onto positively charged silanized glass slides and stained with hematoxylin and eosin, or immunostained using an antibody for PrP (3F4). Immunohistochemical stains for PrP were performed entirely on an automated Discovery XT staining apparatus (Ventana Medical Systems, Oro Valley, AZ) using a DAB Map XT Detection Kit (Ventana Medical Systems). Sections were deparaffinised and then washed in distilled water for 5 min. Epitope exposure was performed by heating sections to 100°C in a citrate buffer for 12 minutes. After treatment with protease 2 (Ventana Medical Systems) for 24 minutes, sections were incubated with anti-PrP 3F4 for 60 min and PrP was detected using HRP-conjugated antibodies and a DAB substrate.
We selected 6 anatomic brain regions in accordance with previous strain typing protocols [46], [47] from 2 hamsters per group. We scored spongiosis on a scale of 0–4 (not detectable, mild, moderate, severe and status spongiosus) and PrP immunological reactivity on a 0–3 scale (not detectable, mild, moderate, severe). A sum of the two scores resulted in the value obtained for the lesion profile for the individual animal. The ‘radar plots’ depict the scores for spongiform changes and PrP deposition. Numbers correspond to the following brain regions: (1) cerebellum, (2) medial thalamus, (3) hippocampus, (4) medial cerebral cortex dorsal to hippocampus, (5) medial cerebral cortex dorsal to septum, (6) white matter at cerebellar peduncles. Investigators blinded to animal identification performed histological analyses.
PrPSc was detected in brain homogenates of inoculated animals by Western blot using mAb 3F4 (primary antibody) and goat anti-mouse Fc (secondary antibody) (Sigma-Aldrich, St. Louis, MO).
To assess the relative amount of sPrPSc and rPrPSc in animals inoculated with the four different inocula described, duplicate PrPSc samples were isolated from the brains of inoculated hamsters using the method of Bolton et al. [44] with slight modifications [20]. The resulting pellets were either dissolved in 200 µL of 6 M guanidinium chloride (Sigma-Aldrich, St. Louis, MO) or resuspended in 0.1% Z 3,14-T-8.5 (0.1% 3-(N,N-dimethylmyristyl-ammonium)propane sulfonate; 20 mM Tris pH 8.5). The 6 M guanidinium chloride solutions were allowed to stand for 24 hours at RT to inactivate the PrPSc [45]. The samples suspended in 0.1% Z 3,14-T-8.5 were sonicated. After sonication, PK was added to make a final concentration of 5 µg/mL. The PK was permitted to react for 1 hour at 37°C. The reaction was quenched by addition of a sufficient amount of phenylmethylsulfonyl fluoride (PMSF) (Sigma-Aldrich, St. Louis, MO) to achieve a final concentration of 1 mM. Enough solid guanidinium chloride was then added to make a 6 M solution, which was left to stand for 24 hours at RT to inactivate the prions. The inactivated prion solutions were precipitated with methanol and the amount of PrP present in the sample was determined by mass spectrometry (vide supra) [20], [21].
All buffers and reagents were prepared fresh for the 2-nitro-5-thiocyanatobenzoic acid (NTCB) reactions. Between 15–20 µg of purified PrPSc was reduced in 2 mM DTT for 1 h at 37°C under denaturing conditions (6 M guanidinium chloride and 100 mM Tris buffer pH 8). In the same buffer, samples were cyanylated in 10 mM 2-nitro-5-(thiocyanato)-benzoate (NTCB) for 30 minutes at RT. After methanol precipitation, samples were redissolved in 6 M guanidinium chloride, 100 mM Tris buffer pH 8. PrPSc, specifically cyanylated at the reactive thiols, was cleaved by alkaline hydrolysis with 150 mM NaOH for 15 minutes at 37°C. The NTCB reaction was terminated by addition of trifluoroacetic acid (TFA) to a final pH of 2–3. PrPSc peptides were isolated from the reaction mixture using C-18 ZipTips (Millipore) according to the manufacturer's instructions. Peptides were eluted from the tips in 10 µl of 50% acetonitrile, 0.1% TFA and used directly for MALDI analysis.
A 2 µL portion of the PrPSc peptide sample was mixed with an equal volume of a saturated solution of sinapinic acid in acetonitrile and 0.1% aqueous TFA (1∶2). One microliter of the mixture was spotted onto a Bruker sample plate, allowed to air-dry, and analyzed using a Bruker Autoflex MALDI instrument in linear mode. The laser frequency was 25 Hz; pulsed ion extraction was set at a value of 140–150 ns. Repeated laser shots, typically 25–30, were averaged.
PK-treated samples were deglycosylated with PNGase F (New England Biolabs, Ipswich, MA, USA) at 37°C for 48 h, according to the manufacturer's instructions, followed by precipitation with ice-cold 85% MetOH. Pellets were boiled in 10 µl of reducing tricine sample buffer (BioRad, Hercules, CA, USA). Tricine SDS-PAGE [48] was carried out using 10–20% Tris-Tricine/Peptide Precast gels (BioRad), in a Criterion electrophoresis system (BioRad). The cathode buffer was Tris-Tricine-SDS buffer (Sigma-Aldrich) and the anode buffer, 100 mM Tris-HCl, pH 8.9. Electrophoresis was performed at a constant voltage of 125 volts for 200 minutes, on ice. After electroblotting on PVDF membranes, these were probed with antibody R1 (a generous gift from Hanna Serban, UCSF), which recognizes residues 226–231. Peroxidase-labeled anti-human antibody was used as the secondary antibody.
|
10.1371/journal.pcbi.1003552 | Coupling of Lever Arm Swing and Biased Brownian Motion in Actomyosin | An important unresolved problem associated with actomyosin motors is the role of Brownian motion in the process of force generation. On the basis of structural observations of myosins and actins, the widely held lever-arm hypothesis has been proposed, in which proteins are assumed to show sequential structural changes among observed and hypothesized structures to exert mechanical force. An alternative hypothesis, the Brownian motion hypothesis, has been supported by single-molecule experiments and emphasizes more on the roles of fluctuating protein movement. In this study, we address the long-standing controversy between the lever-arm hypothesis and the Brownian motion hypothesis through in silico observations of an actomyosin system. We study a system composed of myosin II and actin filament by calculating free-energy landscapes of actin-myosin interactions using the molecular dynamics method and by simulating transitions among dynamically changing free-energy landscapes using the Monte Carlo method. The results obtained by this combined multi-scale calculation show that myosin with inorganic phosphate (Pi) and ADP weakly binds to actin and that after releasing Pi and ADP, myosin moves along the actin filament toward the strong-binding site by exhibiting the biased Brownian motion, a behavior consistent with the observed single-molecular behavior of myosin. Conformational flexibility of loops at the actin-interface of myosin and the N-terminus of actin subunit is necessary for the distinct bias in the Brownian motion. Both the 5.5–11 nm displacement due to the biased Brownian motion and the 3–5 nm displacement due to lever-arm swing contribute to the net displacement of myosin. The calculated results further suggest that the recovery stroke of the lever arm plays an important role in enhancing the displacement of myosin through multiple cycles of ATP hydrolysis, suggesting a unified movement mechanism for various members of the myosin family.
| Myosin II is a molecular motor that is fueled by ATP hydrolysis and generates mechanical force by interacting with actin filament. Comparison among various myosin structures obtained by X-ray and electron microscope analyses has led to the hypothesis that structural change of myosin in ATP hydrolysis cycle is the driving mechanism of force generation. However, single-molecule experiments have suggested an alternative mechanism in which myosin moves stochastically in a biased direction along actin filament. Computer simulation serves as a platform for assessing these hypotheses by revealing the prominent features of the dynamically changing landscape of actin-myosin interaction. The calculated results show that myosin binds to actin at different locations of actin filament in the weak- and strong-binding states and that the free energy has a global gradient from the weak-binding site to the strong-binding site. Myosin relaxing into the strong-binding state therefore necessarily shows the biased Brownian motion toward the strong-binding site. Lever-arm swing is induced during this relaxation process; therefore, lever-arm swing and the biased Brownian motion are coupled to contribute to the net displacement of myosin. This coupling should affect the dynamical behaviors of muscle and cardiac systems.
| Myosin II, the conventional myosin responsible for muscle contraction, generates mechanical force by interacting with actin filament. Our understanding of this actomyosin motor has greatly increased by X-ray analyses of myosin structures [1]–[3] and by electron microscopy (EM) of actomyosin complex [4]–[7]. These structural observations have led to the widely held lever-arm hypothesis [2], [3], in which the change in the nucleotide state in the myosin head is amplified through allosteric communication for rotating the lever-arm region of myosin to exert mechanical force. X-ray and EM data of static protein structures do not, however, provide direct information on how the motor works dynamically. Dynamical behaviors have been observed in single-molecule experiments (SMEs) [8]–[14], among which the Yanagida group [11], [13] analyzed the fluctuating motion of a single subfragment-1 (S1) of myosin and supported the alternative Brownian-motion hypothesis [15]. In this hypothesis, the myosin head stochastically moves along the actin filament with a regular step size of 5.5 nm, which corresponds to the diameter of actin subunit, in both directions toward the plus and minus ends of the actin filament during a single cycle of ATP hydrolysis. In this stochastic walk or effective Brownian motion, the frequency of steps toward the plus end is considerably higher than that of steps to the minus end. This biased Brownian motion enables the search for a stable binding site on the filament, which pulls the filament to exert mechanical force [11], [13].
The thermal Brownian fluctuation of the myosin molecule should also cause the stochastic fluctuation in the direction of lever-arm swing. Even with such Brownian fluctuation of conformation, the lever-arm hypothesis implies that the net displacement of myosin is limited by the allowed angular range of the lever arm. In contrast to this narrow distribution, the net displacement of myosin stochastically varies under the Brownian-motion hypothesis and its distribution is broad and changes flexibly depending on the load applied to the system. These two hypotheses should accordingly show a clear difference in predicting the flexibility and load dependence of the system [16]. In addition, for myosin V, a non-conventional myosin responsible for vesicle transport, SME measurements [17]–[19] have clearly shown that the Brownian motion of the leading head of myosin in searching for the binding location on the actin filament significantly contributes to force generation together with the lever-arm pushing mechanism at the trailing head of myosin. A key issue in understanding the mechanism of actomyosin motors is thus to clarify how and to what extent lever-arm swing and Brownian motion contribute to force generation [16], [20], [21]. In this study, we address this problem by in silico observations of the system composed of a single head (S1) of myosin II and an actin filament.
Analyses of the kinetic cycle of interactions between myosin II and actin filament [22] should help to resolve this problem:(1)Myosin (M) strongly binds to actin filament (A) when no nucleotide is bound to myosin to form the rigor state (A.M). When ATP binds to myosin (A.M.ATP), the myosin detaches from the actin filament (M.ATP). After the bound ATP is hydrolyzed into ADP and Pi , the complex M.ADP.Pi binds to actin to form the weakly bound state (A.M.ADP.Pi), which is transformed to the strongly bound state by the release of Pi (A.M.ADP) and ADP to reach the rigor state again. From observed structures of myosin with various nucleotide analogs [2], [3], [23], it is plausible to assume that the lever arm of myosin in M.ATP and M.ADP.Pi is in the pre-stroke position and the lever arm in other states is in the post-stroke position; therefore, processes 4 and 2 in Eq.1 should correspond to lever-arm stroke during force generation and the recovery stroke, respectively.
Further detailed comparison among kinetic states and structures, however, has raised a question regarding the application of the lever-arm hypothesis [3]. From various observed myosin structures, it is noted that the opening/closure of the nucleotide binding pocket, the lever-arm positioning, and the closure/opening of the 50 kDa cleft of myosin are correlated with one another [23], [24] (See Fig. 1 for an example structure of S1 of myosin II). The resolved structures have shown that Pi in M.ADP.Pi makes the nucleotide binding pocket closed, which tends to maintain the lever-arm in the pre-stroke position and the 50 kDa cleft open. Given that the closure of the 50 kDa cleft has been reported to be necessary for the strong binding of myosin to actin [4]–[7], it is reasonable to assume that M.ADP.Pi weakly binds to actin. The weak binding of M.ADP.Pi to actin has been suggested by kinetic [25]–[27] and structural [28]–[30] measurements. However, for myosin to exert a force using the lever-arm mechanism, myosin must strongly bind to actin before the occurrence of the lever-arm swing. This problem in applying the lever-arm hypothesis may be solved if it is assumed that the 50 kDa cleft of A.M.ADP.Pi is closed, although the pre-stroke open-cleft structure is stable in M.ADP.Pi [31]. If myosin adopts the pre-stroke closed-cleft structure, it should strongly bind to the actin filament, and the subsequent occurrence of the lever-arm swing on the release of Pi should generate mechanical force. The pre-stroke closed-cleft structure may be possible when this structure is stabilized by specific myosin-actin interactions. Although considerable effort has been devoted to detecting the pre-stroke closed-cleft structure [32], there is no direct evidence for its existence thus far [33].
In this study, we develop a theory on kinetic process, which is a dynamical energy landscape theory of actomyosin, without relying on the assumption of a stable pre-stroke closed-cleft structure of myosin. We assume that myosin interacting with actin tends to adopt one of structures observed in previous experiments. We also assume that the structure of myosin with a given nucleotide-binding state shows fluctuating transitions among these conformations, such as are shown by many allosteric proteins in the population-shift or conformation-selection mechanism of allostery [34], [35].
In our previous studies, the theoretical models of movement of myosin S1 were discussed [36], [37]. Molecular dynamics simulation was performed to investigate myosin with the nucleotide-free post-stroke closed-cleft structure [37] and it was shown that the electrostatic interactions at the actin-myosin interface should lead to a globally biased energy landscape of myosin movement toward the strong-binding site on the actin filament and that the stochastic movement of weakly binding myosin follows the gradient of this landscape in the course of relaxation from weak- to strong-binding states; therefore, the relaxation process reproduces the biased Brownian motion observed in SMEs [11], [13]. However, to investigate the roles of this simulated behavior in the kinetic cycle of Eq.1, as noted in [37], we need to extend this method to cases in which the energy landscape is not fixed, but is dynamically changing according to changes in nucleotide state and conformation.
In dynamical energy landscape theory, multiple kinetic states, corresponding to different stages of chemical reactions or other different conditions, are considered and the dynamical switching among landscapes in these states is analyzed [38]–[45]. Here we consider the multiple kinetic states appearing in the course of force generation, called “actomyosin states.” Figure 2 shows the kinetic network among actomyosin states considered in this study. Actomyosin states shown in Fig. 2 are defined by both the conformation and nucleotide state of myosin. We assume that myosin in actomyosin states tends to adopt conformations similar to those observed in X-ray or EM data. Myosin in A.M.ADP.Pi should adopt the pre-stroke open-cleft conformation (Mpre) that is modeled by the X-ray structure of myosin with an ADP.Pi analog, and myosin in A.M.ADP should adopt the post-stroke open-cleft conformation in the X-ray data (Mpost). Myosin in A.M should adopt the post-stroke closed-cleft conformation (Mclosed) obtained by fitting the EM image in the rigor state. See the Methods section for more details on the definitions of these model conformations. In the present study, we distinguish the weakly bound Mclosed from the strongly bound Mrigor in the rigor state. Although both Mclosed and Mrigor have a post-stroke closed-cleft conformation, water molecules that hydrate myosin should be expelled from the interface with actin in transition to the rigor state, which is expressed in the model by a transition from Mclosed to Mrigor.
It should be noted that in A.M.ADP in the absence of Pi, switch-I and switch-II regions of myosin are not bound to the ligand, and therefore, the post-stroke position of the lever arm and the closed 50 kDa cleft are expected to be energetically stable. The open 50 kDa cleft structure, however, should be entropically favorable to form both the post-stroke open-cleft structure and post-stroke closed-cleft structure in A.M.ADP. Therefore, we consider that A.M.ADP fluctuates between A.Mpost.ADP and A.Mclosed.ADP. Though the post-stroke open-cleft structure Mpost has been often referred to as the “near-rigor” or “post-rigor” conformation that appears after leaving the rigor state [46], the post-stroke open-cleft structure is a representative structure of the ADP-bound myosins, and there is no evidence against the appearance of this structure before the rigor state is reached. Therefore, we use Mpost as a structure expected in A.M.ADP. Similarly, we consider that both Mpost and Mclosed appear in A.M. The electron paramagnetic resonance data have shown that coupling between the nucleotide state and conformation is not rigid [47]. We, therefore, assume that myosin with a given nucleotide state can adopt conformations that are expected to appear in the next or in the previous step of ATP hydrolysis as pre-existing or post-existing conformations in the conformation-selection mechanism of allostery. We consider A.Mpost.ADP.Pi to be the pre-existing conformation (the conformation expected to be found in the ADP bound state). In the ADP-bound state, we consider A.Mpre.ADP to be the post-existing conformation (the conformation expected in the ADP.Pi bound state). The rigor state is reached through the conformation-selection mechanism by selecting the pre-existing Mclosed conformation in the weakly bound state. It is assumed that the concentration of ADP or Pi in solution is so low that reverse reactions in steps of ATP hydrolysis are negligible. Thus, we have the network of transitions as shown in Fig. 2. We also assume that the strongly bound state A.Mrigor is stable, and hence, we do not consider the spontaneous loosening of binding from A.Mrigor to A.Mclosed.
For individual actomyosin states, we calculate the free-energy landscape which determines the movement of the myosin head in each of these states. Free-energy landscapes of myosin movement and actin-myosin binding are derived using a coarse-grained model of actomyosin, which represents proteins as chains connecting beads of carbons (s). Forces acting among s of myosin are derived from the Gō-like potential [48], [49], which stabilizes the model myosin structure, Mpre, Mpost or Mclosed. Nucleotide and Mg ion bound to myosin are represented as particles of all nonhydrogen atoms. In this way, different actomyosin states are represented using different Gō-like potentials and different models of nucleotide and Mg. The potential consistently used among actomyosin states is the Gō-like potential for actin, which stabilizes the EM structure of actin filament [50]. As inter-protein interactions, we introduce electrostatic interactions, which are represented by Debye-Hückel potentials, and van der Waals interactions, which are represented by the Lennard-Jones type potentials. Using these potentials, we perform the Langevin molecular dynamics simulation. The setup of the simulation is shown in Fig. 3. An S1 domain of myosin II, comprising a heavy chain, an essential light chain (ELC), and a regulatory light chain (RLC), is placed on the actin filament, which extends along the -axis with its plus-end facing the positive direction. The angle around the -axis is denoted by . The actin filament is connected to the spatially fixed points by springs. By mimicking the setup of the SME [11], the tip of the myosin lever-arm is connected by springs to a line running parallel to the actin filament. Myosin can move freely along this line without any bias either toward the or direction. By monitoring the position of the center of mass of the myosin motor domain (MD) during simulations, we calculate the free-energy landscape in the two-dimensional space of and using the weighted histogram analysis method (WHAM) [51] with umbrella potentials. See the Methods section for the simulation details.
Using the free-energy landscapes thus calculated, the movement of myosin II on the surface of an actin filament is simulated by the stochastic motion of a point at the center of mass of the myosin motor domain. Motion of this point along the free-energy landscape of each state is simulated using the value of the free energy in the Metropolis algorithm. The transition between different actomyosin states because of nucleotide-state change or lever-arm swing is simulated by dynamical switching between free-energy landscapes. Therefore, the point representing the position of myosin moves along the calculated free-energy landscapes and stochastically jumps among them. In this way, we shed light on roles of both the lever-arm swing, occurring during transitions among landscapes, and the biased Brownian motion along individual landscapes.
In Fig. 4, free-energy landscapes of actin-myosin interaction in states A.Mpre.ADP.Pi , A.Mpost.ADP, and A.Mclosed are shown as functions of . In addition, the one-dimensional free-energy landscapes obtained by projecting the two-dimensional landscapes onto the -axis are shown. The calculated free-energy landscapes are almost periodic in the direction because of the helical nature of the EM structure of the actin filament with approximate helical pitch nm. In states with the open 50 kDa cleft structure, A. Mpre.ADP.Pi (top, Fig. 4) and A.Mpost.ADP (middle, Fig. 4), the landscapes have multiple basins located at an interval of 5.5 nm, corresponding to the diameter of the actin subunit. These basins are separated by the low free energy barrier of 1–2 , which should be easily overcome by thermal noise. The lowest free-energy minima on the landscape of A..ADP.Pi and A.Mpost.ADP are positioned at nm and nm, respectively, as shown in Fig. 4.
A large difference from the above two landscapes is found in the landscape of the A.Mclosed state with a closed 50 kDa cleft structure (bottom, Fig. 4). The landscape has an array of basins at positions separated by the size of actin subunit, 5.5 nm, with a global gradient toward the strong binding site at . This prominent feature can be ascribed to complementary matching between the closed-cleft structure of myosin and the actin filament with a heterogeneous distribution of electric charges on its surface. The shear motion between upper and lower 50 kDa subdomains should also contribute to the complementary matching between myosin and actin in A.Mclosed [52]. The arrangement of valleys in the direction is also notable. In A. Mclosed the angle difference between adjacent basins is considerably smaller than the angle expected from the helical structure of the filament, . This narrow distribution of basins results from the interplay among the myosin-actin interactions and the restraints on the motion of myosin and actin. The disagreement between the actin-subunit arrangement and the basin distribution indicates that myosin binds with different orientations to the actin surface in different basins, a difference that should lead to the difference in free energy among these basins. Thus, the strong gradient of the free-energy landscape is coupled with a narrow distribution of basins in the landscape.
The one-dimensional free-energy landscapes in seven states in Fig. 2 are compared in Fig. 5. It is found that the difference in conformation more significantly affects the free-energy landscape than the difference in the nucleotide state. The corresponding two-dimensional landscapes are shown in Fig. S1. We do not consider myosin movement in A.Mrigor, and therefore, the calculation of free-energy landscape in the A.Mrigor state is omitted.
From Figs. 4 and 5, we can deduce the behavior of myosin through kinetic transitions in Fig. 2. A myosin head landing on the actin filament should be attracted to the valley in the free-energy landscape of A.Mpre.ADP.Pi. It should weakly bind there and widely fluctuate among multiple basins of landscapes in A.Mpre.ADP.Pi or A.Mpost.ADP. The most populated - region of the myosin head in A.Mpre.ADP.Pi , A.Mpost.ADP, or other states is the region of high free energy in A.Mclosed.ADP and A.Mclosed landscapes. Thus, after releasing Pi or ADP.Pi , the myosin begins to relax to the more stable low free-energy position at the larger by moving along the actin filament. This movement associates jumps among minima with a regular spacing of approximately 5.5 nm.
The above scenario of myosin movement can be verified by Monte Carlo (MC) simulation. The diffusive motion of the myosin head is simulated by the motion of a point representing the position of the center of mass of the myosin motor domain on the calculated two-dimensional free-energy landscape using the Metropolis algorithm. The trial movement of a point is generated as a step on the lattice with mesh size , where nm and . This trial is accepted when the free-energy change induced by the trial movement is . When , the trial is accepted with probability and rejected with probability . A similar method was used to simulate the movement of kinesin head along the surface of a microtubule [39].
We extend this method by applying it to the problem of multiple landscapes. A point representing the center of mass of the myosin motor domain diffuses along a landscape and jumps from on one landscape to the same on the other landscape with probabilities defined by rates in Fig. 2; with and with . Values of with and with represent chemical reactions and large-scale conformational change, respectively, which should have a 1–10 ms timescale. As discussed in the Methods, Monte Carlo steps (MCS) should correspond to several ms or longer, and hence, with and with should be –. For simplicity, we use either of two values, or . Given that Mpre.ADP.Pi, Mpost.ADP, and Mclosed have been observed in the X-ray and EM analyses, the actomyosin states A.Mpre.ADP.Pi, A.Mpost.ADP, and A.Mclosed should be relatively stable. In the following, values of and are chosen to stabilize the A.Mpre.ADP.Pi state as for , , , , and , and for , , , , and . is the rate of the process of hydrophobic matching between surfaces of proteins and should be faster than the large conformational change of proteins. We accordingly use the value . The calculated results are robust against changes in this parametrization. See Fig. S2 for the results of other choices of values for and .
After the transition from one landscape to the other, the point representing the center of mass of the myosin motor-domain continues to diffuse on the new landscape. Such successive transitions and diffusions are terminated when the trajectory reaches the A.Mrigor state. We assume that this termination is the transition from the lowest free-energy valley of the landscape of A.Mclosed at to A.Mrigor with the rate . See the Methods for more details on the MC simulation. We should note that this MC calculation is based on the approximation that processes occurring during the transition between states, namely the lever-arm swing or chemical reactions, can be decoupled from the motions of actin and myosin within each state. This decoupling should be validated when we can assume separation of timescales among the process between states and motions within states. To evaluate the validity of this assumption, simulations of the coupled processes of transition, conformational fluctuation, and diffusive motion are necessary. A more elaborate molecular dynamics model that allows the examination of such dynamic coupling among processes is being developed [53], and we leave the application of that model to the motor problems as a future project.
Along the MC trajectory, myosin that has begun to interact with the actin filament at an arbitrary position is attracted and weakly bound to the free-energy valley in the A.M.ADP.Pi state, but the position of the myosin largely fluctuates along the -axis while it stays in the weak-binding state. After reaching the A.Mclosed.ADP or A.Mclosed state, the myosin begins to show the biased Brownian motion. Because the closure of the 50 kDa cleft should promote the release of ADP from myosin, we assume that the lifetime of A.Mclosed.ADP is short, and thus, the persistent motion appears in the A.Mclosed state. In the A.Mclosed state, Brownian motion is composed of steps with a regular width of 5.5 nm, and shows both the forward and backward stepping, but is biased toward the forward direction (Fig. 6A). This biased Brownian motion is terminated when myosin reaches the rigor state.
The distribution of myosin displacement was monitored when the large positional fluctuation in the weak-binding state was reduced on the start of the displacement [11]. To compare this measurement, we monitored the simulated displacement after the system enters into the A.Mclosed state by calculating , where and are the position of the myosin motor domain in the rigor state and that at the time when the system enters the A.Mclosed state, respectively. in distinguishes the different strong-binding sites which are almost periodically positioned along the helical actin filament. Shown in Fig. 6B is the calculated distribution of , which consists of two parts, i.e., the major and minor parts. The major part is biased toward with multiple peaks separated by 5.5 nm. The major part of the distribution represents the trajectories that reach . The minor part is the distribution of trajectories that reach the strong-binding site at the periodic location .
In previous SMEs [11], [13], the displacement of S1 has been monitored at RLC near the tip of the lever arm. As will be discussed in the subsection Contribution of the lever-arm swing, the most probable value of the -coordinate near the tip of the lever arm, , is approximately 5–7 nm greater than the -coordinate of the center of mass of myosin motor domain, , in A.Mclosed. Therefore, for comparison with the distribution of displacement of the lever-arm tip, the distribution of Fig. 6B should be shifted by several nanometers in the positive direction. With this correction, the simulated distribution of Fig. 6B reproduces the results of SME of Fig. 5 in [13]. The distribution of Fig. 6B is also consistent with the SME reported earlier [9] although the data has been differently interpreted [9] by disregarding the minor part of the observed distribution. The distributions of simulated with different parameterizations of kinetic rates are compared in Fig. S2, showing that the results are insensitive to differences in these parameters.
The lever-arm swing upon the kinetic transitions in the present scheme also contributes to force generation by displacing the lever-arm tip. Fig. 7 shows the position of the center of mass of the myosin motor domain and the position in RLC near the lever-arm tip represented by (Fig. 7A), and the free-energy landscapes drawn on the plane of in the A.Mpre.ADP.Pi state (Fig. 7B, left) and in the A.Mclosed state (Fig. 7B, right). In A.Mpre.ADP.Pi , myosin only weakly binds to actin to make the free energy insensitive to the angle of myosin to the actin surface. The free-energy basin accordingly spreads in the direction of the axis. In A.Mclosed, in contrast, the free-energy basin is localized at locations with 5–7 nm reflecting the post-stroke position of the lever-arm tip. From Fig. 7B, from estimated difference in location of free-energy basins in two landscapes, the net displacement of the lever-arm tip (A.Mclosed)(A.Mpre .ADP.Pi) is 10–16 nm, in which the contribution of the biased Brownian motion (A.Mclosed) (A.Mpre .ADP.Pi) is 5.5–11 nm and the contribution of the lever-arm swing is 3–5 nm. The finite width of the distribution of is noteworthy because of the stochastic nature of the diffusive motion, and some width of the contribution of the lever-arm swing due to the fluctuating position in the A.Mpre .ADP.Pi state should also be noted. Although the simultaneous measurement of and in SME has not yet been reported, it is important to acquire high resolution data for and to check the validity of the discussed mechanism.
As shown in the previous subsections, the biased Brownian motion of myosin arises from the global gradient of the free-energy landscape of A.Mclosed. Two crucial factors involved in this gradient are (i) the close contact of myosin and actin surfaces, which is allowed to occur only when the 50 kDa cleft of myosin is closed, and (ii) the attractive electrostatic interactions through the contact between myosin and actin [37]. In the following, we show that this contact is formed through the conformational flexibility of actin and myosin.
Some regions of heavy and light chains of myosin are structurally disordered and not determined by X-ray analysis. These disordered regions are spread over N-terminal region (residue number, 1–3), loop 1 (residue number, 205–215), loop 2 (residue number, 627–646), loop 3 (residue number, 572–574), converter (residue number, 732–737), and several regions in the ELC and RLC. In addition, the structure of the N-terminus of actin subunit is inconsistent between X-ray crystallography [54] and EM [50], indicating that this part is also disordered in solution. The importance of loop 2, loop 3, and the N-terminus of actin subunit to actin-myosin binding was shown in our previous molecular dynamics simulation [55]. In Fig. 8A four landscapes are compared with different degrees of allowed fluctuations in these regions. In one landscape, all of the disordered regions fluctuate without the guidance of the Gō-like potential, whereas in the other landscapes, the Gō-like potentials stabilizing the reference structures are assumed to regulate the fluctuation of these regions. The fluctuation of the myosin-actin surface is indispensable for the generation of the global gradient of the landscape (Fig. 8A). When loop 2 and loop 3 structures of myosin are more rigid in the simulation, the global gradient of the landscape considerably decreases, a consequence that should diminish the bias in the Brownian motion. We also find that the flexibility of the N-terminus of actin subunit enhances the biased Brownian motion; with the less flexible N-terminus of actin subunit, the barrier between the minima becomes higher, to allow the rigid N-terminus works to hinder for the diffusive motion of myosin. It would be interesting to investigate these theoretical predictions by observing the movement of myosin by following the introduction of mutations that rigidify the structure of myosin loop regions or of the N-terminus of actin.
The role of electrostatic interactions was investigated in our previous studies [37], [55] by changing the concentration of ions in solution and introducing mutations to change the charge distribution in the model. In addition, in the present simulation, the global gradient in the free-energy landscape decreases by increasing the concentration of counter ions, a result consistent with the experimentally observed decrease in the efficiency of the actomyosin motor in an in vitro motility assay [56] (Fig. 8B). In this study, the simulated results predict that this decrease in motor efficiency should be observed not only for the ensemble of actomyosins but also for SME.
Two major assumptions in the present study are that (i) myosin tends to adopt the conformations determined by X-ray and EM observations, and (ii) myosin fluctuates among these conformations as allosteric proteins fluctuate in the conformation-selection mechanism of allostery. With these assumptions, we found that free-energy landscapes for myosin movement along actin are different in the weak-binding and strong-binding states, necessarily leading to a difference in the most stable positions for myosin, which we called the weak-binding and the strong-binding sites. This difference in binding location forces myosin to move from the weak-binding to the strong-binding sites, according to the kinetic change from weak-binding to strong-binding states. We found that the free-energy landscape of this movement has a global gradient from the weak-binding to strong-binding sites; therefore, myosin shows the biased Brownian motion toward the strong-binding site as has been observed in SME. We also found that this Brownian motion is concomitant with lever-arm swing; therefore, the nm displacement due to the lever-arm swing and the nm displacement due to the biased Brownian motion are coupled with each other to contribute to the net displacement of myosin. The simulated biased Brownian motion explains the SME data, and the theoretical results predicted that the biased Brownian motion and the underlying free-energy landscape are modified by mutagenesis to change the structural rigidity of myosin loop regions or the N-terminus of actin subunit. It is also predicted that the bias in the Brownian motion is weakened due to the increase in counter-ion concentration.
The displacement of S1 due to the combined effects of the biased Brownian motion and lever-arm swing arises from the dynamically changing free-energy landscape. This dynamical free-energy landscape also suggests an intriguing scenario on the mechanism by which myosin binds to actin when the S1 domain is connected to the S2 domain and further to the light meromyosin (LMM) domain. In Fig. 9, the movement of myosin with S2 in a cycle of ATP hydrolysis is illustrated. Here, one of two myosin heads is shown to emphasize the movement in an ATP cycle. Myosin weakly binds to actin in the A.M.ADP.Pi state at (Fig. 9A), begins to move in the direction in the A.Mclosed state (Fig. 9B), reaches the strong-binding site at , and enters the strongly bound rigor state (Fig. 9C). When myosin binds ATP, myosin detaches from actin and the myosin motor domain changes its orientation through the recovery stroke (Fig. 9D). Here, we emphasize the positive role of the recovery stroke. Because the myosin S1 is connected to S2 and LMM, the recovery stroke of the lever arm in a detached state from the actin filament does not shift the position of S2 or LMM but rather shifts the orientation of the motor domain, as illustrated in Fig. 9D. After ATP hydrolysis, the myosin begins to bind to actin. Here, we can expect that the myosin head searches for the next binding site with the swinging motion of S1 and S2 domains (Fig. 9E). With this swinging search with the motor domain oriented as in Fig. 9D, the myosin should have higher binding affinity to the binding site in the next helical pitch rather than to the position in the previous ATP cycle (Fig. 9F). Subsequently, after the release of Pi and ADP, the myosin moves toward the strong-binding site . In this way, the recovery stroke should enhance the net displacement of the myosin head through the multiple cycles of ATP hydrolysis.
This suggested mechanism of myosin II movement is similar to that of processive motion of myosin V. The leading head of myosin V after ATP hydrolysis searches for a binding site to the actin filament through the swinging motion of the neck domain [17], [18]. Because the motor domain has its orientation changed by the recovery stroke in myosin V [57], the binding affinity of myosin to actin is enhanced in the forward more than that in the backward direction along the actin filament [57], [58]. The importance of the motor domain orientation on binding to the actin filament has also been suggested for myosin VI [59].
The present results showed that loop 2 of myosin is the key element in causing the biased Brownian motion. Given that loop 2 of myosin V is longer than loop 2 of myosin II, with a larger number of positive charges, the biased Brownian motion may also contribute to the processive motion of myosin V. As shown by a recent SME [19], the leading head of myosin V finds a position to bind to the actin filament through its random swinging motion, also called the Brownian search-and-catch motion. It will be interesting to investigate whether the leading head of myosin V searches for the strong-binding site through the biased Brownian motion, as discussed in the present study, at the final step of this Brownian search-and-catch process by moving along the actin filament.
The simulated result presented in this paper based on dynamical energy landscape theory is consistent with the observed structural features of myosin and actin, reproduces the SME data, predicts the effects of conformational flexibility and electrostatic interactions, and further suggests a unified mechanism for different members of myosin family. We found that the displacement of myosin head during a cycle of ATP hydrolysis is variable with respect to both the contribution of lever-arm swinging and the biased Brownian motion. These variances should be changed by the load applied to the myosin head in ways characteristic of these contributions. Such dynamically flexible responses should affect the dynamical behaviors of muscle and cardiac systems [60]. The comparisons among these system behaviors, SMEs, and in silico observations should open further avenues to understanding the dynamical physiological phenomena.
We constructed three structural models, Mpre, Mpost, and Mclosed of myosin S1, each of which comprises a heavy chain, an essential light chain and a regulatory light chain. We assumed that Mpre, Mpost, and Mclosed are structures of myosin obtained from chicken skeletal muscle in accordance with single-molecule experiments [11], [13].
Mpre is the pre-stroke open-cleft structure of myosin with the bound analog of ADP and Pi. Because the X-ray structure of myosin of chicken muscle with the analog of ADP and Pi is not yet available, we constructed Mpre using scallop myosin with ADP and VO4 (PDB code: 1QVI) [61] by homology modeling. Using the sequence alignment between sequences of myosin from chicken skeletal muscle and 1QVI and using the structure of 1QVI as a template, Mpre was computationally constructed with the software MODELLER [62]. A vanadium atom, V, was replaced with a phosphorus atom, P, to give Pi. Parts of the myosin structure that were missing because of disorder in the template structure 1QVI were treated as flexible parts in Mpre fluctuating without guidance of the Gō-like potential in the model. Mpost is the post-stroke open-cleft structure of myosin with the bound analog of ADP. Mpost was constructed using chicken skeletal myosin without nucleotide (PDB code: 2MYS) [1], and its binding with ADP was modelled using scallop myosin with ADP (PDB code: 2OTG) [63]. We treated the missing parts in 2MYS (chicken skeletal myosin without any nucleotide) as flexible parts in Mpost fluctuating without guidance of the Gō-like potential. Mclosed is the post-stroke closed-cleft structure extracted from the structure determined by fitting the electron-microscope image of actomyosin complex [7]. Because Mclosed appears in the course of relaxation from the weak to the strong actin-binding states in our simulation scheme, we assumed that structures of loops and other flexible regions of Mclosed are not fixed as in the rigor state. We therefore treated the missing parts in 2MYS as the flexible parts in Mclosed fluctuating without guidance of the Gō-like potential, unless otherwise noted.
We assumed that actin filament is obtained from rabbit skeletal muscle, in accordance with the single-molecule experiment [11]. A structural model of actin filament was represented as a complex of 26 subunits and was reconstructed from the X-ray structure (PDB code: 2ZWH) by positioning the adjacent subunit with 166.4 rotation and 2.759 nm translation [50]. The actin filament constructed in this way shows the 3.2 rotation for the translation of 13 subunits, amounting to 35.867 nm. We represented this structure as one exhibiting helical symmetry with approximate helical pitch 35.9 nm.
These structural models, Mpre, Mpost, , and the model of actin filament, were used as reference structures for the Gō-like potentials (see below).
Each polypeptide chain in the system was represented with residue-level coarse graining as a chain of connected beads of atoms. Bound ligands, Mg+ADP+Pi for the A.M.ADP.Pi states, Mg+ADP for the A.M.ADP states, were represented by all nonhydrogen atoms, whereas the nucleotide-free A.M states lacked bound ligand atoms.
The total potential energy of the actomyosin system, , is given by(2)where is the interaction potential within myosin (including the bound ligands) and within the actin filament, is the interaction potential between myosin and the actin filament, and is the restraint potential on the lever-arm tip of myosin and on the subunits of the actin filament.
As shown in the following, the reference structure defined above is the minimum-energy structure in the interaction potential given by(3)The bond-angle potential is(4)where the bond angle is defined as the angle formed by three successive residues , , and , and the superscript 0 hereafter denotes the values of variables in the reference structure. The dihedral angle potential is given by(5)where is defined as the dihedral angle formed by the four successive residues , , , and . With respect to the contact interactions, all residue pairs are classified as either native or nonnative using the reference structure; for residue pair and within the same chain, if at least one pair of nonhydrogen atoms are within 4.5 Å from each other in the reference structure with , the pair and is considered a native pair. Given that there are multiple subunits within a myosin or an actin filament, residue pairs between different subunits also interact with each other by the contact potential. For the pair and across different subunits in myosin or actin filament, if at least one pair of nonhydrogen atoms are within 4.5 Å from each other in the reference structure, the pair is considered a native pair. Otherwise, a pair within the same chain or a pair across different subunits is a nonnative pair. Contact potential is given by(6)and is given by(7)where(8) is given by(9)where(10)The constants , , , and were defined as 6.67 kcal/mol/rad2, 1.6710 kcal/mol, 8.3310 kcal/mol, 3.3310 kcal/mol and 1.33 kcal/mol/Å2, respectively. The cutoff distance was set to be 4.0 Å.
These definitions of intramyosin or intraactin potential, including the relative strengths of bond angle, dihedral and contact potentials, are similar to those of the Gō-like model [48], [49], except for the following modifications. Bond length between adjacent residues along the polypeptide chain was constrained to by the RATTLE algorithm [64], instead of the spring potential, to ensure the stability of the Langevin dynamics simulation. The contact potential at or was replaced by the spring-like potential to avoid instability in the numerical integration of the Langevin equation.
Ligand contact potential was given by the spring-like potential,(11)The pair of ligand atoms located within 4.5 Å from each other in the reference structure interact with each other by this potential. In addition, if at least one of the atoms in the amino acid residue and an atom in the ligand are located within 4.5 Å from each other in the reference structure, that residue also interacts with the ligand atom by this potential. was 6.67 kcal/mol/Å2.
The interaction at the interface between myosin and actin, , is similar to that in our previous study [37], and is composed of electrostatic and van der Waals interactions:(12)The electrostatic interactions were expressed by the Debye-Hückel potential as(13)where or is the charge of the amino acid residue ( for Asp and Glu, for Lys and Arg, and for His) or the charge of the atom in the ligand ( for each of the the three oxygen atoms in ADP, for each of three oxygen atoms in Pi, and for Mg). The parameters were defined as Å, kcal Å/mol, and = 59.3 Å. The van der Waals interactions were given by the 12-6 type Lennard-Jones potential(14)with(15)where the potential at was replaced by the spring-like potential. The parameters were = 0.015 kcal/mol, = 8.0 Å, and = 1.33 kcal/mol/Å2.
To mimic the experimental setup of the single-molecule experiment [11], we applied spatial restraints to myosin and the actin filament, respectively, as(16)The tip of the myosin lever-arm (residue number 830–843 in the heavy chain and residue number 1–83 in the regulatory light chain) was restrained with the curtain-rail potential(17)where was 0.2 kcal/mol/Å2. The -axis runs parallel to the center line of the reference structure of actin filament, and and are coordinates perpendicular to the -axis. We assumed that the curtain-rail runs helically around the actin filament so that the whole system has the same helical symmetry as the reference structure of the actin filament, a 3.2 rotation for each 35.867 nm. We therefore used(18)with . The rotation of 3.2 per 35.867 nm is so small that the helical arrangement of the curtain-rail is visually indistinguishable from the straight line along the -axis. However, this slightly helical arrangement aids rapid numerical convergence in WHAM by assuring the periodicity of the entire system. In accordance with the single-molecule measurement [11], no force is applied with respect to the movement of the lever-arm tip of myosin along the curtain-rail. All residues in the actin filament were restrained by the potential(19)where was defined as kcal/mol/Å2.
We performed the Langevin molecular dynamics simulation to sample the conformational ensemble of the actomyosin complex. The integration scheme is that of Honeycutt & Thirumalai [65], with the particle mass = 1.0, temperature = 300 K, time step = 0.0175, and friction coefficient = 0.005.
We defined the two-dimensional coordinate system around the actin filament (, ), where is the angle around the -axis. The position of the center of mass of the motor domain (residue number 1–780) of the myosin head is denoted by (, ). Umbrella sampling was used to enhance sampling at the high-free-energy region on the (, ) plane. The region of nm nm and was divided into blocks, where nm. We applied the umbrella potential to enhance the sampling of myosin located in each of the 60 blocks as(20)where nm () and (). The constants and were defined as kcal/mol/Å2 and 10 kcal/mol/rad2. For each of 60 umbrella potentials, we performed 12 independent runs of Langevin dynamics for steps and the data acquired in the first half of the run were discarded. From the data obtained with each of 60 umbrella potentials, we generated the histogram of (, ) with the bin size of nm and . We subsequently combined these data by WHAM [51] to calculate the -dependent free-energy landscape and the two-dimensional free-energy surface on the plane.
When we used only the 60 sets of the data sampled with 60 different umbrella potentials, the resulting landscape was not periodic because of the boundary effect. To avoid this numerical error, we replicated the whole data using the helical symmetry of the system as(21)with , and 2, so that in total five sets of data were used. We confirmed that copying twice in both directions, resulting in five repeats of the same data, yielded a sufficiently periodic free energy landscape.
In each MCS with a time scale of , both a change in the actomyosin state and diffusion in the plane can occur. The procedures in one MCS were as follows: (i) the actomyosin state changed with probability or in Fig. 2, which is defined to be considerably smaller than unity. We assumed that the rates of transitions among actomyosin states are fast () or slow () except for the transition from A.Mclosed to A.Mrigor, where represents the inverse of an MCS. The lifetime of each actomyosin state is determined by whether the rates of approach to or departure from that state are fast or slow. Parameters were chosen to lengthen the lifetime of the A.Mpre.ADP.Pi state: fast for , , , , and , and slow for , , , , and . See Fig. S2 for the other choices of parameter values. (ii) If the transition to the different actomyosin state was not chosen, the diffusion of the myosin head on the two-dimensional free-energy landscape in the current actomyosin state was chosen. The trial movement of the myosin head was represented by motion along a lattice with mesh size . At each trial, the movement of myosin in the direction was chosen with probability , whereas movement in the direction was chosen with probability (see below for the value of ). (iii) The trial move was defined by selecting either of two sites in the direction chosen in (ii) with probability 0.5. (iv) The free-energy difference accompanying the trial move was calculated, and the trial was accepted when or with the probability when , and otherwise rejected. The value of was determined as follows. One step in the direction is the displacement of nm, whereas one step in the direction is the displacement of nm with . Assuming that the Brownian motion of the myosin head is isotropic in the two-dimensional plane of , the average distance after steps is , which yields .
The diffusion constant of the freely diffusing myosin head can be roughly estimated as by considering the myosin head as an ellipsoid moving sidewise with semi-major axes of 8 nm and semi-minor axes of 2.5 nm in water with viscosity 0.89 pNnsnm at 300 K [36]. This value can be used to estimate , the time scale of an MCS. Because the average distance after time steps is given by , we have , which gives ns. Because this value is obtained by assuming the free diffusion of myosin, we should note that this estimate of should give a lower limit. With this estimate, the trajectory of 104 steps as shown in Fig. 6A has a time scale of several milliseconds or longer.
|
10.1371/journal.ppat.1000509 | Membrane-Anchored HIV-1 N-Heptad Repeat Peptides Are Highly Potent Cell Fusion Inhibitors via an Altered Mode of Action | Peptide inhibitors derived from HIV-gp41 envelope protein play a pivotal role in deciphering the molecular mechanism of HIV-cell fusion. According to accepted models, N-heptad repeat (NHR) peptides can bind two targets in an intermediate fusion conformation, thereby inhibiting progression of the fusion process. In both cases the orientation towards the endogenous intermediate conformation should be important. To test this, we anchored NHR to the cell membrane by conjugating fatty acids with increasing lengths to the N- or C-terminus of N36, as well as to two known N36 mutants; one that cannot bind C-heptad repeat (CHR) but can bind NHR (N36 MUTe,g), and the second cannot bind to either NHR or CHR (N36 MUTa,d). Importantly, the IC50 increased up to 100-fold in a lipopeptide-dependent manner. However, no preferred directionality was observed for the wild type derived lipopeptides, suggesting a planar orientation of the peptides as well as the endogenous NHR region on the cell membrane. Furthermore, based on: (i) specialized analysis of the inhibition curves, (ii) the finding that N36 conjugates reside more on the target cells that occupy the receptors, and (iii) the finding that N36 MUTe,g acts as a monomer both in its soluble form and when anchored to the cell membrane, we suggest that anchoring N36 to the cell changes the inhibitory mode from a trimer which can target both the endogenous NHR and CHR regions, to mainly monomeric lipopetides that target primarily the internal NHR. Besides shedding light on the mode of action of HIV-cell fusion, the similarity between functional regions in the envelopes of other viruses suggests a new approach for developing potent HIV-1 inhibitors.
| Acquired immunodeficiency syndrome (AIDS) is a major global health problem, and its causative agent, human immunodeficiency virus (HIV), is extensively studied. To start an infectious cycle HIV must fuse its membrane with that of its host cell. A specific protein on the virus surface facilitates this process by undergoing major conformational changes. Several virus-cell fusion inhibitors target transiently exposed regions during the conformational changes, thereby preventing progression of the fusion process. Here, we focused on a specific fusion inhibitor peptide having two distinct binding sites and modes of inhibitions. By simple chemical modifications we demonstrate a shift between these two modes of inhibition. Most importantly, we reveal novel details regarding the conformational changes during the fusion process. Furthermore, the chemical modifications extremely enhanced the fusion inhibitory potency of the peptide. Lastly, since the fusion process of HIV shares common features with diverse biological processes, our results might contribute to their research and therapeutic efforts as well.
| HIV-1, like other enveloped viruses, utilizes a protein embedded in its membrane, termed envelope glycoprotein 1(ENV), to facilitate the fusion process [1],[2],[3]. The ENV is organized as trimers on the membrane of the virus, and is composed of two non-covalently associated subunits. The surface subunit (SU), gp120, mediates host tropism [4],[5]), whereas the transmembrane subunit (TM), gp41, is responsible for the actual fusion event (reviewed in [6]). The extracellular part of gp41 is composed of several functional regions including the fusion peptide (FP), the N-terminal heptad repeat (NHR), the C-terminal heptad repeat (CHR), and the pre-transmembrane (PTM) domain.
The ability of the virus to fuse its own membrane with that of the hosting cell is due to conversion among three identified ENV conformations. Initially, the envelope subunits are in a metastable native conformation [7], in which gp41 is considered to be sequestered by gp120. Binding of gp120 to specific cell receptors involves conformational changes in both subunits, resulting in the pre-fusion conformation [7],[8] in which gp41 is exposed and extended, leading to insertion of the FP into the host cell membrane [9]. Additional conformational changes produce the post fusion conformation [10],[11], where a trimeric central coiled-coil is created by three NHR regions. These three CHR regions are packed in an anti-parallel manner into conserved hydrophobic grooves exposed on the surface of the central NHR coiled-coil. A complex representing this structure has been resolved by X-ray crystallography [12],[13], and is designated as the “six helix bundle” (SHB) or “core” structure. Similar bundles are created in intracellular vesicle fusion by SNARE proteins demonstrating a common mechanism in diverse systems [14],[15].
Inhibition of HIV-1-mediated fusion has been demonstrated by several N- or C-peptides: peptides that originate from the endogenous NHR or the CHR sequence of gp41, respectively [7],[16]. The common model is that C-peptides bind the endogenous NHR region in the pre-fusion conformation, thereby blocking core formation [17],[18],[19]. N-peptides, on the other hand, have two distinct modes of inhibitory action: binding of the endogenous CHR region in the pre-fusion conformation, thereby blocking core formation, and binding the endogenous NHR region to disrupt the creation of the internal NHR coiled-coil [20]. Previously it has been demonstrated that anchoring of inhibitory, expressed CHR peptides, to the membrane of cells can increase their inhibitory activity, as well as aid in deciphering the intermediate steps in the viruses' fusion [9],[21]. We have demonstrated earlier that conjugation of fatty acids to peptides can sufficiently anchor a short CHR-peptide to the membrane of cells, dramatically increase its inhibitory activity, and reveal the boundary of the core structure in a dynamic fusion process [22]. The observation that the increase in the inhibitory activity was significantly more pronounced when the fatty acid was attached to the C-terminus compared with the N-terminus supported a preferred orientation of the CHR peptide towards the endogenous pre-hairpin conformation. Here we address the role of the orientation of membrane bound NHR peptides during the ongoing fusion event and its implication on the understanding of the molecular mechanism of gp41 fusion.
Fmoc amino acids including lysine with a 4-Methyltrityl (MTT) side chain protecting group and Fmoc Rink Amide MBHA resin were purchased from Nova-biochem AG (Laufelfinger, Switzerland). Other peptide synthesis reagents, fatty acids octanoic acid (C8), dodecanoic acid (C12), and hexadecanoic acid (C16), LPC (lysophosphatidylcholine), and PBS were purchased from Sigma Chemical Co. (Israel). DiD (DiIC18(5) or 1,1′-dioctadecyl-3,3,3′,3′,-tetramethylindodicarbocyanine, 4-chlorobenzenesulfonate salt), DiI (1,1′- dioctadecyl-3,3,3′,3′,0 tetramethylinocarbocyanine perchlorate) lipophilic fluorescent probes and NBD-F (4-fluoro-7-nitrobenzofurazan) were obtained from Biotium (California, USA). Buffers were prepared in double-distilled water.
Cell culture reagents and media were purchased from Biological Industries Israel (Beit Haemek LTD). All cell lines were obtained through the NIH AIDS Research and Reference Reagent Program, Division of AIDS, NIAID, NIH. Jurkat E6-1 cells were from Dr. Arthur Weiss [23], Jurkat HXBc2 (4) cells expressing HIV-1 HXBc2 Rev and ENV proteins were from Dr. Joseph Sodroski [24], TZM-bl cells were from Dr.John C. Kappes, Dr. Xiaoyun Wu, and Tranzyme Inc [25],[26], and HL2/3 cells were from Dr. Barbara K. Felber and Dr. George N. Pavlakis [27]. Cells were cultured every 3 to 4 days, and maintained in RPMI-1640 or DMEM supplemented with the appropriate antibiotics at 37°C with 5% CO2 in a humidified incubator. For ENV expression, Jurkat HXBc2 (4) cells were transferred to medium without tetracycline three days prior to the experiments.
GCN4 trimer, C34, and N36 were synthesized on Rink Amide 4-Methylbenzhydrylamine (MBHA) resin by using the Fmoc strategy as previously described [28]. C-terminally conjugated N36 peptides contain a lysine residue at their C-terminus with an MTT side chain protecting group, enabling the conjugation of a fatty acid that required a special deprotection step under mild acidic conditions (2×1 min. of 5% TFA (trifluoro acetic acid) in dichloro methan (DCM) and 30 min. of 1% TFA in DCM). Conjugation of a fatty acid to the N-terminus was performed using standard Fmoc chemistry. Addition of the NBD (emission-530 excitation-467) fluorescent probe to the N- or C-terminus of selected peptides was performed using 3 equivalents of NBD-F in a 2% diisopropylamin (DIEA) solution in DMF for one hour. All peptides were cleaved from the resin by a TFA: DDW: TES (93.1∶4.9∶2 (v/v)) mixture, and purified by reverse phase high performance liquid chromatography (RP-HPLC) to >95% homogeneity. The molecular weight of the peptides was confirmed by platform LCA electrospray mass spectrometry.
The protocol utilizing Jurkat E6-1 and Jurkat HXBc2 cells for a cell-cell fusion assay was previously described [29]. In short, Jurkat E6-1 and Jurkat HXBc2 cells were labeled with DiI and DiD lipophilic fluorescent probes, respectively. The two cell populations were co-incubated, in a ratio of 1∶1, for 6 h in the presence of eight different concentrations of the inhibitory peptides. Prior to measurements the cells were washed, spinned, dissolved in PBS, and put on ice. Cells co-incubated without the presence of peptides served as an optimal fusion reference. Unlabeled cells that were handled similarly served as an intrinsic fluorescence control. Cells labeled separately with DiI or DiD were used to adjust the optimal separation of fluorescent signals. Jurkat HXBc2 cells labeled with DiI were co-incubated with Jurkat HXBc2 cells labeled with DiD for a fusion background that was subtracted from the measurements of the experiment. The following alterations were applied to the original protocol: (i) 5 µL of a 1 mg/mL DiI or DiD solution in dimethylsulfoxide (DMSO) was added to 1 mL of 4×106 cells/mL Jurkat E6-1 or Jurkat HXBc2 cells, respectively. (ii) For each data point 150,000 events were collected. Measurements were performed on a FACSort machine, upgraded to a FACSCalibur cell analyzer (Becton Dickinson). Fitting of the data points was performed according to the equation derived from Hills' equation:In this equation B is the maximum value, therefore it equals 100% fusion, A is the IC50 value, and c represents Hill's coefficient, in this particular case: the inhibitory oligomeric state of the peptide. For the fitting, we uploaded the X and Y values of the raw data (after subtracting the background) into a nonlinear least squares regression (curve fitter) program that provided the IC50 value (A of the equation), as well as the c value.
For triple staining, the same cell-cell fusion inhibition assay experiment as above was performed in the presence of NBD-labeled peptides. Cells labeled separately with DiI or DiD, and unlabeled cells in the presence of an NBD-labeled peptide were used to compensate for the optimal separation of the three fluorescent signals. For each data point 500,000 events were collected. The eight different possible combinations (triple, NBD, DiI, DiD, NBD+DiI, NBD+DiD, DiI+DiD, no label) were defined in the analysis software and the percentage of each one was calculated. The percentage of NBD labeling (peptide) on all cell types in relation to all available labeled cells in the system was calculated. This analysis provided us with the percentage of cells labeled with NBD-peptide.
Additionally, the percentage of NBD labeling (peptide) in cells labeled with DiD (effector) or DiI (target) cells was further calculated. Analysis of the data enabled us to examine the relative binding of labeled peptides to different cell populations, namely, target or effector cells.
CD measurements were performed on an Aviv 202 spectropolarimeter. The spectra were scanned using a thermostatic quartz cuvette with a path length of 1 mm. Wavelength scans were performed at 25°C, the average recording time was 15 sec., in 1 nm steps, the wavelength range was 190–260 nm. Peptides were scanned at a concentration of 10 µM in HEPES buffer (5 mM, pH 7.4) and in a membrane mimetic environment of 1% LPC in double distilled H2O.
To scrutinize the effect of anchoring N36 to the membrane, we conjugated octanoic, dodecanoic, and palmitic acids to the N-terminus of N36 (Table 1). The resulting peptides C8-N36, C12-N36, and C16-N36 were examined in a cell-cell fusion inhibition assay and the results are shown in Figure 1. A correlation was observed between the length of the conjugated fatty acid and the inhibitory activity of the N- conjugated N36 peptides. N36, C8-N36, C12-N36, and C16-N36 exhibited IC50 values of 488±119, 222±56, 190±21, and 72±27 nM, respectively. Interestingly, AcN36 was not active up to 2000 nM; therefore we refer to it as inactive. This correlates with previous studies demonstrating an IC50 of 16000±2000 nM and 584±46 nM for the acetylated and non-acetylated forms of N36, respectively [20],[30]. Overall, our data reveal that the anchoring of N36 to the membrane significantly increases its inhibitory activity.
To examine the importance of the proper orientation of the N36 peptide in relation to the pre-fusion conformation, we also conjugated octanoic, dodecanoic, and palmitic acids to the C-terminus of modified N36, termed N36M (Table 1). The parental peptide and the resulting fatty acid-conjugated peptides N36M, N36M-C8, N36M-C12, and N36M-C16 (Table 1) were examined in a cell-cell fusion inhibition assay and the results are presented in Figure 1. Likewise, a correlation was observed between the length of the conjugated fatty acid and the inhibitory activity of the C-conjugated N36 peptides. N36M, N36M-C8, N36M-C12, and N36M-C16 exhibited IC50 values of 531±48, 354±25, 241±89, and 159±47 nM, respectively. Since acetylating N36 abrogates its activity we added an acetyl group to N36M-C12 and N36M-C16 resulting in AcN36M-C12 and AcN36M-C16. Both lipopeptides were examined in a cell-cell fusion inhibition assay and exhibited IC50 values of 226±38, and 125±51 nM, respectively. Since these values are similar to those of N36M-C12 and N36M-C16 we can conclude that the charge in their N-terminus does not influence their inhibitory ability, contrary to N36.
Interestingly, there was only a slight difference between the activities of N- and C-terminally conjugated peptides having the same fatty acid. This was in contrast with the results obtained with the C-helix peptide, in which there was a marked difference (∼30-fold) between them [22]. Thus, we can conclude that the length of the fatty acid is important, and it is correlated to the inhibitory activity, whereas primarily, the orientation of the peptides is not critical for their activity pattern.
Representative experiments showing the inhibitory activity curves of N36 and its N-terminally fatty acid-conjugated analogs is presented in Figure 2A. It reveals different shapes of the inhibition curves for the different peptides shifted from sigmoid through a median shape to hyperbolic. A sigmoid shape can be explained by the tendency of N36 to oligomerize. Therefore, we speculated that the different binding curves might be attributed to a different inhibitory oligomeric state of the peptides. Consequently, for optimal fitting, we employed an equation that contains a cooperativity parameter, indicative in this case, to the inhibitory oligomeric state of the peptide. Therefore, after a fit is achieved, the c value represents the oligomeric state of the peptide. The values of the oligomerization parameters (averaged from at least four independent experiments) for the different peptides are presented in Figure 2B. The c values for the N- conjugated N36 peptides, namely: N36, C8-N36, C12-N36, and C16-N36 are 2.67, 2.61, 1.77, and 1.47 respectively. The c values for the C- conjugated N36 peptides, namely: N36M, N36M-C8, N36M-C12, and N36M-C16 are 3.19, 2.82, 1.67, and 1.19 respectively. The c parameter for the original peptide (N36 or N36M) was compared to the c parameter of its longest fatty acid conjugate (C16-N36 or N36M-C16) by the nonparametric Mann-Whitney test, the two sided significance was p = 0.016, demonstrating statistical significance for these results. These data reveal an interesting shift in the oligomerization tendency. It suggests that for the native peptides, N36, and N36M, the tendency is for the trimeric form. The longer the fatty acid the lower is the oligomerization value until it almost reaches a monomer with the C16-N36, and N36M-C16 peptides.
We tested whether the attachment of the fatty acids to the peptides allowed their anchoring to the cell membrane by utilizing a triple staining flow cytometry assay that incorporates fluorescently labeled target cells, effector cells, and inhibitory peptides [22]. This assay allowed the determination of the IC50 of the peptides, as well as monitoring the percentage of cells labeled with the different peptides (Table 2). We analyzed the most and least active peptides, namely, NBDN36 (parallel in its inhibitory activity to AcN36), NBDN36M-C16, and C16-N36NBD (Figure 3). The NBDGCN4 peptide served as a negative control for a non-binding peptide, whereas, C16-NBDGCN4 served as a positive control for a strongly binding peptide. As expected, both are not inhibitors (data not shown). The data reveal a direct correlation between the activity of the N-helix peptides and their global binding to the cells.
We determined the secondary structure of the most active and inactive peptides in solution to find out whether this feature correlates with their activity pattern. N36 and N36M exhibited α-helical structures in solution, whereas the structure of AcN36, C16-N36, and N36M-C16 was undefined (Figure 4). The peptides' ability to create a core structure with C34 in solution was also monitored. The CD signal of each peptide was measured and this signal was added to the signal of C34. This calculated combined signal would represent the signal in the case that the peptides do not interact with each other. This signal was compared to the actual signal monitored upon co-incubation of the two peptides together. If the two peptides interact one with each other, we would expect to see a difference between the two signals. AcN36, in contrary to the results presented in a previous study [31], and C16-N36 were unable to create a core structure, whereas N36 and N36M-C16 did interact with C34 (Figure 4) [32],[33]. Note that Chan et al have used the C-34 and the N-36 peptides both in their acetylated forms and obtained a stable core. Here we obtained a stable core with both peptides in their non-acetylated forms, but we could not get a stable core with one acetylated and one non-acetylated peptides, the reason for which is not clear. The structure of the peptides alone, and their ability to create a core structure with C34 was also measured in a membrane mimetic environment (Figure 4). Under these conditions, all the peptides exhibited α-helical structures. However, with all peptides, the non-interactive signal overlapped the experimental signal. In this case, the overlap does not necessarily mean that there is no creation of the core structure. Since all peptides have strong helical signals by themselves they could create a core structure without an observed change in the secondary structure. Overall, these data demonstrate that the structure of the peptides and their ability or inability to create a core structure with C34 (in solution or in a membrane mimetic environment) cannot account for their activity pattern.
In order to investigate further the mechanism of inhibition we utilized known N36 mutants [20] (see Table 3 for sequences). The first was N36 MUTe,g which contains mutations in its e and g positions. These mutations preserve its ability to self-assemble into trimers, but it cannot interact with the CHR. The second mutant was N36 MUTa,d which contains mutations in it's a and d positions knocking out its ability to interact with itself, thus leading to inability to create the internal coiled-coil. These mutants demonstrated that the NHR can inhibit by preventing the formation of the viral NHR coil-coiled (probably as a monomer or dimer), or by binding to the CHR domain to prevent SHB formation (probably as a trimer). We conjugated a palmitic acid to the N or C- terminus of both of them (Table 3) and determined their IC50 inhibitory values. As expected, N36 MUTa,d was inactive alone and when conjugated to palmitic acid, because it could not bind itself, as well as CHR, therefore both modes of inhibitions could not take place. Strikingly however, the attachment of palmitic acid to N36 MUTe,g caused an increase of 7-fold to 100-fold in its IC50 compared to the soluble peptide, depending on the directionality of the conjugation. N36 MUTe,g, C16-N36 MUTe,g, and N36 MUTe,g-C16 exhibited IC50 values of 936±36, 162±4, and 8.8±4 nM, respectively (Figure 5). Such preference was not observed with the wild type N36 which preserve binding to the CHR region. The data analysis suggested a trimeric and monomeric modes of inhibition for the wild type N36 and its palmitic acid conjugates, respectively (Figure 2B). Here, N36 MUTe,g, C16-N36 MUTe,g, and N36 MUTe,g present oligomerization parameters values of 1.4, 0.77, and 1.4 respectively (Figure 5B), suggesting primarily a monomeric mode of inhibition. The c parameter of N36 MUTe,g, C16- N36 MUTe,g, and N36 MUTe,g-C16 was compared to the c parameters of N36, N36M, C16-N36, N36M-C16, and to themselves by the nonparametric Mann-Whitney test. Even though our sample size is small, out of 14 comparisons only one did not obey our predictions.
Our cell-cell fusion assay is based on lipophilic fluorescent probes. Therefore, there is a risk that the inhibitory results that we obtained are due to hemifusion. In order to exclude this possibility, we performed a reporter gene cell-cell fusion assay for representative peptides as a proof of concept. The gene reporter assay is based on activation of HIV long terminal repeat-driven luciferase cassette in TZM-bl (target) cells by HIV-1 tat from the HL2/3 (effector) cells. The peptides that we examined were: N36, N36M, N36M-C16, and N36 MUTe,g-C16. Their original IC50 values (nM) were: 488, 531, 159, and 8.8 respectively, in comparison to: 472, 333, 128, and 8 respectively, in the gene reporter assay experiment. Since the values were comparable we conclude that the inhibitory results obtained with our cell-cell fusion assay represent full fusion and not hemifusion.
To examine whether the peptides have an enhanced tendency to bind the cells with the receptors (target cells), or those with the ENV glycoprotein (effector cells), in a dynamic fusion process, we employed a triple staining assay. Fluorescently labeled peptides were incubated with differently labeled effector and target cells, exactly according to the protocol of the cell-cell fusion assay. The fusion was allowed to take place and then the sample was washed and measured by FACS. Further analysis, as specified in the Materials and Methods section, enabled us to compare the relative level of the peptide's binding for each cell population (Figure 6). The NBDGCN4 peptide served as a negative control for a non-binding peptide, whereas C16-NBDGCN4 served as a positive control for a strongly binding peptide without preference for a specific cell population. A line is drawn in each panel to emphasize where we would expect the data in case there is no preference among the different populations. Since the NBDGCN4 peptide does not bind the membranes, all the data points are concentrated in the lower left-hand corner. We can conclude that (in the same conditions as for the experiments determining the inhibitory activity of the peptides) there is a tendency of the conjugated N36 peptides to reside more on target than on effector cells.
Using synthetic peptides with homologous sequences to endogenous domains within gp41 is a powerful tool to decipher the molecular mechanism of HIV-cell fusion. Among these peptides the NHR and CHR play a crucial role. Studies with soluble CHR derived peptides support the current model in which gp41 adopts an extended conformation in the pre-fusion step, inserts the fusion peptide into the target membrane while the NHR forms a trimeric coil-coiled structure. A critical step toward membrane fusion is the collapse of this structure to form the SHB. CHR-derived synthetic peptides can prevent SHB formation by competing with the endogenous CHR domain for the binding of the NHR trimers (Figure 7). For such an inhibition to occur, CHR needs to bind in an antiparallel manner to the NHR. To support this mechanism, we have previously anchored short CHR peptides to the cell membrane by palmitic acid conjugation. The CHR derived lipopeptides had 30-fold higher inhibitory activity when attached via their C-terminus (antiparallel to the endogenous NHR), compared to the N-terminus. That study also demonstrated the C-terminal boundary of the six helix bundle [22].
NHR peptides display a distinct feature in comparison to the CHR peptides in their ability to self oligomerize in solution [34],[35]. Thus, they can bind to two endogenous domains of gp41 in the pre-fusion extended conformation [20]. Synthetic NHR can bind the endogenous NHR to prevent the formation of the coil-coiled NHR trimer probably as dimers or as monomers (Figure 7). In addition, NHR can bind the CHR and hence prevent endogenous SHB core formation (Figure 7). Binding to the CHR region depends on the ability of the NHR to homo-oligomerize [20], probably as a trimer. In support of this, enhancing the trimeric tendency of N36 increases its inhibitory activity [36],[37]. The accepted model is that this trimeric N36 form is primarily responsible for binding of the CHR region. In both of these mechanisms the directionality of the NHR towards the internal pre-fusion conformation seems to be important. To test this hypothesis we conjugated fatty acids with increasing lengths to the N or C-terminus of N36, and the inhibitory activity of the resulting lipopetides was examined in a cell-cell fusion assay. Importantly, the IC50 of the resulting lipopeptides increased significantly in correlation to the length of the fatty acid (Figure 1), as well as to their ability to bind cells which was examined by a triple staining assay (Figure 3). However, in contrast with the CHR, we found that the directionality of the attachment of N36 was not critical, since the attachment of the same fatty acid to N36 increased its inhibitory activity similarly, independent of whether it was attached to the C- or N-terminus (Figure 1). These findings suggest a planary orientation of the endogenous NHR region, as well as the N36 lipopeptides, on the cell membrane. Indeed, previous studies have revealed that NHR derived peptides can bind and assemble on a membrane and adopt an α-helical structure [38],[39],[40],[41],[42],[43]. Since it is less likely that internal coiled-coil will disassemble after its creation, we suggest that a loose extended conformation is created after the conformational changes induced by the receptors and co-receptors binding. In this conformation the FP is inserted into the host cell membrane but the internal coiled-coil is not formed yet (“loose” Pre-fusion in Figure 8). Then, the NHR coiled-coil is formed which leads to its parallel orientation towards the membrane, and finally folding into the post-fusion conformation. In this model the peptides with the long fatty acid will create a chimeric coiled-coil with endogenous NHR leading to an altered sequence of events as is presented in Figure 8 at the bottom. However, it is possible that the conjugated peptides inhibit partially also from solution. We suggest that these gp41 conformations: the loose pre-fusion conformation and the NHR region lying parallel to the cell membrane, are additional intermediate conformations during the fusion process.
We observed two different shapes of the inhibition curves of N36 and its fatty acid derivatives: sigmoid for the N36 and short fatty acids conjugated peptides, in contrary to hyperbolic for the longer fatty acid conjugates (see Figure 2A). We utilized a derivative of Hills' equation for the fitting of the experimental cell-cell fusion assay data, and for extracting the IC50 value, in which the Hill coefficient represents the oligomeric tendency of the peptide in the inhibition process. Examining the oligomeric parameters revealed an interesting trend. The soluble unmodified N36 and N36M peptides act as trimers, whereas the strongly membrane bound lipopeptides C16-N36 and N36M-C16 act as monomers (see Figure 2B). We speculated that N36 mostly binds the CHR as a trimer (However, the monomeric fraction of these peptides can also bind the endogenous NHR) while C16-N36 mainly binds the endogenous NHR as monomers.
To further support this, we conjugated palmitic acid to the N- or C-terminus of two previously studied N36 mutants: (i) N36 MUTe,g which contains mutations in its e and g positions resulting in its inability to interact with the CHR, and (ii) N36 MUTa,d which contains mutations in it's a and d positions knocking out its ability to interact with itself, leading to inability to create the internal coiled-coil [20]. Fatty acid conjugation to N36 MUTa,d could not compensate for the inhibitory obligatory requirement of N36 self binding. In contrast, fatty acid conjugation dramatically increased the inhibitory activity of N36 MUTe,g (up to 100-fold). In this case the C-terminal anchored N36 MUTe,g-C-16 is about 20-fold more active than the N-terminal anchored C-16-N36 MUTe,g. Importantly, N36 MUTe,g, C16-N36 MUTe,g, and N36 MUTe,g can bind the NHR but not the CHR. The fitting of their inhibitory curves reveal that they bind the endogenous NHR region in a monomeric form (Figure 5B).
The enhanced inhibitory activity of the conjugated peptides can be accounted for by: (i) increased local concentration of the conjugated peptides on the membrane surface resulting in increased accessibility near the fusion site. (ii) If conjugation of a fatty acid indeed changes the tendency of the peptide into a monomeric inhibitory mode of action, then less peptides are required to exert the same inhibitory effect thus reducing the IC50 value, and (iii) When N36 inhibits it can bind simultaneously to the NHR and CHR regions of the same pre-fusion structure. In contrast, when a peptide can only bind one target site (like N36 MUTe,g or the monomeric conjugated peptides) a lower concentration exerts the same effect thus reducing the IC50 value. Interestingly, analysis of the inhibitory curves of C-helix peptides also reveal, as expected, a monomeric mode of action (data not shown). Since, similarly to the anchored N36 MUTe,g, a C-helix peptide can bind only the NHR, the inhibitory activities of N36 MUTe,g-C16 and e.g. T-20 are similar and in the low nanomolar range.
Combining these results it seems that N36, N36M or the lipopeptides with the short fatty acids primarily target the endogenous CHR region as trimers, and that conjugation of a long fatty acid leads to a shift toward a lower oligomerization requirement for the inhibition reaction, thereby primarily targeting the endogenous NHR region. Apparently, the membrane bound peptide does not depend on trimerization for the inhibition activity similarly to the N36 peptide in solution; membrane binding compensates for the trimerization requirement. Strengthening this assumption is another interesting finding - a tendency of the N36 conjugates to reside more on the target cells that occupy the receptors than on the effector cells (Figure 5). This feature was detected by a triple staining assay performed under the same protocol conditions utilized for the cell-cell fusion assay.
A new anti-HIV-1 therapeutic category classified as fusion inhibitors emerged to the HAART (Highly Active Antiretroviral Therapy) with the entry of a C-peptide named enfuvirtide. The potential of C- and N-peptides to inhibit the fusion process of the virus was discovered simultaneously. Nevertheless, most efforts were aimed at developing C-peptides as drugs. This was due to inferior inhibitory activities demonstrated by N-peptides, which were attributed to their tendency to form weakly active oligomers. Studies that have demonstrated improved inhibition of N-peptides are rare and include (i) Stabilization of a specific, usually a trimeric, coiled-coil NHR complex, by fusion to unrelated coiled-coils [36], by covalently connecting NHR regions [37],[44],[45], or by combining different methods including point mutations in specific heptad repeat positions [17]. (ii) Creation of an incomplete core complex [46]. (iii) Abolishing the CHR binding capability by altering the e and g positions of the N36 heptad repeat [20], resulting in enhanced inhibition, probably due to reduced aggregation. Most of these methods involve elaborate techniques. Here we have demonstrated a significantly enhanced inhibitory activity of an N-peptide by a simple chemical reaction that involves the attachment of a fatty acid to N36. The similarity between functional regions in the envelopes of many viruses suggests a possible new therapeutic approach.
In summary, taking all of the results into consideration leads us to suggest that our peptides demonstrate a shift in the inhibitory mode of action from mainly a trimeric, oligomeric N36/N36M complex, which can target either the internal NHR coiled-coil or the CHR region, to monomeric lipopetides that mostly target the internal NHR coiled-coil. Additionally, the similar inhibitory effect of the N- and C-terminally conjugated peptides suggests that the mode of inhibition involves a planary peptide orientation on the membrane's surface indicating a possible additional intermediate conformation during the fusion process. (Figure 8). Importantly, this study demonstrates that a simple chemical conjugation of fatty acids to N36 can significantly increase its inhibitory activity.
|
10.1371/journal.pgen.1004492 | Tethering Sister Centromeres to Each Other Suggests the Spindle Checkpoint Detects Stretch within the Kinetochore | The spindle checkpoint ensures that newly born cells receive one copy of each chromosome by preventing chromosomes from segregating until they are all correctly attached to the spindle. The checkpoint monitors tension to distinguish between correctly aligned chromosomes and those with both sisters attached to the same spindle pole. Tension arises when sister kinetochores attach to and are pulled toward opposite poles, stretching the chromatin around centromeres and elongating kinetochores. We distinguished between two hypotheses for where the checkpoint monitors tension: between the kinetochores, by detecting alterations in the distance between them, or by responding to changes in the structure of the kinetochore itself. To distinguish these models, we inhibited chromatin stretch by tethering sister chromatids together by binding a tetrameric form of the Lac repressor to arrays of the Lac operator located on either side of a centromere. Inhibiting chromatin stretch did not activate the spindle checkpoint; these cells entered anaphase at the same time as control cells that express a dimeric version of the Lac repressor, which cannot cross link chromatids, and cells whose checkpoint has been inactivated. There is no dominant checkpoint inhibition when sister kinetochores are held together: cells expressing the tetrameric Lac repressor still arrest in response to microtubule-depolymerizing drugs. Tethering chromatids together does not disrupt kinetochore function; chromosomes are successfully segregated to opposite poles of the spindle. Our results indicate that the spindle checkpoint does not monitor inter-kinetochore separation, thus supporting the hypothesis that tension is measured within the kinetochore.
| The spindle checkpoint monitors tension on chromosomes to distinguish between chromosomes that are correctly and incorrectly attached to the spindle. Tension is generated across a correctly attached chromosome as microtubules from opposite poles attach to and pull kinetochores apart, but are resisted by the cohesin that holds sister chromatids together. This tension generates separation between kinetochores as pericentric chromatin stretches and it also elongates the kinetochores. To monitor tension, the checkpoint could measure the separation between kinetochores or the stretch within them. We inhibited the ability of pericentric chromatin to stretch by tethering sister centromeres to each other, and we asked whether the resulting reduction in inter-kinetochore separation artificially activated the spindle checkpoint. Inhibiting inter-kinetochore separation does not delay anaphase, and the timing of mitosis was the same in cells with or without the spindle checkpoint, showing that the checkpoint is not activated. Inhibiting chromatin stretch does not alter the function of kinetochores as chromosomes are still segregated correctly, nor does it hinder the checkpoint. Cells whose sister kinetochores are held together can still activate the checkpoint in response to microtubule depolymerization. Our results indicate the spindle checkpoint does not monitor inter-kinetochore separation and likely monitors tension within kinetochores.
| Faithful chromosome segregation is essential. Mistakes lead to aneuploidy [1], cancer progression [2], and birth defects [3]. To ensure proper division of chromosomes, eukaryotes have evolved the spindle checkpoint, which monitors the kinetochore, a large multi-protein complex that assembles on centromeric DNA and attaches microtubules to chromosomes. In Saccharomyces cerevisiae, the budding yeast, the kinetochore consists of over 65 proteins that are assembled on the conserved 125 bp centromere [4]. The spindle checkpoint delays the onset of chromosome segregation until all chromosomes have attached their two sister kinetochores to microtubules emanating from opposite poles (bi-orientation) [5], [6]; it is activated by unattached kinetochores [5], [7] and lack of tension at the kinetochore [8], [9]. Morphologically, the checkpoint regulates the transition between metaphase, when the pairs of sister chromatids are aligned equidistant from the two poles, and anaphase, when the sisters split apart and are pulled to opposite poles.
Bi-oriented kinetochores are under tension: microtubules pull them towards the poles, but the chromosomes they lie on are held together by cohesin. In metaphase, this tension can be seen as separation of GFP-labeled centromeres [10], [11] and by elongation of the kinetochores, detected by measuring the separation between different kinetochore proteins [12]–[14]. In budding yeast, removing tension (by preventing replication or uncoupling sister chromatids) activates the spindle checkpoint and arrests cells in mitosis [9]. An unpaired, tensionless chromosome in praying mantid spermatocytes delays cell division, and applying tension to this chromosome allows cells to enter anaphase [8]. Although there is debate about whether the release of tension, or the subsequent release of microtubules from the kinetochore, generates the molecular signal that arrests cells in mitosis, it is clear that kinetochores can monitor tension, thus controlling the stability of microtubule attachment and progress through mitosis. The release of chromosomes and subsequent cell cycle arrest by the spindle checkpoint requires the activity of Sgo1 and the protein kinase, Ipl1/Aurora B [15]–[19].
Where does the spindle checkpoint measure tension? There are two possible locations: between the two sister kinetochores (inter-kinetochore, L1 in Figure 1A) or within an individual kinetochore (intra-kinetochore, L2 in Figure 1A). Inter-kinetochore tension could be measured by the stretching of pericentric chromatin [11], or by a protein spring that spans the distance between kinetochores, such as PICH [20], a protein seen to span the inter-kinetochore gap in HeLa cells [21], [22] (Figure 1B). Intra-kinetochore stretch could be detected by monitoring changes in the distance between different parts of the kinetochore or conformational change in a single molecule. For either model, stretch stabilizes microtubule attachment to the kinetochore and relaxation destabilizes attachment and activates the checkpoint [12], [13], [23] (Figure 1B).
We manipulated budding yeast chromosomes to determine whether inter- or intra-kinetochore stretch regulates the spindle checkpoint (Figure 1C). By binding the tetrameric form of the GFP-labeled Lac repressor to an array of Lac operators, we held sister centromeres together (and measured their separation), inhibiting inter-kinetochore separation as cells entered mitosis. Despite the inhibited inter-kinetochore stretch, sister chromatids still separated on schedule, even though our manipulation left cells capable of assembling functional kinetochores and activating the spindle checkpoint. Because inhibiting inter-kinetochore separation does not slow mitosis, we believe that the spindle checkpoint senses tension by monitoring events within the kinetochore.
Tension on bi-oriented chromosomes allows the spindle checkpoint to distinguish between correct and incorrect attachments. Tension increases the separation between the centromeres and the kinetochores that have been assembled on them [10], [11] (L1 in Figure 1A) and the separation between components in a single kinetochore [12]–[14] (L2 in Figure 1A), but we do not know which distance the checkpoint monitors. To reduce the inter-kinetochore distance (L1), we tethered the sister chromatids of Chromosome III to each other by placing Lac operator (LacO) arrays on either side of the centromere and expressing two alternative versions of the Lac repressor. The tetrameric Lac repressor (LacI4) can bind simultaneously to two chromatids thus holding them together. The dimeric form of the repressor (LacI2) [24] is a control; it binds the Lac operator, but the two DNA binding domains must bind to the same operator, preventing the dimer from holding two DNA molecules together. It has been previously demonstrated that the tetrameric Lac repressor can hold homologous sister chromosomes together during meiosis in budding yeast while the dimeric Lac repressor cannot [25]. Both repressors were fused to GFP to see the centromeric DNA. Centromeric separation gives rise to two GFP dots [10], [11], and one GFP dot indicates two centromeres separated by less than the resolution of light microscopy, which is theoretically 200 nm, but is probably closer to 350 nm in our hands (Figure 2B). Both repressors contained two point mutations (P3Y and S61L) in the DNA-binding domain to produce the tightest binding affinity of all characterized Lac repressors (Kd≈10−15 M) [26].
We asked if the tetrameric Lac repressor inhibits centromere separation in metaphase. Cells were synchronized in G1 by treating them with the mating pheromone, alpha factor, released from this arrest, and allowed to proceed to a metaphase arrest, caused by removal of Cdc20, an essential activator of the anaphase promoting complex (APC, Figure 2A). Cells expressing GFP-LacI2 or GFP-LacI4 were sampled every 30 minutes for 3 hours and examined by fluorescence microscopy. Their centromeres were scored as stretched apart (2 GFP dots) or unstretched (1 GFP dot) (Figure 2B). We initially placed a Lac operator array on only one side of the centromere, but we found that a single array did not inhibit the separation of the centromeres (Figure 2C). Both dimer- and tetramer-expressing cells containing an array on one side of the centromere had equivalent percentage of visibly separated centromeres at all time points; there was no statistical difference between the two populations (p>0.35 at all time points, Student's t-test).
To better tether the two chromatids together, we placed Lac operator arrays on both sides of the centromere (Figure 2D). For the first 30 minutes after their release from alpha factor, control (GFP-LacI2) or tethered (GFP-LacI4) cells both showed little centromere separation (<10% stretched) consistent with cells being in S phase and lacking a spindle. At 60 minutes, cells were entering mitosis: 50±3% of control, GFP-LacI2 cells (n>100) had 2 GFP dots whereas only 24±2% of tethered GFP-LacI4 cells (n>100) had 2 dots (p<0.005, Student's t-test, Figure 2D). Throughout the remaining time points, approximately 50% of control GFP-LacI2 cells had 2 dots, similar to previous studies [10], [11]. Cells expressing GFP-LacI4 had significantly lower percentage of visible inter-kinetochore separation at all time points (p<0.005, Student's t-test), but the fraction rose during the metaphase arrest from 24±2% at 60 minutes to 42±2% at 180 minutes (p<0.005, Student's t-test). This experiment shows that the tetrameric Lac repressor can reduce inter-kinetochore separation only if Lac operator arrays are placed on both sides of the centromere, and reveals that this effect is primarily kinetic: the fraction of cells with visibly separated centromeres rises slowly during a prolonged metaphase arrest.
We interpret the reduction in the fraction of cells with 2 GFP dots as evidence that the tetrameric Lac repressor is tethering the chromatids together, inhibiting the stretch of a correctly bi-oriented chromosome whose sister chromatids have attached to both poles. However, it is possible that the tetrameric Lac repressor generates fewer cells containing 2 GFP dots because it disrupts kinetochore assembly or slows error correction mechanisms in a way that the dimeric Lac repressor does not. If tetrameric Lac repressor disrupts kinetochores or inhibits error correction, a higher frequency of GFP-LacI4 bound chromosomes should be mis-segregated compared to GFP-LacI2 bound chromosomes. To test the segregation of GFP-LacI2 and GFP-LacI4 bound chromosomes, cells were arrested in anaphase using a temperature sensitive cdc15-2 allele that inhibits mitotic exit [27]. Cells were synchronized in G1 with alpha factor, raised to the restrictive temperature, washed and released at the restrictive temperature to arrest cells in anaphase (Figure 3A). Cells were collected for scoring three hours after release from their G1 arrest, allowing cells to proceed to and arrest in anaphase as previously described [9], [27], [28]. Cells were stained with DAPI to confirm their arrest. Anaphase cells are large-budded and have DNA masses in each cell (Figure 3B); 99±1.5% of cells scored displayed this morphology. Correct segregation of the GFP-LacI bound chromosome was scored by the presence of one GFP dot in each mother and daughter cell, and mis-segregation was scored by one cell possessing both copies of the chromosome (two GFP dots in one cell) (Figure 3C). As a control, the segregation of GFP-labeled Chromosome III was also measured in cells with a conditional centromere. The GAL1 promoter was placed upstream of CEN3; when cells are grown in glucose, the promoter is silent and the centromere functions normally (Figure 3D). When cells are grown in galactose, transcription initiated from the GAL1 promoter disrupts centromere function and the chromosome is mis-segregated a high frequency [29]. Similar to previous studies using the conditional centromere [28], we found that 96±1% of cells grown in glucose correctly segregated the chromosome, but correct segregation occurred in only 41±6% of cells grown in galactose (Figure 3E). The presence of tetrameric Lac repressor did not disrupt chromosome segregation; both GFP-LacI2 and GFP-LacI4 bound chromosomes segregated correctly in 92±3% of cells. There was no statistical difference between cells grown in glucose, cells with GFP-LacI2, and cells with GFP-LacI4, but all were significantly different from cells grown in galactose (p≤0.003, Student's t-test). These results indicate that the presence of tetrameric Lac repressor does not disrupt kinetochore assembly or interfere with the correction of erroneous attachments, suggesting that the reduction in the fraction of metaphase-arrested cells with 2 GFP dots (Figure 2D) represents chromosomes that are correctly attached to opposite poles but cannot stretch apart due to the tethering effect of the tetrameric Lac repressor.
Does reduced inter-kinetochore separation produced by binding the tetrameric Lac repressor near the centromere activate the spindle checkpoint and thus delay the onset of anaphase? Cells were synchronized in G1 with alpha factor, washed and released to proceed through the cell cycle under conditions where they produce Cdc20, activate the APC, enter anaphase, and divide. Samples were taken every 10 minutes, fixed, and visualized to score mitotic progression (Figure 4A). Cells were scored for anaphase by the segregation of their GFP-labeled chromosome (Figure 4B). The separation of sister centromeres that indicates bi-orientation is always less than 1 µm, whereas the separation associated with anaphase is always greater than 2 µm, making it easy to rigorously distinguish the centromere separation associated with metaphase bi-orientation from the chromosome segregation of anaphase. The fraction of anaphase cells falls at the end of the experiment because cells divide, producing two daughter cells, each containing a single GFP dot. Control cells expressing GFP-LacI2 began to enter anaphase 40–50 minutes post-release from G1 and peaked with approximately 80% of cells in anaphase between 60 and 70 minutes. By 100 minutes, nearly every cell had exited mitosis (Figure 4C). Cells expressing GFP-LacI4 showed the same pattern of mitotic progression as control cells; they entered anaphase, reached a peak fraction of anaphase cells, and had fully exited mitosis at the same time as the GFP-LacI2 control (Figure 4C). At each time point, there was no statistically significant difference between control and tethered cells, suggesting that inhibition of chromatin stretch does not activate the spindle checkpoint. Since some of the cells that express GFP-LacI4 have not achieved the metaphase separation of sister centromeres after two hours in metaphase-arrested cells (Figure 2D), but cells that are allowed to pass through mitosis all complete anaphase within 90 minutes, we conclude that the failure to achieve metaphase centromere separation does not prevent entry into anaphase.
It is possible, however, that both GFP-LacI2 and GFP-LacI4 cells activated the spindle checkpoint and experienced mitotic delay. To rule out this possibility, we removed Mad2, an essential component of the spindle checkpoint, from both dimeric and tetrameric Lac repressor strains. All four strains (GFP-LacI2, GFP-LacI4, GFP-LacI2 mad2Δ, and GFP-LacI4 mad2Δ) moved through mitosis on the same time scale, with the peak of anaphase 60–70 minutes after release from G1 and with no statistically significant difference between any of the four strains (Figure 4C). These results show that neither the dimeric or tetrameric Lac repressor cause a mitotic delay by activating the spindle checkpoint.
We wanted to eliminate the possibility that our manipulations had interfered with the checkpoint in either of two ways. The first is that introduction of the tethering components (Lac operator and either form of the Lac repressor) might disrupt the spindle checkpoint. The second is that tethering sister centromeres might activate the checkpoint and, as a result, strains containing the tetrameric Lac repressor could only be produced by selecting cells that have mutationally or epigenetically inactivated the checkpoint. To confirm that strains expressing either form of the Lac repressor can still activate the spindle checkpoint, cells were synchronized in G1 with alpha factor and released into the microtubule-depolymerizing drugs benomyl and nocodazole (Figure 5A). Treatment with these drugs activates the spindle checkpoint, preventing cells from going through mitosis and causing them to arrest as large-budded cells [5]. Approximately 90% of dimeric and tetrameric repressor-containing cells reached the large-budded stage 120 minutes after being released from G1 into microtubule poisons and remained arrested at this stage for the duration of the experiment (Figure 5B). Cells that lacked Mad2 (GFP-LacI2 mad2Δ and GFP-LacI4 mad2Δ) did not arrest; after peaking at a value of 90% at 120 minutes, the fraction of large-budded mad2Δ cells declined to 55% at 180 minutes and 20% at 240 minutes (p≤0.002 for all time points, Student's t-test), compared to the MAD2 cells, 90% of which remained large-budded in the presence of the drugs. The difference between mad2Δ and MAD2 cells was statistically significant at 180 and 240 minutes post-release (p≤0.005, Student's t-test). These results show that cells expressing the dimeric and tetrameric forms of the Lac repressor remain capable of activating the spindle checkpoint and arresting the cell cycle. To demonstrate that Lac repressor-containing cells can inactivate the spindle checkpoint and resume mitosis, cells expressing the dimeric or tetrameric Lac repressor were synchronized in G1 with alpha factor, released into benomyl and nocodazole. After 90 minutes of drug treatment, the cells were washed and transferred to drug-free media (Figure 5C). During drug treatment, no dimeric or tetrameric-expressing cells entered anaphase (0% anaphase cells through T = 90 minutes), but after drug wash-out (marked by red arrow) both dimeric and tetrameric cells recovered from the mitotic arrest and began entering anaphase (Figure 5D). By 150 minutes after their release from G1 arrest, approximately 30% of both GFP-LacI2 and GFP-LacI4 cells had entered anaphase. This result shows that both strains have functional spindle checkpoints that can be inactivated to allow cells to resume mitosis.
The spindle checkpoint ensures that all chromosomes are properly attached to the spindle; it monitors microtubule attachment to kinetochores and the tension generated when sister kinetochores attach to opposite spindle poles. We found that the binding of the Lac repressor to LacO arrays surrounding a budding yeast centromere holds sister kinetochores close together and we asked whether the checkpoint monitors tension within the kinetochore (L1 in Figure 1A) or responds to the distance between sister kinetochores (L2 in Figure 1A). Holding sister centromeres together did not activate the checkpoint, suggesting that the checkpoint senses tension by monitoring events within the kinetochore rather than responding to reduced distance between sister centromeres.
We compared the behavior of cells expressing tetrameric and dimeric forms of the Lac repressor to determine the effect of slowing the sister centromere separation associated with bi-orientation. By 60 minutes after their release from G1 into a metaphase arrest, the cells expressing GFP-LacI2 had reached a steady state, with half of them showing two GFP dots. This value is similar to previous observations [10],[11] and reflects oscillations in the distance between sister centromeres that can take their separation below the level detectable by light microscopy (“breathing”) [11], [30], [31]. The tetrameric repressor (GFP-LacI4) reduced the fraction of cells with visibly separated GFP dots (Figure 2D). Their percentage increased from 24% to 42% during the two hours the cells spent in metaphase, suggesting that spindle forces can gradually overcome the Lac repressor's tether, despite this tether being the tightest binding version of the Lac repressor [26]. We attribute the increase in the fraction of cells with one GFP dot to the tetrameric repressor holding chromatids together that are correctly attached to opposite poles (Figure 1C). To eliminate the possibility that tetrameric repressor increased the fraction of cells with one GFP dot cells by disrupting kinetochores or correction of erroneous attachments, we showed that the rate of chromosome mis-segregation is not increased in cells expressing the tetrameric repressor compared to control cells and those expressing the dimeric repressor (Figure 3E). Because the assay we used cannot reliably detect frequencies of chromosome mis-segregation below 5%, we cannot exclude the possibility that the presence of the tetrameric Lac repressor does not elevate the frequency of mitotic chromosome loss above the normal rate of 10−5/cell division. But we are confident that the long delay in separating sister chromatids in the cells expressing the tetrameric repressor is not due to their failure to attach to opposite spindle poles.
By repeating our experiment in cells that could enter anaphase, we showed that inhibiting sister centromere separation did not activate the spindle checkpoint to the point that delayed entry into anaphase. Cells expressing the dimeric and tetrameric forms of the Lac repressor progressed through mitosis indistinguishably: 60 minutes after their release from G1 arrest, most of the cells were in anaphase, even though there is a marked difference between the degree of inter-kinetochore stretch (23% of tetramer- versus 50% of dimer-expressing cells, Figure 2D) at this time in cells that have been arrested in metaphase. Observing the same kinetics of anaphase in cells expressing dimeric and tetrameric forms of the Lac repressor shows that inhibiting inter-kinetochore separation and thus the stretch of pericentric chromatin does not delay the cell cycle or the ability of microtubule-dependent forces to move kinetochores in anaphase (Figure 4C).
To eliminate the possibility that the dimeric and tetrameric versions of the Lac repressor were activating the spindle checkpoint, we tested the effect of removing Mad2, an essential component of the checkpoint. With either form of the repressor, the timing of mitosis is unchanged when the spindle checkpoint was deleted (Figure 4C), demonstrating that neither form activates the checkpoint. We also checked that our strains had a functional checkpoint. Cells expressing either form of the Lac repressor arrested as large-budded cells in response to microtubule depolymerization (Figure 5B), and the cells only entered anaphase once the microtubule-depolymerizing drugs were removed (Figure 5D).
Our results suggest that the spindle checkpoint does not monitor the distance between sister kinetochores. We cannot make this a rigorous conclusion because the tetrameric Lac repressor reduces inter-kinetochore separation rather than abolishing it. We can only detect that a higher fraction of centromere pairs are separated by a distance smaller than the resolution limit of our microscope, and despite the presence of the tetrameric Lac repressor, some cells still manage to produce visible, metaphase separation between sister centromeres. Nevertheless, we might expect that some of the cells that express the tetrameric repressor have their sister centromeres close enough together to activate the spindle checkpoint and thus that some of the cells in this strain would enter anaphase more slowly than the control strain expressing the dimeric repressor. We see no such effect, leading us to argue that the checkpoint does not monitor inter-kinetochore distance. We assayed for spindle checkpoint activation by mitotic progression; cells that had activated the checkpoint should be delayed in entering anaphase [5], [6]. The sensitivity of our assay would reveal if cells with tethered kinetochores are delayed in metaphase by 10 minutes or more, but we cannot rule out transient checkpoint activation on a shorter time scale. Unfortunately, no other method for assaying spindle checkpoint activation would provide greater resolution for activation caused by a single chromosome in budding yeast. Unlike higher eukaryotes, methods such as visualizing Mad1 or Mad2 bound to individual kinetochores are not feasible in budding yeast because kinetochores are too clustered to distinguish individual kinetochores, and localization of these checkpoint proteins to kinetochores has only been demonstrated in response to global spindle defects [32], [33]. Nevertheless, because it takes much longer to overcome the tether in metaphase arrested cells (Fig. 2D), than it takes the same cells to proceed through an unrestrained mitosis (Fig. 4C), we argue that the presence of the tether does not substantially activate the spindle checkpoint.
If the checkpoint does not monitor events between sister centromeres, it must respond to changes within the kinetochore. Maresca and Salmon [12] showed that treating Drosophila melanogaster tissue culture cells with taxol reduces inter-kinetochore but not intra-kinetochore stretch and does not activate the spindle checkpoint. Uchida et al. [13] showed that treating HeLa cells with low nocodazole concentrations reduces intra-kinetochore but not inter-kinetochore stretch and does activate the checkpoint. Our studies agree with the conclusion that the checkpoint responds to events within kinetochores rather than between them: we find that inhibiting chromatin stretch does not activate the checkpoint, and our approach avoids the potential side effects of altering microtubule dynamics with drugs, and isolates chromatin stretch from other effects on spindle structure and dynamics.
Kinetochores can elongate under tension [12]–[14]. In Drosophila S2 cells, unattached kinetochores measure 65±31 nm from the inner centromere protein, CENP-A, to the outer kinetochore protein, Ndc80. When attached and bi-oriented, this distance increases by an average of 37 nm [12]. Kinetochores could elongate by two mechanisms: altering their composition [34] or changing the conformations and contacts of individual proteins. Studies using immuno-electron and fluorescent microscopy showed that inner kinetochore proteins CENP-A, -C, and -R deform under tension, and CENP-T elongates, separating its N- and C-termini [35]. The outer domains of the microtubule-binding Ndc80 complex has also been shown to move 15 nm further away from the inner kinetochore upon bi-orientation [14], perhaps by straightening of a long coiled-coil domain broken by a flexible, elbow-like hinge [36].
Two different mechanisms have been proposed for the link between kinetochore elongation and the activity of Ipl1: relaxing the kinetochore activates Ipl1, or it allows an already activated kinase better access to its substrates. In budding yeast, Bir1 and Sli15 (Survivin and INCENP in higher eukaryotes), members of the chromosomal passenger complex that localize and activate Ipl1, help link centromeres and microtubules [37], [38]. Studies on SLI15 and BIR1 mutants have led to the proposal that these proteins activate Ipl1 on relaxed kinetochores [37]. Recently, it has been shown that Sli15's ability to cluster Ipl1 together rather than its ability to localize the kinase to the centromere may be sufficient for distinguishing between correct and incorrect attachments [39]. There is also evidence supporting a constitutively active kinase that is separated from its substrates when the kinetochore is stretched: the phosphorylation of an Ipl1/Aurora B target depends on its distance from the kinase, located in the inner kinetochore, and repositioning the kinase closer to the outer kinetochore destabilizes microtubule attachments and activates the checkpoint [40].
Our results in yeast corroborate other work arguing that the spindle checkpoint measures the effects of tension within kinetochores. Monitoring the kinetochore means that the checkpoint would not activate in response to the observed variations in the distance between sister chromatids, but would detect mono-oriented chromosomes. Preventing false alarms from a tensiometer at the kinetochore would requires it to have one of two properties to keep the checkpoint from activating as the distance between sister centromeres fluctuates: 1) the extensible element within the kinetochore would have to have a lower spring constant than the linkage between the centromeres to make sure the tensiometer remained stretched, or 2) the conformational change that activated the checkpoint would have to be slower than the variations in the overall force separating the sister centromeres. Distinguishing between these possibilities will require further investigation of kinetochore dynamics and biochemistry.
Strains used in this study are listed in Table 1; all were constructed in W303 (ade2-1 his3-11,15 leu2-3,112 trp1-1 ura3-1 can1-100) using standard genetic techniques. Lactose operator arrays containing 256 repeats of the operator were integrated either upstream of the centromere or on either side of the centromere on Chromosome III. Both arrays were integrated approximately 1500 bp from the centromere. Dimeric control strains contained a C-terminal truncation mutant of the Lac repressor (LacI2) that cannot cross-link two arrays; experimental cells contained the wild-type version of the Lac repressor capable of tetramerizing and cross-linking two arrays (LacI4) [24]. Both versions of the repressor were placed under the HIS3 promoter and were fused via their N-terminus to monomeric yeast optimized GFP. Cells were either grown in Synthetic Complete media (2% glucose) lacking histidine (SC-HIS) or Synthetic Complete media (2% glucose) lacking histidine and methionine (SC-HIS-MET) at 30°C to promote expression of the Lac repressor under the HIS3 promoter. YPD containing 1-(butylcarbamoyl)-2-benzimidazolecarbamate (benomyl) and nocodazole was prepared by heating YPD to 65°C and adding dimethyl sulfoxide (DMSO) 10 mg/ml stocks of benomyl drop-wise to a final concentration of 30 µg/ml; media was cooled to 37°C for drop-wise addition of DMSO 10 mg/ml stock of nocodazole to a final concentration of 30 µg/ml. All drugs and chemicals were purchased from Sigma Aldrich.
Strains were grown in SC-HIS-MET at 30°C and maintained in log phase for 24 hours before the experiment. Log phase cells (∼5×106 cells/ml) were arrested in G1 with 10 µg/ml alpha factor (Bio-Synthesis) for 3 hours. After confirmation of arrest by light microscopy, cells were washed three times with YPD to remove alpha factor and released into SC-HIS media containing methionine (250 µg/ml). Media lacking methionine allows cells to grow, but media containing methionine inhibits expression of Cdc20 from the MET promoter and induces metaphase arrest. Cells were grown at 30°C for 3 hours, and samples were collected every 30 minutes (see Figure 2A). Samples were fixed with formalin (see below) and stored at 4°C for imaging. Using fluorescence microscopy to visualize GFP-tagged chromatids, samples were scored for the presence of one or two GFP dots; two dots indicates stretched chromatids.
Strains were grown in SC-HIS plus 2% raffinose at 23°C and maintained in log phase for 24 hours before the experiment. Log phase cells (∼5×106 cells/ml) were arrested in G1 with 10 µg/ml alpha factor (Bio-Synthesis) for 3 hours at 23°C. Cells were transferred to either SC -HIS+2% galactose+10 µg/ml alpha factor to induce the GAL1 promoter or to SC-HIS+2% glucose+10 µg/ml alpha factor to repress the promoter, and G1 synchronization continued an additional hour at the restrictive temperature (37°C). After confirming the arrest by light microscopy, cells were then washed three times in YEP, and incubated for a further three hours in either SC-HIS+2% glucose or 2% galactose at 37°C. Under these conditions, cells proceed through the cell cycle and arrest at anaphase, as large-budded cells because of the cdc15 mutation (see Figure 3A). Samples were sonicated, fixed with formalin (see below), and stored at 4°C for imaging. Cells were scored for chromosome segregation based the position of the two chromatid copies of GFP-labeled chromosome III. Correct chromosome segregation produces one copy of the chromosome (one GFP dot) in both the mother and daughter cells, whereas incorrect chromosome segregation leads to two GFP dots in a single cell. Anaphase arrest was confirmed by staining fixed cells with ProLong Gold antifade reagent with DAPI (Life Technologies); 100 cells were scored in three independent trials for DNA masses in both mother and daughter cells.
Strains were grown in SC-HIS at 30°C and maintained in log phase for 24 hours before the experiment. Log phase cells (∼5×106 cells/ml) were arrested in G1 with 10 µg/ml alpha factor (Bio-Synthesis) for 3 hours. After confirmation of arrest by light microscopy, cells were washed three times with YPD to remove alpha factor and released into SC-HIS media. Cells were grown at 30°C for 3 hours, and samples were collected every 10 minutes (see Figure 4A). Samples were sonicated, fixed with formalin (see below), and stored at 4°C for imaging. After 60 minutes, 10 µg/ml alpha factor was added to prevent additional entry into a second mitosis during the experiment. Samples were scored for mitotic progression by cell morphology and position of GFP-tagged chromatids. Anaphase was scored as large-budded cells with GFP-tagged chromatids separated into mother and daughter cells.
Strains were grown in SC-HIS at 30°C and maintained in log phase for 24 hours before the experiment. Log phase cells (∼5×106 cells/ml) were arrested in G1 with 10 µg/ml alpha factor (Bio-Synthesis) for 3 hours. After confirming the arrest by light microscopy, cells were washed three times with YPD to remove alpha factor and released into YPD containing 30 µg/mL 1-(butylcarbamoyl)-2-benzimidazolecarbamate (benomyl) and 30 µg/mL nocodazole prepared as described above. In Figure 5B, cells were grown in the drugs at 30°C for 4 hours with samples collected every 60 minutes and scored for the percentage of large-budded cells. In Figure 5D, cells were grown in the drugs at 30°C for 90 minutes then washed three times with YPD and released into drug-free YPD for an additional 60 minutes of growth at 30°C. Samples were taken every 10 minutes post-release from G1, fixed with formalin (see below) and scored for anaphase, identified as large-budded cells with GFP-tagged chromatids separated into mother and daughter cells.
Samples for imaging were fixed with 10% formalin added directly to growth media containing cells (final concentration of 1%), incubated for 10 minutes at room temperature, washed with 0.1M KH2PO4 pH 8.5, washed with 1.2M Sorbitol+0.1M KH2PO4 pH 8.5, resuspended in 1.2M Sorbitol+0.1M KH2PO4 pH 8.5, and stored at 4°C. Images were acquired at room temperature (25°C) using a Nikon Eclipse Ti-E inverted microscope with a 60× Plan Apo VC, 1.4 NA oil objective lens with a Photometrics CoolSNAP HQ camera (Roper Scientific). Metamorph 7.7 (Molecular Devices) was used to acquire images. Fixed samples were imaged in 1.2M Sorbitol+0.1M KH2PO4 pH 8.5 buffer on Concanavalin A-coated coverslips (VWR) adhered to glass slides (Corning). Exposure times were 10 ms for differential interference contrast and 300 ms for fluorescence.
|
10.1371/journal.pcbi.1003198 | Genome-Wide Signatures of Transcription Factor Activity: Connecting Transcription Factors, Disease, and Small Molecules | Identifying transcription factors (TF) involved in producing a genome-wide transcriptional profile is an essential step in building mechanistic model that can explain observed gene expression data. We developed a statistical framework for constructing genome-wide signatures of TF activity, and for using such signatures in the analysis of gene expression data produced by complex transcriptional regulatory programs. Our framework integrates ChIP-seq data and appropriately matched gene expression profiles to identify True REGulatory (TREG) TF-gene interactions. It provides genome-wide quantification of the likelihood of regulatory TF-gene interaction that can be used to either identify regulated genes, or as genome-wide signature of TF activity. To effectively use ChIP-seq data, we introduce a novel statistical model that integrates information from all binding “peaks” within 2 Mb window around a gene's transcription start site (TSS), and provides gene-level binding scores and probabilities of regulatory interaction. In the second step we integrate these binding scores and regulatory probabilities with gene expression data to assess the likelihood of True REGulatory (TREG) TF-gene interactions. We demonstrate the advantages of TREG framework in identifying genes regulated by two TFs with widely different distribution of functional binding events (ERα and E2f1). We also show that TREG signatures of TF activity vastly improve our ability to detect involvement of ERα in producing complex diseases-related transcriptional profiles. Through a large study of disease-related transcriptional signatures and transcriptional signatures of drug activity, we demonstrate that increase in statistical power associated with the use of TREG signatures makes the crucial difference in identifying key targets for treatment, and drugs to use for treatment. All methods are implemented in an open-source R package treg. The package also contains all data used in the analysis including 494 TREG binding profiles based on ENCODE ChIP-seq data. The treg package can be downloaded at http://GenomicsPortals.org.
| Knowing transcription factors (TF) that regulate expression of differentially expressed genes is essential for understanding signaling cascades and regulatory mechanisms that lead to changes in gene expression. We developed methods for constructing gene-level scores (TREG binding scores) measuring likelihood that the gene is regulated based on the generative statistical model of ChIP-seq data for all genes (TREG binding profile). We also developed methods for integrating TREG binding scores with appropriately matched gene expression data to create TREG signatures of the TF activity. We then use TREG binding profiles and TREG signatures to identify TFs involved in the disease-related gene expression profiles. Two main findings of our study are: 1) TREG binding scores derived from ChIP-seq data are more informative than simple alternatives that can be used to summarize ChIP-seq data; and 2) TREG signatures that integrate the binding and gene expression data are more sensitive in detecting evidence of TF regulatory activity than commonly used alternatives. We show that this advantage of TREG signatures can make the difference between being able and not being able to infer TF regulatory activity in complex transcriptional profiles. This increased sensitivity was critically important in establishing connections between disease and drug signatures.
| The specificity of transcriptional initiation in the genomes of eukaryotes is maintained through regulatory programs entailing complex interactions among transcription factors (TF), epigenetic modifications of regulatory DNA regions and associated histones, chromatin-remodeling proteins, and the basal transcriptional machinery [1]. High-throughput sequencing of immuno-precipitated DNA fragments (ChIP-seq) provides means to assess genome-wide expression regulatory events, such as TF-DNA interactions [2]. Sophisticated statistical methodologies have been developed for identifying TF binding events in terms of “peaks” in the distributions of ChIP-seq data [3]–[8]. The evidence provided by ChIP-seq binding data that a gene's expression is regulated by a TF is a function of the number of peaks, their intensity and proximity to the transcription start site (TSS) [9]. Furthermore, binding of a transcription factor in a gene's promoter alone does not always result in transcriptional regulation. In the case of highly studied pleiotropic regulator ERα, transcriptional regulation depends on the presence of specific co-factors as well as on the type of activating ligand [10], [11]. Therefore, the identification of true regulatory TF-gene relationships requires per-gene summaries/scores measuring the totality of the evidence in ChIP-seq data, integrated with measurements of gene expression levels.
Current approaches to summarizing binding peaks in order to correlate TF binding with transcriptional changes range from simple summaries in proximal gene promoter (e.g. maximum peak height within a narrow region around the promoter) [12]–[14] to weighted sums of peak heights where weights are inversely proportional to the distance of the peak to the gene's TSS [9], [15]. Currently used distance-based weights are dependent on TF-specific tuning constants established through ad-hoc examination of the distribution of the peaks [9], [12], [13].
Dysregulation of transcriptional programs is intimately related to the progression of cancer [16], [17] and other human diseases [18], [19]. Modulating the behavior of specific TFs is a popular strategy for developing new disease treatments [20]–[23]. Genome-wide transcriptional profiles associated with a disease phenotype provide indirect evidence of TF involvement in the etiology of the disease. The most common strategy of implicating TF involvement is by computational analysis of genomic regulatory regions of differentially expressed genes [24]–[27]. However, such strategies are not effective when the search needs to include distant enhancers and when concurrent activity of multiple regulatory programs lead to “messy” transcriptional signatures. ERα-driven proliferation is one such case where the involvement of ERα regulatory program has been difficult to identify in resulting transcriptional profiles using the DNA binding motif analysis [27].
We have developed a comprehensive statistical framework for assessing True REGulatory (TREG) TF-gene interactions by integrated analysis of ChIP-seq and gene expression data. In the first step we introduce a novel two-stage mixture generative statistical model for summarizing “peaks” within 2MB window centered around a gene's TSS. Fitting this two-stage model yields scores and associated probabilities of regulation based on ChIP-seq data alone (ie TREG binding profile). We show that our approach produces effective summaries for a TF with binding sites clustered in close proximity of TSS (E2F1) and a TF known to exhibit regulation through binding to distant enhancers (ERα).
In the second step we integrate the TREG binding profile with a differential gene expression profile to create an integrated TREG signature of TF regulatory activity. We use TREG signatures to detect faint signals of ERα regulation in “messy” transcriptional signature, and demonstrate how such analysis can yield better drug candidates than simply correlating transcriptional signatures of the disease and the drug activity [28]–[30].
An overview of the TREG framework is shown in Fig. 1. We start with “peaks” extracted from ChIP-seq binding data and differential gene expression profile that eventually yield the integrated TREG signature of TF activity (Fig. 1A). The foundation of the TREG framework consists of two statistical mixture modules. The first mixture model describes the distribution of functional and non-functional “peaks” in ChIP-seq TF-gene binding data (Fig. 1B). Based on this model, we derive the TF-specific distance weights and construct gene-level binding scores (TREG binding scores) measuring the likelihood that a gene is regulated by the given TF. The second mixture model describes the distribution of TREG binding scores for regulated and non-regulated genes (Fig. 1C). This second model provides us with gene-level probabilities that genes are regulated by a specific TF based on the ChIP-seq data alone. TREG binding scores and associated gene-level probabilities for all genes make up the TREG binding profile. The TREG binding profile and differential gene expression profiles are integrated using Generalized Random Set (GRS) methodology [31] to produce an integrated genome-wide TREG signature of the TF activity (Fig. 1D). The TREG signature of ERα is used to demonstrate involvement of its regulatory activity in complex transcriptional profiles and to mine Connectivity Map Data for inhibitors of its activity.
We assume that observed peaks consist of two populations: Functional peaks that are more likely to occur closer to TSS and whose distance to TSS is distributed as an exponential random variable; and, Non-functional peaks that are randomly occurring throughout the 2 million base pair genomic region centered around the TSS, and whose distances to TSS are distributed as a uniform random variable. The distances to TSS of all peaks are then distributed as a mixture of the exponential and the uniform distribution (Fig. 1, Eq1), where π is the proportion of functional peaks among all observed peaks. We define the TREG binding score for gene g as the logarithm of the weighted average of peak intensities, using the probability of the peak belonging to the population of “functional peak” as weights (Fig. 1 Eq3).
We assessed the effectiveness of the TREG binding score by comparison to the simple scoring method based on the maximum peak intensity (MPI) within a window of specific size around TSS. The two types of scores were evaluated by comparing the enrichment of genes with high evidence of TF binding among genes differentially expressed in appropriately matched experiments. For gene expression data, we identified genes differentially expressed (two-tailed FDR<0.01) 24 h after treating MCF-7 cell line with estradiol (E2) with and without pre-treating the cell line with Cycloheximide (CHX) [27]. CHX is an inhibitor of protein biosynthesis in eukaryotic organisms. Treatment with E2 after pre-treatment with CHX (E2+CHX) resulted in differential expression of genes presumed to be directly regulated by ERα; whereas after E2 treatment without CHX, the majority of differentially expressed genes were secondary target genes functionally enriched for cell-cycle genes and reflective of the rapid proliferation resulting from the E2 treatment [27]. For the TF binding data, we used ChIP-seq analysis of the key proliferation regulator E2f1 in growing mouse embryonic stem (ES) cells [32], and ERα binding 1 h after treating MCF-7 cells with estradiol [10]. ChIP-seq data at 1 h hour after treatment with E2 is correlated with gene expression changes 24 h after treatment because of the expected time-delay between ERα binding to a gene promoter and the observable change in the gene's expression level.
Among differentially expressed genes, enrichment of genes with high TREG binding scores was statistically significant for both E2F1 and ERα in both experiments (Table 1). Fig. 2 shows the relative levels of enrichment for maximum peak intensity (MPI) score over the range of window sizes around TSSs in comparison to the TREG binding score. Simple MPI scores never attain the level of statistical significance of enrichment attained by TREG binding scores. Furthermore, the performance of the simple score is heavily dependent on the specific size of the window used, and expectedly, the optimal windows are TF–specific. The optimal window size for E2f1 and ERα is around 1 kb and 50 kb respectively, with maximum statistical significance of enrichment attained for the simple score reaching 42% and 80% of the TREG binding score significance, respectively. Similar results were obtained using unweighted sum and linear-weighted sum of TF binding peak intensity scores (supplementary results in Text S1 and Fig. S1). This indicates that TREG binding scores not only provide the best correlation with expression changes, but they also obviate the need of knowing the right window size to use in deriving the summary measure of TF binding. The calculation of TREG binding scores does not include any free parameters that need to be specified in ad-hoc fashion, such as the length of the genomic region around TSS for simple scores, or the ad-hoc weighting parameters used in similar scores before [9], [15].
Having constructed gene-specific TREG binding score, our goal was to estimate gene-level probabilities of “functional interaction” between a TF and a gene based on these scores. The histogram of the TREG binding scores (Fig. 1C) clearly shows two populations of TREG binding scores. One population with a majority of TREG binding scores being close to zero, representing genes with low likelihood of functional TF-gene interaction, and the other populations with TREG binding scores distributed in bell-shaped form around the mean slightly higher than 2, representing functional interactions. Therefore, we assume that TREG binding scores come from two populations: Scores significantly greater than zero representing functional TF-gene interactions which are distributed as a Normal random variable; and, scores close to zero representing non-functional interactions which are distributed as an exponential random variable. Assuming that the proportion of TREG binding scores corresponding to functional interactions is η, the distribution of all TREG binding scores is a mixture of Normal and exponential probability distribution functions (Fig. 1 Eq4). The probability that a TREG binding score for gene g (Sg) is functional is defined as the probability of Sg belonging to the normal component (Fig. 1 Eq5). The set of TREG binding scores and associated probabilities of the score indicated functional TF-gene interaction for all genes in the genome (Sg, pg), g = 1,…,G, is the TREG binding profile.
Identifying genes that both have high probability of “functional” TF binding and are differentially expressed is complicated by the need to set arbitrary thresholds for statistical significance. We have previously developed a method, based on the Generalized Random Set (GRS) analysis that obviates the need for such thresholds when assessing concordance of two differential gene expression profiles [31]. Here we apply the GRS framework to assess the concordance between the TREG binding profile and the differential gene expression profile (Fig. 1 Eq6) (details in Text S1), and to identify genes with statistically significant concordance. The results (Table 2) of the analysis generally followed the results based on designating differentially expressed genes (Table 1) with the levels of statistical significance being orders of magnitude higher in the GRS concordance analysis. We demonstrate that GRS is producing expected distribution of p-values under the null hypothesis by systematically examining empirical cumulative distribution functions (ECDFs) of p-values after randomly permuting gene labels in TREG binding profile before GRS analysis (supplementary results in Text S1, Fig. S2). We also compared the results of GRS analysis with the thresholding approach based on TREG binding probability where gene was placed in the “regulated” group if the corresponding TREG probability (pg) was greater than 0.95. Results were similar to the GRS analysis (supplemental results Text S1). However, we also show that in the situations when binding signal is relatively “faint”, GRS is likely to outperform thresholding approach (Text S1, Fig. S3). Since these are situations in which the method of concordance analysis will make the difference, the GRS is still likely the better default choice for performing the concordance analysis.
Finally, we integrate at the gene level TREG binding profiles with differential gene expression profiles as the contribution of an individual gene to the overall concordance in the GRS concordance statistics eg (Fig. 1 Eq7). The statistical significance of gene-level GRS statistics is assessed by associated resampling-based p-values (see methods) which define gene-specific TREG concordance scores (tg, Fig. 1, Eq8). The vector of such scores for all genes represents the TREG signature of TF activity (Fig1 Eq9).
We examined the ability of TREG binding profiles and TREG signatures to identify genes regulated by ERα and E2f1. Fig. 3A contrasts the statistical significance of the enrichment by the computationally predicted ERα targets from MSigDB database [33] based on E2+CHX differential gene expression profile (Diff Exp), ERα TREG binding scores (TREG bind) and integrated TREG signature (TREG sig). In this setting, MSigDB targets provide a “noisy” gold standard since the perfect gold standard does not exist. While all three data types provided statistically significant enrichment, the integrated TREG signature showed the highest statistical significance of the enrichment. The overall relationship between the TREG binding scores, statistical significance of differential gene expression (−log10(p-value) E2+CHX) and the statistical significance of TREG concordance scores (ERα TREG score (sg)) is shown in Fig. 3B. The “statistically significant” (p-value<0.001) TREG concordance scores (red dots in Fig. 3B) required both, a high TREG binding score and a high statistical significance of differential expression. Similar analysis of the E2f1 TREG signature showed a similar pattern (Fig. 3C and D), although the overall statistical significance of enrichment was much higher for all three data types. These results show that integrated TREG signatures are more informative of the regulatory TF-gene relationships than expression or TF binding data alone. TREG binding scores, gene specific concordance statistic, and TREG concordance scores for all genes are given in the Table S1.
We further examined ERα and E2F1 TREG signatures to determine molecular pathways and biological processes regulated by these two TFs and to evaluate benefits of such integrated signatures. We assessed the enrichment of genes with high TREG concordance scores in lists of genes related to the prototypical function of ERα and E2F1. For the ERα signature the list consisted of genes associated with the Gene Ontology term “cellular response to estrogen stimulus”, and for the E2F1 with the term “regulation of mitotic cell cycle”. In both cases, integrated TREG signatures showed significantly higher statistical significance of enrichment than either TREG binding scores or differential gene expressions (Fig. 4). Unsupervised enrichment analysis of the two signatures revealed that biological processes specifically associated with ERα signature were related to the development of the mammary gland (Fig. 5A). Moreover, significant associations between ERα-regulated genes and some key developmental processes could not have been established using either TF binding or gene expression alone. Likewise, processes related to mitotic cell cycle were most highly associated with E2f1 signature (Fig. 5B). Results of enrichment analysis for all GO terms are provided in Table S2.
To assess the reproducibility and specificity of our results, we constructed TREG binding signatures for all 494 TF ChIP-seq datasets in the Genome Browser ENCODE tables [14], [34]. Two gene expression profiles in our analysis (E2+CHX and E2) were then systematically compared with 494 ENCODE TREG binding profiles. Top 10 most concordant profiles are shown in Fig. 6. Results show that ENCODE ERα binding profiles correlates equally well with E2+CHX profile as did our original TREG profile (Fig. 6A). Furthermore, all five ENCODE ERα binding profiles correlated better with E2+CHX profile than any other ENCODE profile. Similarly, ENCODE binding signatures most concordant with E2 profile (Fig. 6B) included E2F4, E2F1 and MYC which are all known to be important cell cycle regulators. The statistical significance of the concordance was again similar to the levels we observed with the E2f1 binding profile in mouse embryonic stem cells. These results indicate that reproducibility of TREG results across different ChIP-seq datasets and its ability to identify key transcriptional regulators for a given profile. Results of the concordance analysis for all ENCODE TREG profiles are in Table S3.
The ultimate goal of the TREG framework is to facilitate identification and characterization of signatures of TFs regulating disease-related differential gene expression profiles (DRGEP). Here we demonstrate the power of TREG signatures and TREG binding scores in elucidating the faint signals of ERα activity in two complex DRGEPs, the response of MCF-7 cell line 24 hours after treatment with E2 [27] and differences between ER− and ER+ breast tumors [35]. In both of these DRGEPs, the signal of direct ERα regulation is “drowned out” by the strong secondary proliferation-related transcriptional signature, and the standard enrichment analysis of computationally predicted ERα targets in MSigDb fails to find evidence of ERα regulation (Fig. 7). However, the GRS concordance analysis with both TREG binding scores and TREG signatures are highly statistically significant, and the TREG signature which integrated binding and transcriptional evidence again shows the highest statistical significance of concordance (Fig. 7). Additional discussion of these results is provided in supplementary results (Text S1).
We used the ERα TREG signature to mine a collection of differential gene expression profiles in GEO datasets (GDS signatures), and differential gene expression profiles of small drug perturbations (CMAP signatures) [29], for evidence of ERα regulatory activity. Fig. 8 shows differential gene expression levels of top 10 GEO profiles and top 10 drug perturbations based on the statistical significance of the concordance between the ERα TREG signature and each differential gene expression profiles. In both situations the top transcriptional profiles are obviously related to the ERα activity demonstrating the precision of the TREG signature in this setting. Additional results related more specifically to disease-associated GEO profiles are given in the supplementary results (Text S1).
The problem of identifying functional TF targets that regulate gene expression, in a specific biological context, requires joint considerations of both TF DNA-binding data and the target gene's expression changes. We described a statistical framework for quantifying the evidence of TF-gene interaction from ChIP-seq data, and integrating them appropriate gene expression data to construct genome-wide signatures of TF activity.
Two main findings of our study are that 1) TREG binding scores derived from ChIP-seq data alone are more informative than simple alternatives that can be used to summarize ChIP-seq data; and 2) TREG signatures that integrate the binding and gene expression data are more sensitive in detecting evidence of TF regulatory activity than available alternatives. We show that this advantage of TREG signatures can make the difference between being able and not being able to infer TF regulatory activity in complex transcriptional profiles. This increased sensitivity also showed to be critical in establishing connections between disease and drug signatures that would not be possible using currently available strategies.
Identifying the role of specific TFs in producing disease-related transcriptional profiles is of vital importance for understanding the molecular mechanisms underlying disease phenotype. Although it is possible to obtain direct measurements of TF activity in disease samples [45], such ChIP-seq profiling is technically challenging and systematic profiling of many different TFs is not feasible. Therefore, the ability to infer the role of a TF from the transcriptional profiles remains challenging. The most common strategy of implicating TF involvement is by computational analysis of genomics regulatory regions of differentially expressed genes [24]–[27], or by searching for enrichment of known targets among differentially expressed genes [46]. Here we present an alternative strategy relying on direct concordance analysis between TREG signatures of TF activity and disease-related transcriptional profiles. When searching for evidence of regulation by the TF with functional binding sites in distant enhancers, such as ERα, and “messy” transcriptional signatures resulting from activity of multiple regulatory programs, our approach dramatically improves the precision of the analysis.
Our results indicate that TREG signatures derived from in-vitro experiments (ERα; MCF-7 cells), and even from a different organism (E2f1; mouse) provide effective means for analyzing transcriptional profiles derived from human tissue samples. This would indicate that TF binding profiles coming from any biological system under which TF shows signs of activity might be sufficiently informative to construct TREG signatures. In this context the recently released ENCODE project data [14], [34] may be turned into a powerful tool for detecting TF activity. As a step in this direction, we have created 494 TREG binding profiles using the ENCODE ChIP-seq data and made it available from the support web-site (http://GenomicsPortals.org). Complementary gene expression data generated by directly perturbing specific TFs, such as shRNA knock-downs and overexpression experiments can be used to construct TREG signatures. For example, transcriptional signatures of such systematic perturbations that is being generated by NIH LINCS project (http://LincsProject.org) could provide complementary transcriptional profiles for ENCODE ChIP-seq data.
Our methods are complementary to methods used to analyze the recently released ENCODE project data [14], [34]. For some experimental conditions, the ENCODE project provides additional data types that can be used in assessing the functionality of TF binding peaks, such as distribution of specific epigenetic histone modifications. For discussion on how to possibly incorporate this additional information within TREG methodology, please see supplemental discussion (Text S1).
Up-regulated expression of proliferation genes is a hallmark of neoplastic transformation and progression in a whole array of different human cancers [47]. While the core transcriptional signature of proliferation is recognizable in a wide range of biological systems and diseases, the events and pathways that drive the transcriptional program of proliferation vary widely. Increased expression of proliferation-associated genes has been associated with poor outcomes in breast cancer patients [48]–[54]. However, the driver mechanisms in many aggressive cancer types are poorly understood. Inhibiting known driver pathways, such as ERα signaling in breast cancer often leads to treatment resistant tumors due to activation of alternative, poorly understood driver pathways [55], [56]. Using the signatures of such “driver events/pathways” we can identify candidate drugs capable of inhibiting them. In our analysis of ERα activity in ER+ breast cancers we showed that such an approach can highlight connections between disease and drug candidates that would be missed by simply correlating disease and drug transcriptional signatures [28]–[30].
We assume that observed peaks consist of two populations: Functional peaks that are more likely to occur closer to TSS and whose distance to TSS is distributed as an exponential random variable with the parameter λ; and, non-functional peaks that are randomly occurring throughout the 2 million base pair genomic region centered around the TSS, and whose distances to TSS are distributed as a uniform random variable. The distances to TSS of all peaks are then distributed as a mixture of the exponential and the uniform distribution (Fig. 1, Eq1), where π is the proportion of functional peaks among all observed peaks, a is the distance of a peak to the gene's TSS, is the probability density function (pdf) of the exponential random variable (rv) with the location parameter λ, and is the pdf of a uniform rv on the interval (−106, 106). We use the standard Expectation-Maximization (EM) algorithm [57] to estimate the parameters of this mixture model (π,λ) for each TF. Given the estimates we calculate the posterior probability for peak i with distance ai from a TSS to belong to the population of “functional peaks” (Fig. 1 Eq2). Suppose now that for a gene g, ng is the number of peaks within the 1MB window around its TSS (1MB upstream to 1MB downstream), is the peak intensity (ie, the maximum number of overlapping reads over all positions within the peak), and is the distance to TSS of the kth such peak (k = 1,…,ng). We define the TREG binding score for the gene g as the logarithm of weighted average of peak intensities, using the probability of the peak belonging to the population of “functional peak” () as the weight (Fig. 1 Eq3).
We assume that TREG binding scores come from two populations: Scores significantly greater than zero representing functional TF-gene interactions which are distributed as a Normal random variable; and, scores close to zero representing non-functional interactions which are distributed as an exponential random variable (histogram in Fig. 1B). Assuming that the proportion of TREG binding scores corresponding to functional interactions is η, the distribution of all TREG binding scores is a mixture of Normal and exponential probability distribution functions (Fig. 1 Eq4), where S is the TREG binding score, is pdf of the exponential random variable with the location parameter ψ, and is the pdf of a Normal random variable with mean μ and variance σ2. We again use the standard EM algorithm to estimate the parameters of this mixture model (η, ψ, μ, σ2) for each TF. Given the estimates , the probability of a TREG binding score for gene g (Sg) being functional is defined as the probability of Sg belonging to the normal component (Fig. 1 Eq5). The set of TREG binding scores and associated probabilities of the score indicated functional TF-gene interaction for all gene in the genome (Sg, pg), g = 1,…,G, is the TREG binding profile. Additional discussion of motivations for the choice of specific distributions is provided in supplemental methods (Text S1).
Details of the EM algorithm are provided in supplemental methods (Text S1).
The enrichment of genes with high TREG and MPI scores among differentially expressed genes (Table 1, Fig. 2) was performed using the logistic regression-based LRpath methodology [58]. LRpath does not require thresholding on binding scores but uses such scores as the continuous variable that explains the membership of a gene in the “differentially expressed” category. Similarly, LRpath was used to analyze enrichment of differentially expressed genes among genes associated with GO terms in Fig. 5 and 6.
When performing concordance analysis between TREG binding profiles and the two differential gene expression profiles of interest (E2+CHX and E2) (Table 2) and constructing TREG signatures in Fig. 4,5, and 6, we used two-tailed p-values not distinguishing between induction and repression activity. When comparing TREG signatures with other DRGEPs (Table 3, and Fig. 7 and 8), we account for directionality of gene expression changes by using single-tailed p-values for increase in gene expression. This is necessary to account for the directionality of the concordance between the TREG signature and the DRGEPs. The ERα TREG signatures for this analysis was constructed by the GRS concordance analysis (Fig. 1D) between ERα TREG binding profile and the single tailed p-values for statistically significant up-regulation of gene expression after E2+CHX treatment of MCF-7 cell line. The genes used for plotting heatmaps in Fig. 8 were then selected based on the gene-specific p-values of concordance (p-value(eg), Fig. 1D) being <0.001 (Table S5). The concordance between this ERα TREG signature, and GEO/CMAP transcriptional signatures was performed again using the GRS analysis.
The description, location and processing of the ChIP-seq and gene expression datasets are provided in supplemental methods (Text S1).
All computational methods are implemented in the R package treg which can be downloaded from our web site (http://GenomicsPortals.org). The package also contains processed ChIP-seq data for ERα [10], E2f1 and 15 other transcription factors [32], as well as TREG signatures for ERα and E2f1, and transcriptional signatures derived from GEO GDS datasets and CMAP drug signatures. We have previously described derivation of CMAP signatures [31]. All functional enrichment analyses were performed using the LRpath methodology [58] as implemented in the R package CLEAN [59].
|
10.1371/journal.ppat.1002923 | Structure and Assembly of a Trans-Periplasmic Channel for Type IV Pili in Neisseria meningitidis | Type IV pili are polymeric fibers which protrude from the cell surface and play a critical role in adhesion and invasion by pathogenic bacteria. The secretion of pili across the periplasm and outer membrane is mediated by a specialized secretin protein, PilQ, but the way in which this large channel is formed is unknown. Using NMR, we derived the structures of the periplasmic domains from N. meningitidis PilQ: the N-terminus is shown to consist of two β-domains, which are unique to the type IV pilus-dependent secretins. The structure of the second β-domain revealed an eight-stranded β-sandwich structure which is a novel variant of the HSP20-like fold. The central part of PilQ consists of two α/β fold domains: the structure of the first of these is similar to domains from other secretins, but with an additional α-helix which links it to the second α/β domain. We also determined the structure of the entire PilQ dodecamer by cryoelectron microscopy: it forms a cage-like structure, enclosing a cavity which is approximately 55 Å in internal diameter at its largest extent. Specific regions were identified in the density map which corresponded to the individual PilQ domains: this allowed us to dock them into the cryoelectron microscopy density map, and hence reconstruct the entire PilQ assembly which spans the periplasm. We also show that the C-terminal domain from the lipoprotein PilP, which is essential for pilus assembly, binds specifically to the first α/β domain in PilQ and use NMR chemical shift mapping to generate a model for the PilP:PilQ complex. We conclude that passage of the pilus fiber requires disassembly of both the membrane-spanning and the β-domain regions in PilQ, and that PilP plays an important role in stabilising the PilQ assembly during secretion, through its anchorage in the inner membrane.
| Many bacteria which cause infectious disease in humans use large fibers, called pili, to attach to the surfaces of the cells of the host. Pili are also involved in a particular type of movement of bacteria, termed twitching motility, and the uptake of DNA into the bacterial cell. They are made up of thousands of copies of a specific pilin protein. The process of assembly of pili is complicated: it requires the cooperative action of a group of proteins which span both the inner and outer membranes in bacteria. Here we have determined the structure of part of the machinery which forms a channel between both membranes. One of the proteins, PilQ, is organized in a segmental way, being divided into separate domains which are jointed, hence allowing them to move relative to each other. We infer that this movement is critical to the functioning of the channel, which must open up to allow passage of the pilus fiber. We suggest that the function of the other protein we have studied, PilP, is to maintain the PilQ assembly during pilus secretion.
| Type IV pili are long (1–5 µm), mechanically strong polymers which extend from the surfaces of many Gram-negative bacteria, including Neisseria meningitidis, Pseudomonas aeruginosa and Vibrio cholerae [1], [2]. They are known to mediate a variety of functions, including attachment to host cell surface receptors during infection [3], natural DNA competence [4] and a phenomenon termed twitching motility, a flagellum-independent process which enables some bacteria to move rapidly (1 µm/s−1) across surfaces [5]. The pilus fiber consists principally of subunits of pilin (PilE in N. meningitidis), a small protein which adopts an α/β fold and assembles into a helical structure which confers mechanical strength on the assembly [6], [7], [8]. Twitching motility is associated with a notable feature of type IV pili: an ability to retract rapidly at a rate of approximately 1,000 pilin subunits per second, generating a powerful mechanical force which has been measured at up to 100 pN per fiber [9], [10].
The secretins are a large and diverse family of integral outer membrane (OM) proteins which comprise key components of the type II and type III secretion systems, as well as the biogenesis systems for type IV pili and filamentous bacteriophage [11]. Three-dimensional reconstructions of secretin structure by electron microscopy have revealed that they adopt multimeric structures, characterized by the formation of large chambers which lie within the periplasm. Our previous work on PilQ from Neisseria meningitidis showed a dodecameric structure, with a chamber sealed at both ends [12]. Studies on the type II secretion system (T2SS) secretins PulD [13] and, more recently VcGspD which is responsible for the secretion of V. cholerae toxin, revealed a cylindrical-shaped structure with 12-fold symmetry enclosing a large chamber which is open at the periplasmic end but closed at the OM [14]. The structure of a type III secretion system (T3SS) secretin can also be extracted from the 10 Å resolution cryoelectron microscopy density map of the Salmonella needle complex: this shows the secretin in an open state, with the needle passing through both ends of the chamber [15].
Figure 1A shows a schematic illustration of the domain structure of N. meningitidis PilQ and two prototypical T2SS and T3SS secretins. All share a well conserved C-terminal region which spans the membrane and is responsible for oligomerization [13], [16], [17], [18], [19]. The central and N-terminal regions are more diverse; crystal structures of the N0, N1 and N2 domains from the T2SS and T3SS secretins have been reported, GspD [20] and EscC [21]. The structure of each domain is well conserved, and is based on a core fold of two α-helices packed against a three-stranded β-sheet. Docking of a model based on the N0/N1/N2 GspD crystal structure into the VcGspD cryoelectron microscopy electron density map established that these domains extend into the periplasm and form the sides of the secretin chamber [14].
A number of proteins are known to interact with secretins, either for the purposes of assembly, OM insertion or mediation of function once the mature protein has been formed. Pilotin proteins are responsible for membrane targeting of secretins: the interaction sites of some have been mapped to the extreme secretin C-terminus, and their recognition of T2SS and T3SS secretins has recently been revealed at the structural level [22], [23]. At least two proteins, PilW and Omp85, are known to promote assembly of the PilQ oligomer [24], [25]. Other proteins seem to play a more direct structural role: PilP is a lipoprotein which binds to PilQ and is essential for type IV pilus (TFP) formation [26], [27], [28]. It has an N-terminal lipid attachment site, followed by an unstructured N-terminal region and a C-terminal globular domain which adopts a lipochalin-like β-structure [29], [30] (Figure 1B). It has been shown to be located in the inner membrane [27]. Recent evidence has established that the fold of the PilP C-terminal domain is similar to that adopted by the HR domain from the T2SS GspC protein, which is known to bind to its cognate secretin, GspD [31]. A crystal structure of a complex between the GspC HR domain and the N0/N1 domains from GspD revealed a binding site formed by the edge-on association of β-strands from GspC and the GspD N0 domain [31]. PilP is among a group of ‘core’ proteins which are essential for assembly of TFP in Neisseria meningitidis [26]. Recent evidence from studies in Pseudomonas has shown that PilP also binds to PilO and PilN, two integral inner membrane proteins which are essential for pilus formation [30]. PilP thus forms a link between the OM, through its interaction with PilQ, and the inner membrane components of the type pilus biogenesis system. Unlike some other bacterial secretion systems, however, there is currently little structural information on the way in which the TFP biogenesis proteins assemble.
Here we report the structural determination of the PilQ periplasmic domains by using a combination of NMR and homology modelling. The original reconstruction of the PilQ oligomer which we reported was generated using cryonegative stain [12]; whilst this served to define the overall dimensions and structure of the complex, it cannot reliably be used for automated docking of constituent domains into the density map. We therefore also report a new 3D reconstruction of the PilQ oligomer, generated by single particle averaging from cryoelectron microscopy data of unstained specimens, and use this to dock the domain structures and generate the dodecameric assembly. Finally, we use a combination of NMR chemical shift perturbations and modelling to generate the complex formed between the first α/β domain in PilQ and the C-terminal domain of PilP. We propose that the segmental organization of the domain structure within PilQ is intrinsic to its ability to open up and form a channel to allow entry of the pilus fiber into the chamber, and its subsequent passage across the periplasm and OM.
Bioinformatic studies suggested that the N-terminal regions of TFP-dependent secretins generally contained one or two putative domains, predicted to be rich in β-sheet and characteristically different from the α/β domains observed in T2SS and T3SS secretins [32]. We therefore adopted a cloning and expression strategy which over-produced these β-domains from TFP-dependent secretins originating from a number of different Gram-negative bacteria, including N. meningitidis, P. aeruginosa, Aeromonas hydrophila, Xanthomonas campestris and Xylella fastidiosa. We generally found the B2 domain more amenable to over-production and purification than B1 (Figure 1A), and obtained good quality NMR spectra from a construct spanning residues 224 to 329 in N. meningitidis PilQ (B2PilQ224–329; Figure 1A). NMR spectra of the 13C/15N uniformly labelled sample exhibited characteristic shifts of a well-folded predominantly β-strand structure, confirmed by 1H, 13C and 15N assignment of native sequence (92.3% complete). The solution structure of the second β-domain revealed an eight-stranded β-sandwich structure which is a novel variant of the HSP20-like fold (Figure 2A). The most similar fold identified within the SCOP database [33] is the CS domain from the human Sgt1 kinetochore complex [34]. The β-domain fold is larger, however, and includes two additional β-strands, such that β5 is paired with β6, rather than β4, as is the case with the CS domain (Figure S1).
A comparison of the sequences of the second β-domains from PilQ in different Gram-negative bacteria revealed a high degree of conservation within the region between β4 and β5, including the short β5′ strand (Figure S2). This observation was highlighted by application of the program CONSURF [35], which maps sequence conservation on to protein structure; in this case sequences from 63 different TFP-dependent secretins are mapped on to the surface of the B2 domain (Figure 2B). Strikingly, the most highly conserved residues map to a single patch on the domain surface, incorporating Lys232 from β1 with Asp281 and Phe282 from the β4/β5 loop. The implication is that this patch forms a binding site, possibly to another unidentified TFP biogenesis protein.
In contrast to the B2 domain, attempts to over-produce the B1 domains from several sources generally met with limited success: protein products were either produced in low yield and/or exhibited poor stability. The best progress was made with the B1 domain from Aeromonas hydrophila: assignment of the NMR spectra and use of chemical shift indices show that the A. hydrophila B1 domain consists of nine β-strands (Figure S3). The poor stability of this single domain precluded the collection of the high quality NOEs required for structural determination. Nevertheless, the similarities in secondary structure between the B1 and B2 domains determined by the NMR chemical shift indices suggest that they share a common origin, as seems to be the case with the repeated N0/N1/N2 domains within the N-terminal sections of the T2SS and T3SS secretins [11]. Most TFP-dependent secretins contain two β-domains, although the first β-domain is missing from some (eg Xylella fastidiosa). It is noteworthy that residues which are highly conserved in the B2 domain (Figure 2B) are not found to be so in the B1 domain and vice versa. In addition, an interesting variation in neisserial PilQ is the presence of low complexity repeat sequences, termed small basic repeats (SBRs), which lie between the B1 and B2 domains and have been shown to influence the efficiency of TFP formation [36]. The presence of such polymorphic repeat elements is unprecedented within the secretin family. As we show below, electron density within the cryoelectron microscopy map for the whole PilQ oligomer cannot accommodate 12 copies of the B1 domain if it folds into a compact, globular structure similar to the B2 domain, so it may be the case that the B1 domains adopt a partially unfolded state in the assembled oligomer.
Secondary structure predictions and sequence alignments suggested the existence of two domains which are likely to adopt a variant of the α/β-type fold identified in other secretins [20], [21]. In a similar approach to that employed for the β-domains, single and multiple domain fragments from different bacteria were over-produced, purified and analysed by NMR. A two domain fragment from N. meningitidis, N0N1PilQ343–545 (Figure 1A), exhibited well dispersed NMR spectra: it was subsequently assigned and its secondary structure determined (Figure S4). Both the N0 and N1 domains are folded, but N1 contains a long random coil extension of over 35 amino acids at its C-terminal end. The very intense peaks from this region obscured many of the peaks from the folded domain of N1 and precluded extraction of the high quality NOEs required for a complete structure determination of the N0/N1 tandem domains. Using the Chemical Shift Index (CSI) information as a marker for the domain boundaries, a smaller fragment was produced which encompassed only the first domain (N0PilQ343–442) and its NMR structure determined by conventional methods using NOE restraints. The high quality structure adopts a fold similar to the N0 domains identified from GspD and EscC [20], [21] (Figure 3A; Table 1). Comparison of the spectra from the single and double domain protein samples verified that the chemical shifts from common residues in the first domain are very similar in both samples (not shown). A striking and novel feature of the domain structure is the presence of an α-helix at the C-terminus of this domain (circled in Figure 3A): from sequence alignments, this appears to be a general feature of the TFP-dependent secretins and is absent from other secretin types. The structure of the N1 domain was constructed using the CSI data, CS-ROSETTA and homology modeling, based on the crystal structure of the same domain from EscC [21] (Figure 3B).
Analysis of the 15N-1H residual dipolar couplings (RDCs) indicated that the N0 and N1 domains have no fixed orientation relative to each other in solution: it was therefore not possible to obtain a common orientation in the alignment tensor frame for the N0 and N1 domains from the RDC measurements. However, the rotation correlation times, calculated from the 15N T1 and T2 values obtained separately for the single N0 domain (τc∼9.6 ns) and the N0/N1 double domain (τc∼14 ns), suggest that the N0 and N1 domains do not tumble completely independently. It is likely that the helical part of the linker between the two domains reduces the flexibility in this region. We therefore generated 100 structures of the N0/N1 double domain using CS-ROSETTA [37], with varying inter-domain orientations. The relevant section of the cryoelectron microscopy density map was then used to identify the cluster of structures which gave the best fit, as well as satisfying other constraints (see below). Interestingly, the relative orientation of the PilQ N0 and N1 domains bears a closer similarity to that observed in the T3SS secretin EscC [21], rather than the T2SS secretin GspD [20]. Clearly, crystal packing constraints and other factors can also influence relative domain orientations. Nevertheless, our observations do lend weight to the idea that the flexibility of the N0/N1 secretin domains could be an integral part of their function.
PilP77–164 is a recombinant fragment which corresponds to the C-terminal domain of the PilP lipoprotein (Figure 1B). Titration of unlabelled N0N1PilQ343–545 into 15N-labelled PilP77–164 identified, from the chemical shift changes [29], [38], a patch of residues on the PilP domain surface involved in binding. These were concentrated mainly into an area around the β1–β2 hairpin in the PilP77–164 structure (Figure 4A). The reverse experiment, where unlabelled PilP77–164 was titrated into 15N-labelled N0N1PilQ343–545, demonstrated that it is the N0 domain, rather than the N1 domain, which is involved in recognition of PilP77–164 (Figure S5). The experiment was repeated using the single N0 domain, N0PilQ343–442, and similar results were obtained. The largest chemical shift attenuations mapped to one side of the structure, concentrated around the first α-helix and β-strand in the fold of PilQ (Figure 4B). Similar experiments titrating PilP77–164 into the B1B2PilQ24–329 and B2PilQ224–329 domains did not show any evidence of binding (not shown).
The identified residues involved in binding on the surface of each protein were used as input into the restraint-driven docking programme HADDOCK [39], [40] to generate a structural model for the PilP77–164:N0N1PilQ343–545 complex. The largest HADDOCK-generated cluster bore marked similarities to the GspC-GspD complex [31]. However, upon further analysis of the HADDOCK-generated structures, a side chain was found to be artificially fixed in position by the rigid body docking procedure, interfering with the protein-protein interface. To allow for greater residue flexibility, the NMR restraints from N0PilQ343–442 and PilP77–164 (PDB 2IVW), together with five intermolecular edge-on backbone hydrogen bond restraints (derived from the favored HADDOCK structure and related GspC-GspD complex), were input into CNS1.2 [41] to generate the final model for the complex (Figure 4C). The binding site is centred around an edge-on interaction between the first two β-strands in each domain. Residue conservation was mapped on to the N0N1PilQ343–545 structure using CONSURF [35], in a similar manner to its implementation for the B2 domain (above), and provided evidence that the proposed binding site for PilP is moderately or well conserved within type IV pilus-dependent secretins (Figure 4D). We conclude that the C-terminal domain of PilP (Figure 1B) recognises the N0 domain from N. meningitidis PilQ in a similar manner to that for GspC and its cognate GspD secretin. This is therefore a further example of the congruence between the type II secretion and type IV pilus biogenesis systems.
We have previously reported on the structure of the intact N. meningitidis PilQ oligomer, using negative stain-based methods [12], [42]. This work established that PilQ forms a dodecamer, in common with the T2SS secretins [13], [14]. In order to generate a structure which would allow docking of the domain structures presented above, we determined a 3D reconstruction of the complete PilQ dodecamer by cryoelectron microscopy. PilQ particles were well dispersed and clearly identifiable (Figure 5A). Single particle selection of 25,303 particles generated a good range of top, side and intermediate views (Figure 5B). The final structure, measuring 155 Å in height and 110 Å at its widest external extent, forms a shell around a large internal chamber (Figure 6A). The chamber is sealed at both ends, and a cut-away view shows evidence for distinct and separate structures within the density map (marked on the right hand side of Figure 6A). From our previous work [12], [42], [43], and comparisons with the structures of other secretins, we ascribe the flattened disc of density at the top of the structure to the membrane-spanning C-terminal domain, which is highly conserved within the secretin family. Our work above has established that PilQ, in common with the other secretins, adopts a ‘string of beads’ type domain organisation. Combining this evidence, we deduce that the structure lining the walls of the chamber, outlined in yellow in Figure 6A, can be reasonably ascribed to the N0/N1 domains. The N-terminal region, encompassing the β-domains would, therefore, form the part of the oligomer which closes the chamber at the bottom (outlined in orange in Figure 6A).
Alignment of the PilQ density map with the T2SS secretin VcGspD [14] shows some key structural differences between the two. PilQ is more compact and, critically, closed at the base, where VcGspD has a flared, open gateway to the secretin chamber (Figure 6B). We attribute this difference to the presence of the B1 and B2 domains in PilQ, which are absent from VcGspD (Figure 1A). The periplasmic gate structure found in VcGspD, which bisects the chamber and effectively divides it into two, is absent from PilQ (Figure 6B). A superposition of PilQ on to the 10 Å resolution cryoelectron microscopy structure of the T3SS needle complex from Salmonella [15] enabled a comparison with the structure of a secretin in the open form. The InvG secretin component from the needle complex forms a cylindrical structure which is open at both ends, to allow assembly of the needle fiber (outlined in blue in Figure 6C). Such a comparison suggests that both the top and bottom parts of PilQ must open up to allow passage of the type IV pilus fiber, in keeping with our previous observation that TFP can bind into the PilQ chamber when added in vitro [43]. Direct comparisons of domain assignments to respective density maps were complicated by possible differences in detergent mass associated with the transmembrane regions, and the large amount of predicted coil or unstructured polypeptide in secretin sequences, with associated uncertainty about the degree to which these regions may contribute to observed density. Nevertheless, it is clear that significant structural differences exist between different secretin types, and also that such structures must be dynamic to allow passage of secreted pilus fibers and exoprotein substrates.
Structures of the B2 domain (B2PilQ224–329) and N0/N1 double domain (N0N1PilQ343–545) were docked into the cryoelectron density map using MULTIFIT, a program which has been shown to work well for structures with multiple components, even with low resolution maps [44]. In addition to optimal fit to the density and minimization of steric clashes, further constraints were applied to differentiate between multiple potential solutions. First, fitting was confined to the relevant sections of the map for each domain, as shown in Figure 6A. Second, orientations of the N0/N1 structure which placed the N1 domain closer to the membrane-spanning region were favoured. Third, the PilP binding site needed to be exposed on the outer surface, in keeping with our previous demonstration that this is the case [27]. Some orientations were also precluded because they created steric clashes between PilP and adjacent PilQ molecules. Fourth, the distance between the C-terminus of the second β-domain and the N-terminus of the N0/N1 double domain needed to be lower than the maximum span which could be plausibly bridged by the missing residues. This latter criterion ruled out an ‘inverse’ orientation of the second β-domain, in which the direction of the last β-strand is towards the base of the PilQ oligomer (i.e. the N-terminal end). These constraints were applied to the highest scoring solutions obtained from MULTIFIT [44], and succeeded in identifying a unique solution for the locations of both B2PilQ224–329 and N0N1PilQ343–545 which satisfied all the criteria (Figure 7A). A striking feature of the resulting assembly is the location of the C-terminal helix in the N1 domain, which is orientated vertically, lining the sides of the top of the chamber and presumably forming a link to the transmembrane domain at the C-terminus. Although the B2 domain fitted extremely well into the relevant part of the map, there was insufficient volume remaining to accommodate a further 12 copies of the B1 domain, if it is assumed that it adopts a similar folded, globular structure. As discussed above, however, our structural work on several such domains from different bacteria did not identify any that were completely folded. We therefore propose that the first β-domain adopts a partially folded structure in the PilQ oligomer, sufficient to contribute some density to the map, but have omitted it from our model as it remains poorly defined at present.
Using the model for the PilP77–164:N0N1PilQ343–545 complex determined above, the PilP C-terminal domain can be placed onto the PilQ assembly (Figure 7B). PilP projects outward from the assembled PilQ complex, in an orientation which is different from the T2SS GspD-GspC complex: in that case, GspC was placed closer towards the interior of the secretin chamber [31]. It is also readily apparent that the PilP C-terminal domain lies close to the B2 domain, essentially sandwiched as a ‘wedge’ between the N0 and B2 domains (Figure 7C). An assembled TFP fiber measures 60 Å in diameter [45]; passage of the pilus fiber would therefore require movement of the PilQ C-domain (Figure 6A), as well as the B1/B2 domains and possibly also the linker between the N0 and N1 domains (Figure 7C). One obvious function of PilP, therefore, is to stabilize the PilQ oligomer during secretion, preventing disassociation and consequent disruption of the channel.
Recent structural work has started to shed some light on secretins and the way in which they mediate the transition of exoproteins across the OM. A question of particular importance is how secretins are able to function in several different secretion systems. Our work here has highlighted a critical adaptation of TFP-dependent secretins which is not found in members of the family elsewhere: the presence of separate β-domains which are involved in closing the chamber at its periplasmic end. The β-domains appear to be uniquely adapted for this purpose and must, by inference from our previous observations on the filling of the PilQ chamber with TFP [43], be involved in gating the entry of pilin or an assembled pilus fiber. A prevailing theme in structural studies on secretins is the modular organisation of their domains. Here we provide evidence that, even in the central part of the chamber where the gap for passage of the pilus fiber is at its widest, there must be some movement to accommodate the pilus fiber during secretion and retraction (Figure 7C). We do note, however, that the type IV pili in N. gonorrhoeae can undergo a force-induced narrowing to a form with a diameter reduced by 40% [46]. We cannot exclude the possibility, therefore, that the PilQ chamber could house the pilus fiber in an intermediate and narrower state. Flexibility of movement between adjacent domains, which we have demonstrated experimentally for the N0 and N1 domains, is likely to be a critical part of secretin function. There is also evidence that secretins somehow recognize their secreted substrates [14], [47]. These observations suggest a model in which the secretins associated with different secretion systems have diversified by modification of their periplasmic domains, and it seems likely that this is where the specificity for recognition of their secreted substrates resides. Such specificity may be necessary in organisms such as P. aeruginosa, which have the capacity to express more than one secretin and may therefore require mechanisms to distinguish between them.
The B2 domain sequence is well conserved in PilQ sequences from other bacteria (Figure S2), suggesting that our observations can be generalised, at least to type IVa pilus-dependent secretins [2]. It is less clear, however, whether type IVb pilus-dependent secretins adopt the same domain organization as shown in Figure 1A. Sequences of BfpB, from E. coli, and TcpC, from V. cholerae, did not align well with the neisserial B2 domain sequence, leaving this as an open question at present. The type IVb pilus-dependent secretins differ in other respects: they have lipid attachment sites at the N-terminus, for example, and no readily apparent equivalent of the PilP lipoprotein.
Our previously reported structural studies on N. meningitidis PilQ by electron microscopy were carried out using negatively stained specimens [12], [42], whereas the current structure has been determined in the absence of stain in vitreous buffer. To date, the best structure available for the PilQ oligomer was obtained using cryo-negative stain, a procedure which involves addition of a negative stain reagent (ammonium molybdate) to the sample before freezing. The additional contrast obtained using negative staining led to a higher quoted resolution value (12 Å) than that cited here for a low contrast, unstained sample (19 Å) but the fitting of domains into a low resolution structure requires a good representation of the true distribution of protein density across 3D space. The resulting map records the molecular envelope well, but not the internal hydrophobic features of a protein which exclude the stain. The structure reported by Collins et al. [12] was adequate to delineate the general structural features of the PilQ oligomer but could not reliably be used for automated docking using MULTIFIT [44], or similar programs, which make no allowance for the contribution of negative stain. Additionally, positive staining of hydrophilic regions of protein may sometimes occur, resulting in an incorrect envelope and a protein deficit where protein density should actually be observed. Finally, the staining pH and ionic strength are usually under non-physiological conditions, resulting in structural changes in the protein that may be artefactual. We therefore argue that the current structure, although it is at lower resolution than that reported by Collins et al., is nevertheless a much better map into which domains can be fitted. A second difference between the two structures concerns the symmetry applied: C12 symmetry was apparent in the structure studied by Collins et al., but C4 symmetry was applied as a more conservative option, given that the C4 signal was stronger and the apparent partial squaring of particles within the data. Since then, much stronger evidence has emerged for C12 symmetry of secretins [14]. Application of C12 symmetry in the refinement of either C4- or C12-symmetric preliminary models led to convergence of the structure during refinement, validating the imposition of C12 symmetry on the structure presented here.
There are a number of well documented similarities between the proteins involved in TFP biogenesis and the T2SS: these include not just the secretins and cytoplasmic ATPases, but also structural components such as the cytoplasmic protein PilM, which has a similar fold to the T2SS protein EpsL [48]. Here we have shown that PilP binds to the N0 domain of PilQ in a similar manner to the recognition of the GspD secretin by GspC [31]: the analogy therefore extends from similarity in fold between the two pairs of proteins, to a similarity in their mode of recognition. This provides further weight to the view that the two secretion systems are evolutionarily related. There are also important differences between the two systems, however. GspC is a multidomain protein, with a transmembrane helix and a C-terminal PDZ domain, as well as the HR domain which is similar in fold to PilP. PilP is also membrane-associated, but through a lipid anchor which is covalently attached to its N-terminus. Between the lipid attachment site and the beginning of the globular domain fold at the C-terminus, there is a proline-rich sequence comprising some 60–70 residues which is unstructured, at least in the N. meningitidis protein [29]. Sequence alignments and secondary structure predictions suggest that this is also the case in other Gram-negative pathogens (not shown). Work on PilP from P. aeruginosa has established that it also binds to the inner membrane proteins PilN and PilO, probably through the unstructured N-terminal region [30]. This result has also been confirmed recently in N. meningitidis [49], and through pull-down experiments with N. meningitidis PilP in solution (our unpublished data). Why is it the case that expression of PilP is critical to TFP assembly in N. meningitidis [26]? Our structure-based model of the PilP:PilQ complex, combined with these other recent observations, suggests that it could play a key role in maintaining assembly of the PilQ oligomer during pilus fiber secretion. There would be much reduced contact between adjacent PilQ monomers in the oligomer, once the C-domain and B1/B2 domains have opened up (Figure 6A). We note that none of the secretin periplasmic domains studied to date form dodecamers when expressed separately in recombinant form, suggesting that the interactions between adjacent monomers in this part of the oligomer are generally weak. PilP, on the other hand, is linked to the inner membrane through its lipid moiety and interaction with PilO and PilN, through its flexible N-terminus [48], [49]. Our current hypothesis is that PilP is needed to maintain the integrity of the PilQ oligomer during secretion, and that it does this by effectively forming a bridge between the PilQ periplasmic domains and the inner membrane.
The large periplasmic chamber formed by PilQ is reminiscent of similar structures found in other OM protein secretory complexes. The Wza translocon for capsular polysaccharides, for example, forms a more elongated chamber but it is also sealed at the periplasmic end [50]. The type IV secretion system complex spans the entire periplasm and, in this case, is a double walled structure with an opening on the cytoplasmic side [51]. Similar studies on the TFP biogenesis system have been complicated by difficulties in isolation of correctly folded and assembled full length PilQ in recombinant form, and in reconstituting the core secretion platform from purified inner and OM components. Our deconstruction of the PilQ-PilP binding site and ability to reassemble the PilQ-PilP complex therefore represents a first, but crucially important, step on the pathway to reassembling this complex molecular machine.
Protocols for expression and purification of all proteins used in this study are described in Text S1.
3 µl samples were applied undiluted to Quantifoil R 1.3/2 holey carbon-coated EM grids and blotted using Whatman No.1 filter paper (2×1 sec blots) at 90% humidity and then frozen in liquid ethane using a Vitrobot plunge freezing system (FEI, Hillsboro, OR). Cryo-EM was performed using a Tecnai F20 200 kV EM operating in low dose mode at 200 kV. Micrographs were recorded using a Gatan 4 k×4 k CCD at underfocus in the range 1–5 µm and with a calibrated magnification corresponding to 4.53 Å/pixel at the specimen level. Images were recorded under low-dose mode with an overall electron dose of 20- 25 electrons/Å2. Particles were selected into 64×64 pixel boxes (equivalent to 290×290 Å) from the digital micrographs using the EMAN software package [67] and masked with a circular mask of radius 131 Å. After correction of the microscope contrast transfer function (CTF), and removal of outlier particles (based on size), a final dataset of 25,303 particles were used to calculate the low resolution 3D structure of PilQ. An initial model was generated by selection of small (<0.5%) subsets of particles with the strongest n-fold symmetry and strongest bilateral symmetry, and then calculating a noisy 3D structure assuming an orthogonal relationship between the two sets of particles. (EMAN command startcsymm). Based on prior work [12] we generated preliminary models for both C4 and C12 symmetry. Iterative refinement of the initial structures was subsequently carried out using the entire dataset, and using both C4 and C12 symmetry for refinement of each model. Comparison of projections of the 3D structures with the corresponding particle class averages, showed a good agreement with the C12 symmetric structure (Figure 5). Moreover applying C12 symmetry in the refinement of either C4- or C12-symmetric preliminary models led to convergence. Estimation of the resolution of the final structure using the same method applied by Collins et al. [12], measuring the value at which a comparison of the Fourier shell correlation (FSC) of one half of the dataset with the other reaches 0.5, gave a value of 1/19 Å−1. Application of the more recently introduced, and more conservative, rmeasure software [68], gave a value of 1/33 Å−1 resolution. Maps derived by electron microscopy were displayed with the CHIMERA software package [69]. The PilQ density map was deposited in the EMDataBank with accession code EMD-2105 and coordinates for the modelled PilQ:PilP complex are available as PDB deposition 4AV2.
|
10.1371/journal.ppat.1000654 | Autogenous Translational Regulation of the Borna Disease Virus Negative Control Factor X from Polycistronic mRNA Using Host RNA Helicases | Borna disease virus (BDV) is a nonsegmented, negative-strand RNA virus that employs several unique strategies for gene expression. The shortest transcript of BDV, X/P mRNA, encodes at least three open reading frames (ORFs): upstream ORF (uORF), X, and P in the 5′ to 3′ direction. The X is a negative regulator of viral polymerase activity, while the P phosphoprotein is a necessary cofactor of the polymerase complex, suggesting that the translation of X is controlled rigorously, depending on viral replication. However, the translation mechanism used by the X/P polycistronic mRNA has not been determined in detail. Here we demonstrate that the X/P mRNA autogenously regulates the translation of X via interaction with host factors. Transient transfection of cDNA clones corresponding to the X/P mRNA revealed that the X ORF is translated predominantly by uORF-termination-coupled reinitiation, the efficiency of which is upregulated by expression of P. We found that P may enhance ribosomal reinitiation at the X ORF by inhibition of the interaction of the DEAD-box RNA helicase DDX21 with the 5′ untranslated region of X/P mRNA, via interference with its phosphorylation. Our results not only demonstrate a unique translational control of viral regulatory protein, but also elucidate a previously unknown mechanism of regulation of polycistronic mRNA translation using RNA helicases.
| All viruses rely on host cell factors to complete their life cycles. Therefore, the replication strategies of viruses may provide not only the understanding of virus pathogenesis but also useful models to disentangle the complex machinery of host cells. Translation regulation of viral mRNA is a good example of this. Borna disease virus (BDV) is a highly neurotropic RNA virus which is characterized by persistent infection. BDV expresses mRNAs as polycistronic coding transcripts. Among them, the 0.8 kb X/P mRNA encodes at least three open reading frames (ORFs), upstream ORF, X, and P. Although BDV X and P have opposing effects in terms of viral polymerase activity, the translational regulation of X/P polycistronic mRNA has not been elucidated. In this study, we show an ingenious strategy of translational control of viral regulatory protein using host factors. We demonstrate that host RNA helicases, mainly DDX21, can affect ribosomal reinitiation of X via interaction with the 5′ untranslated region (UTR) of X/P mRNA and that the downstream P protein autogenously controls the translation of X by interfering with the binding of DDX21 to the 5′ UTR. Our findings uncover not only a unique translational control of viral regulatory protein but also a previously unknown mechanism of translational regulation of polycistronic mRNA using RNA helicases.
| The control of translation initiation on mRNA is one of the most fundamental processes in the regulation of gene expression. Most eukaryotic mRNAs initiate translation via the so-called “scanning mechanism”, in which the 40S ribosomal subunit binds to the cap structure at the 5′-terminus of mRNA and slides to the proximal AUG codon [1]. In this mechanism, translation initiation from the downstream AUGs is generally inefficient. Thus, the eukaryotic cellular genes are transcribed individually, generating monocistronic mRNAs. On the other hand, many animal viruses produce polycistronic mRNAs and express efficiently functionally different proteins from a single mRNA molecule [2]–[5], suggesting that eukaryotic ribosomes have the potential to initiate the translation of downstream ORFs, under the control of sequence- and/or structure-dependent features of the mRNAs.
Polycistronic coding by mRNAs is a means of coordinating the expression of more than two proteins, which are arranged in tandem or overlapping in a single mRNA molecule [6],[7]. Analysis of polycistronic mRNAs therefore provides a better understanding of the regulatory mechanisms of ribosomal scanning during mRNA translation. In the leaky scanning mechanism, ribosomes bypass the first start codon when the context is poor and thus reach a start codon further downstream. Some viruses, such as Sendai virus and papillomaviruses, use such mechanisms to enable a multifunctional mRNA to express several proteins with different functions in viral replication [8]–[10]. Another strategy for translation of downstream cistrons from an mRNA is termination/reinitiation, is the major method of translation of prokaryotic and some viral mRNAs [11]–[13]. In this case, ribosomes resume the scanning of the mRNA and reinitiate translation efficiently at a downstream AUG codon, following the termination of an upstream cistron. Although eukaryotic ribosomes are in general unable to reinitiate downstream cistrons on an mRNA, it is also true that about 10 to 30% of eukaryotic mRNAs contain upstream AUG codons (uAUG), which have the capacity to initiate translation of a short upstream ORF (uORF), usually consisting of fewer than 30 codons [14]–[16]. The uORF-mediated reinitiation of downstream ORFs also has been demonstrated in eukaryotic mRNAs [17]–[20], suggesting that ribosomal termination/reinitiation may be a key mechanism for the regulation of complex gene expression in eukaryotic cells. However, we know little about the molecular mechanisms underlying the regulation of ribosomal initiation in the translation of polycistronic mRNA, especially how eukaryotic viruses use translational regulation in the expression of viral proteins.
Borna disease virus (BDV) is a non-segmented, negative-sense RNA virus that belongs to the Mononegavirales and which is characterized by highly neurotropic and persistent infection. BDV replicates and is transcribed in the cell nucleus and employs several unique strategies for gene expression [21],[22]. One of the most striking characteristics of this virus is that all of the BDV transcripts have polycistronic coding capacity. The shortest, 0.8 kb transcript of BDV, X/P mRNA, encodes at least three ORFs: uORF, X, and P in the 5′ to 3′ direction (Figure 1A). The X and P ORFs produce major viral proteins, which overlap by 215 nucleotides (nt) [23],[24]. In contrast, the uORF, whose stop codon overlaps the X translation start codon (X-AUG) by one nt, UGAUG, (Figure 2A), encodes an 8 amino acid peptide, the expression of which has not yet been shown in infected cells. BDV X is a negative regulator of viral polymerase activity, while the P phosphoprotein is a necessary cofactor of the polymerase complex [25],[26]. Thus, the expression ratio between X and P is critical for viral polymerase activity [25]–[28]. Previous studies revealed that, despite an optimal sequence context for initiation of X compared to P, translation of X seems to be suppressed at an early stage of viral infection and gradually increases along with the establishment of persistent infection (Figure S1) [29],[30]. This finding allows us to hypothesize that translational regulation of such a short polycistronic mRNA may have evolved to control rigorously the ratio between X and P in the infected cell nucleus and is essential for the maintenance of the persistent infection. Recent studies have suggested that the 5′ untranslated region (5′ UTR) of X/P mRNA plays a critical role in the translational regulation of X from the polycistronic mRNA [28],[31]. However, the translation mechanism used by the X/P polycistronic mRNA has not been determined in detail.
In this study, we demonstrate the autogenous translational regulation of the X/P polycistronic mRNA mediated by host RNA helicases. We show that DDX21, also known as RNA helicase II/Gu, is a regulator of ribosomal reinitiation of X via interaction with the 5′ UTR of X/P mRNA (X/P UTR) and that expression of the downstream P protein may regulate the translation of X by interfering with the binding of DDX21 to the 5′ UTR. Our results provide not only a unique insight into translational control of a viral polycistronic mRNA but also a novel role for RNA helicase in the regulation of ribosomal reinitiation during eukaryotic mRNA translation.
To investigate translational regulation of the X/P polycistronic mRNA, we first used a plasmid, pX/Pwt [30], which encodes a cDNA clone corresponding to the X/P mRNA, and assessed whether this plasmid is able to reproduce the translational regulation of X/P mRNA independently of BDV infection. Upon transfection into COS-7 and OL cells, the plasmid produced efficiently both X and P, and expression of X appeared to increase following the course of time after transfection (Figure 1B), similar to the expression dynamics of X in BDV-infected cells (Figure S1) [30],[32]. Previous studies revealed that P translocates to the cytoplasm from the nucleus via interaction with X [28],[30],[33]. As shown in Figure 1C, although the cells transfected with pX/Pwt exhibited the nuclear distribution of P at an early time after transfection, P was shown to move to the cytoplasm of the cells expressing X at 72 h post-transfection (arrows). These observations suggested that the X/P mRNA by itself regulates the translation of X independently of BDV infection and, therefore, that pX/Pwt provides a useful tool to investigate the translational regulation of the polycistronic mRNA.
To understand the role of the uORF in the translation of the X and P ORFs, we generated a series of mutant plasmids (Figure 2A) and examined the expression of X and P by Western blotting at 12 h post-transfection, at which point the expression of X had not yet been upregulated. Mutants with the uAUG replaced by TAG or TTG exhibited markedly increased expression of X compared to the wt plasmid (Figure 2B and 2C, lanes 1 and 2). In addition, the involvement of uORF in the translation of X was demonstrated by using a series of deletion mutants of the 5′ UTR (Figure S2). In contrast, the translation of P was reduced to approximately 50% of the wt plasmid (lanes 1 and 2). In addition, changing the initiation codon of X (X-AUG) to TTG in the above uAUG mutants recovered the expression level of P to the equivalent of the wt plasmid (lanes 3 and 4). Furthermore, single mutants, which lack only the X ORF, with substitution of the X-AUG by AGC or ACG produced P at levels comparable to the wt plasmid with complete abolition of the expression of X (lanes 5 and 6). These observations suggested that the uAUG is recognized efficiently by scanning ribosomes and that the presence of the uORF seems to downregulate the basal expression level of X, while enhancing the translation efficiency of the P ORF to a level equivalent to an mRNA lacking both the uORF and X ORF.
To determine how translation of the X and P ORFs is initiated from the X/P mRNA, we next introduced mutations into the termination codon of the uORF. A mutant in which the uORF stop codon had been changed to TTA expressed both X and P at equivalent levels to the wt plasmid (lane 7). Furthermore, a 3-nt downstream extension of the uORF termination codon also appeared not to influence the translation of either protein (lane 8). Meanwhile, downstream extension of the uORF to 29 nt reduced the expression of X, but not P, to 70% of the wt level (lane 9). Interestingly, a 164 nt extension of the uORF, so that a stop codon is introduced within the P ORF, showed significant decreases (∼30%) in the expression of both X and P (lane 10). The introduction of premature stop codons within the uORF also reduced the expression level of X (Figure S3). These results indicated that the presence of the uORF termination codon in close proximity to the X-AUG seems to be important for efficient translation of the X ORF. Furthermore, termination of the uORF before initiation of the P ORF is required for the expression of P. We also demonstrated that P is unlikely to be expressed by ribosomal shunting or internal ribosome entry site-mediated mechanism (Figure S4). Taken together, both the X and P ORFs were shown to be translated predominantly by ribosomal reinitiation, dependent on uORF termination, and the remainder of the translation (approximately 30%) might be initiated by leaky scanning of upstream start codons.
The results shown above suggest that uORF-termination/reinitiation may play a key role in the regulation of translation of X. Therefore, we sought next to investigate the factors that influence ribosomal reinitiation at the X ORF. At first, we examined the role of the peptide predicted to be produced by the uORF. However, we could not detect any effects of this predicted peptide on the translation of the downstream ORFs (Figure S5).
We next examined the effect of the protein encoded downstream, P, because this may accumulate in the nucleus in association with the expression of the X/P mRNA [34]. We generated mutant plasmids, pX/PΔP, in which the initiation codon of P, P-AUG, was substituted by TTG, and puORF-X/PΔP, in which the uORF was fused in-frame to the X ORF in pX/PΔP, and assessed the production of X at 48 h post-transfection, at which point P should have accumulated sufficiently in wt plasmid-transfected cells. As shown in Figure 3A, translation of X was significantly higher in cells transfected with pX/Pwt than with pX/PΔP. In contrast, initiation of translation of the uORF-X fusion protein appeared not to be affected by deletion of the P-AUG. These results indicated that expression of P upregulates the translation of the X ORF without affecting the ribosomal initiation from the uAUG. To verify this, a plasmid expressing P, pcP [33], was cotransfected with pX/PΔP or puORF-X/PΔP. Interestingly, expression of X, but not the uORF-X fusion protein, was enhanced markedly in a dose-dependent manner by co-expression of P, while the nucleoprotein (N) of BDV failed to enhance the expression of X in transfected cells (Figure 3B). A ΔP mutant based on a 164-bp uORF-extension plasmid (Figure 2A, lane 10), puORF164ΔP, also upregulated the expression of X in the presence of P (Figure 3C), although the basal expression level of X in the puORF164ΔP is significantly lower than that in pX/PΔP both with or without P. These data suggested that P influences the ribosomal initiation at the X-AUG predominantly by the uORF-termination-coupled reinitiation mechanism, while a leaky scanning mechanism may be involved to some extent in the upregulation of the translation of X by P.
We set out to address next the question of how the expression of P can enhance the reinitiation of translation of X. First, we investigated the possibility that P may interact directly with the X/P mRNA and then influence ribosomal reinitiation at the X ORF. We could not demonstrate, however, a direct interaction between P and the X/P mRNA by immunoprecipitation (IP)-RT-PCR analysis (data not shown). Second, it might be possible that P affects the functions of eukaryotic initiation factors (eIFs), such as eIF2α and eIF3. However, an interaction between eIFs and BDV P was not demonstrable in cells transfected with BDV P (Figure S6A). Furthermore, expression of eIF2α, 2Bε and 3A, as well as the serine phosphorylation level of eIF2α, appeared not to be changed in cells expressing P (Figure S6B), indicating that the expression of P is unlikely to affect the quantitative and qualitative properties of eIFs.
We therefore considered the possibility that P enhances the translation of X indirectly by interaction with cellular factor(s), which may affect ribosomal reinitiation at the X ORF. To investigate this, we conducted first an in vitro translation assay using in vitro transcribed X/P mRNAs encoding firefly luciferase fused with the X ORF and cellular extracts of OL cells. As shown in Figure 4A, luciferase activity was markedly reduced in the presence of the nuclear extract, but not the cytoplasmic extract, demonstrating that the nucleus may contain factor(s) that suppress translation initiation from the X-AUG. Interestingly, a mutant X/P RNA, which lacks a 38-mer from the 5′ end of the UTR, retained luciferase activity even in the presence of the nuclear extract (Figure 4A), suggesting that nuclear factors may influence the translation of X via interaction with the 5′ UTR. To verify this, we used 48-mer decoy RNAs, which represent the X/P UTR and, as a control, the 5′ UTR of another BDV polycistronic mRNA (M/G UTR). As shown in Figure 4B, incubation with the X/P UTR decoy RNA, but not the M/G UTR control, interfered with the inhibitory effect of the nuclear extract in a dose-dependent manner. We also performed an RNA-electrophoretic mobility shift assay (RNA-EMSA) using 32P-labeled riboprobes corresponding to the X/P UTR and M/G UTR to determine whether the nuclear extract interacts specifically with the X/P UTR. As shown in Figure 4C, we found that the X/P UTR riboprobe forms complexes with the nuclear extract (arrows) and that an excess of cold riboprobe efficiently interferes with complex formation. On the other hand, the M/G UTR probe failed not only to generate clear complexes with the extract (Figure 4C) but also to interfere with the complex formation of X/P UTR probe (Figure 4D). All these observations suggested the presence of nuclear factors that inhibit the translation of the X ORF via interaction with the 5′ UTR of X/P mRNA.
To investigate in more detail the involvement of nuclear factors in the translational regulation of X, we tried to identify the nuclear factors using a stepwise purification assay and RNA-affinity columns coupled with short (20 mer) and full-length (48 mer) X/P UTR RNAs (see Materials and Methods). The nonspecific RNA binding was visualized using an RNA-affinity column coupled with M/G UTR control RNA. As shown in Figure 5A, seven specific bands were detected by SDS-PAGE. The bands were excised and digested with trypsin and analyzed further by LC-MS/MS. We found that these bands represent DDX21, DDX50, DNA topoisomerase 1 (TOP1), hnRNPQ1/2 and nucleolin (Figure 5A). The accuracy of this analysis was confirmed by Western blotting with antibodies specific for each protein (Figure 5B).
Among the X/P UTR-binding proteins (UBPs), direct interaction has been demonstrated only between nucleolin and TOP1 [35]. Coimmunoprecipitation experiments using Flag-tagged UBPs revealed interactions among DDX21, nucleolin, and TOP1 (Figure 6A and 6B). In addition, interactions of DDX50 and hnRNPQ1 with nucleolin and DDX21, respectively, were demonstrated when hemagglutinin (HA)-tagged proteins were expressed as targets (Figure 6C and 6D). Furthermore, we demonstrated that nucleolin interacts with DDX21 through its C-terminal region using a pull-down analysis with GST- and His-fused recombinant proteins (Figure S7). These observations indicated that the UBPs interact with each other and may have been isolated from the affinity columns as complexes.
To determine which UBPs contribute dominantly in binding to the X/P UTR, we performed IP-RT-PCR analysis using BDV-infected OL cells and X/P mRNA-specific primers. As shown in Figure 7A, X/P mRNA was amplified clearly from the cells transfected with Flag-tagged DDX21 and nucleolin. Furthermore, RNA EMSA using GST-fused DDX21 and nucleolin revealed that DDX21 interacts directly only with the X/P UTR (Figure 7B), while nucleolin binds to both the X/P UTR and the control M/G UTR (arrowhead). We also found by competitive EMSA that DDX21 binds more strongly to the X/P UTR than nucleolin (data not shown). Nucleolin binds a wide variety of DNA/RNA molecules and is known usually to work in concert with other proteins, which may provide the functional specificity [36],[37]. Along with these properties of nucleolin, our results strongly suggested that DDX21 is a core protein that interacts with the X/P UTR.
DDX21 is known to have RNA helicase activity, which may link to RNA-folding and/or -unwinding through its binding directly to target RNA elements [38]. These observations led us to hypothesize that the interaction of DDX21, along with other UBPs, causes a structural change of the X/P UTR, impacting on ribosomal reinitiation at the X ORF. To determine whether DDX21 alters the structure of the 5′ UTR, therefore, we performed an in vitro RNA folding assay using a 32P-labeled X/P UTR probe. At first, we monitored the mobility of the X/P UTR riboprobe in a 12% native PAGE with or without boiling. As shown in Figure 8A and 8B, the probe produced low mobility bands after boiling and quick-cooling (lane 2, arrows), in addition to major bands (arrowheads), which are also seen in the gel without boiling (lane 1). The intensities of these low mobility bands were relatively stable during the equilibration period after the quick-cooling step (data not shown), indicating that the low mobility bands represent the extended form of X/P UTR RNA.
To determine the effect of DDX21 on X/P UTR structure, we added GST-fused DDX21 to the RNA probe during the equilibration period. When the X/P UTR probe was reacted with the active DDX21, high mobility bands were observed in the gels (Figure 8A and 8B, asterisks). On the other hand, the heat-inactivated DDX21 failed to produce such bands (Figure 8A). The major bands were restored even in the presence of proteins, when the re-boiling process was conducted after the equilibration step (Figure 8B, double-asterisks). These results suggested that DDX21 caused the folding of the UTR probe and produce the high mobility bands in the gels. Therefore, DDX21 is likely to cause structural alteration of the 5′ UTR of the X/P mRNA.
To examine the effect of DDX21 on the translation of the X ORF, we performed a coupled assay of in vitro RNA binding and in vitro translation using in vitro transcribed X/P mRNA and recombinant DDX21. As shown in Figure 9A, incubation with DDX21 reduced the translation not only of X, but also of P, from the X/P mRNA. This result suggested that DDX21 inhibits ribosomal scanning through its binding to the X/P mRNA, resulting in suppression of translation of both X and P. However, the difficulty of the reaction conditions, which must be suitable for in vitro RNA binding and in vitro translation reactions in the same tube, as well as the possibility that the effect of DDX21 on ribosomal scanning may require its interaction with other UBPs, suggested that this in vitro assay is insufficient to determine completely the role of DDX21 in translation. Furthermore, although we generated short-interfering RNAs for UBPs, including DDX21, expression of the siRNAs appeared to induce nonspecific inhibition of the translation of other mRNAs (data not shown). Therefore, we sought to investigate further the effect of DDX21 on the translation of the X ORF by focusing on its interaction with P. P is phosphorylated and acts as a protein kinase substrate, inhibiting the phosphorylation of host proteins to modify their functions [39],[40]. A recent study demonstrated the phosphorylation of DDX21 [41]. Furthermore, the phosphorylation of RNA helicases, such as nucleolin, is known to be critical for RNA-binding activity [42]. Therefore, it is tempting to speculate that interference with phosphorylation by P affects the ability of DDX21 to bind to the X/P UTR. To address this, we examined whether the phosphorylation of DDX21 is affected by the expression of P. OL cells were transfected with wt or mutant P, PS26/28A, in which two major phosphorylation sites (Ser26, Ser28) were substituted by alanine [39],[43], and the phosphorylation of DDX21, as well as nucleolin, was monitored. Although the expression levels of the UBPs were unchanged by the expression of P (Figure 9B), the phosphorylation levels of both DDX21 and nucleolin decreased clearly in the cells transfected with wt P, but not with PS26/28A (Figure 9C).
To investigate whether the hypophosphorylation of DDX21 in the presence of P modulates its RNA binding activity, we extracted Flag-tagged DDX21 from the cells transfected with wt P or PS26/28A and then estimated its binding ability to the 32P-labelled X/P UTR probe using an in vitro RNA binding assay. As shown in Figure 9D, Flag-tagged DDX21, as well as nucleolin, from wt P-transfected cells exhibited significant reduction of binding to X/P UTR. The binding activities of the tagged proteins from the cells transfected with PS26/28A were significantly higher than those with wt P, suggesting that interference with phosphorylation by P decreases the RNA binding activity of DDX21. Therefore, we finally examined whether phosphorylation of P directly affects translation of the X ORF. Consistent with Figure 3B, the expression of X was significantly upregulated in the pX/PΔP-transfected cells in the presence of wt P in a dose-dependent fashion, whereas the PS26/28A mutant was not able to upregulate the translation of X (Figure 9E). Altogether, these results suggested that BDV P may inhibit the binding of DDX21 to the 5′ UTR by interfering with its phosphorylation, resulting in the upregulation of the ribosomal reinitiation from the X-AUG.
In this study, we demonstrated translational regulation of polycistronic mRNA in a unique animal RNA virus. The BDV X/P polycistronic mRNA encodes three overlapping ORFs within a short, 0.8 kb sequence. We showed that the X and P ORFs are translated predominantly by a reinitiation strategy, following the termination of translation of the uORF, although a leaky scanning mechanism is implicated to some extent in the translational processes. In this study, we employed an RNA polymerase II-controlled vector for expression of the X/P mRNA in transfected cells. We have carefully investigated the expression, as well as the structure, of the transcripts from pX/P plasmid DNAs in each experiment and then verified that our system could recreate the translational regulation of X/P mRNA in BDV-infected cells (data not shown). Currently, two types of reinitiation mechanism have been identified in eukaryotic and viral mRNAs [2],[3],[6],[11],[17],[18]. The first type of mRNAs contain short uORFs (<30 codons) upstream of the main ORFs. In this mechanism, the efficiency of reinitiation is controlled by the length of the uORF and by the intercistronic region, an appropriate distance being necessary for the recharging of reinitiation factors, including eIF2 and Met-tRNAiMet, to the ribosomes. Cellular mRNAs such as C/EBP and AdoMetDC, are representative examples of this type of regulation [17],[44]. In the X/P mRNA, initiation of translation of the P ORF may be mediated by this type of reinitiation mechanism. The scanning ribosomes, which travel continuously on the mRNA after termination of translation of the uORF, must recharge the initiation factors on the intercistronic region between the uORF and P ORF and efficiently initiate translation from the P-AUG. Note that the expression level of P is quite invariant, with or without translation of X, if the uORF is present (Figure 2), indicating that the number of ribosomes, which move continuously along the mRNA after uORF termination, is relatively constant on the X/P mRNA. This may be the mechanism underlying the stable and persistent expression of P in infected cells.
The second type of reinitiation mechanism involves mRNAs containing long 5′ ORFs, which usually encode functional proteins. These mRNAs display only short intercistronic distances between the upstream and downstream ORFs, or even may overlap. It has been shown that efficient reinitiation in this mechanism is determined by the stability/mobility of ribosomal complexes to allow reinitiation at the downstream initiation codon [17],[18]. Among viral mRNAs, segment 7 of influenza B virus [45], the ORF-2 of the M2 gene of respiratory syncytial virus (RSV) [46], and the 3′ terminal ORF (VP2) of feline calicivirus (FCV) [47],[48] represent are examples of this type of reinitiation process. Our experiments revealed that reinitiation of the X ORF may be regulated by this type of mechanism, although the uORF encodes only a short and, probably, non-functional peptide. Interestingly, the uORF and X ORF feature an overlapping stop-start codon, UGAUG, as found in other viral polycistronic mRNAs [47],[49],[50]. This feature indicates that the overlapping stop-start codon of the X/P mRNA may play a key role in the regulation of translation of the X ORF. Previous studies revealed that genes divided by such an overlapping stop-start codon are expressed predominantly by termination-coupled translation, in which translation of the downstream ORF is initiated by ribosomes which have read the uORF and stalled at the overlapping stop-start codon [48],[51]. The downstream extension of the termination signal of the uORF in the X/P mRNA significantly reduced the expression of X, suggesting that ribosomal reinitiation from the X-AUG is also carried out by the coupled translation mechanism associated with uORF termination.
The regulation of ribosomal movement/stability around the overlapping stop-start codon must be crucial for controlling the translation of the downstream ORF. The scanning ribosomes, which have not dissociated from the mRNA after stalling at the uORF termination codon, may be reutilized efficiently for the reinitiation of translation of the X ORF. In favor of this hypothesis, we found that host nuclear factors influence ribosomal initiation of the X ORF through interaction with the 5′ UTR and identified RNA helicase complexes, mainly involving DDX21. DDX21 is a DEAD-box RNA helicase that localizes to the nucleoli and is involved in ribosomal RNA synthesis or processing [38],[52],[53]. Although detailed functions of DDX21 have not been elucidated yet, this helicase appears to fold or unwind RNA or ribonucleoprotein structures through regulation of RNA-RNA or RNA-protein interaction [38],[52],[53]. We found that DDX21 may be a scaffold protein that interacts with the X/P UTR, among the UBPs, and causes structural alteration of the 5′ UTR. Numerous reports have demonstrated that RNA secondary structure contributes to translational control by affecting the constancy of ribosomal scanning on mRNAs or the recognition of initiation signals [6],[54]. The ribosomes may stack or pass through the initiation codons if secondary structures are formed around the initiation site, leading to enhancement or reduction of the translation efficiency of the ORFs. Therefore, it is conceivable that structural modification of the X/P UTR by DDX21 and the UBPs decreases the ribosomal reinitiation at the X ORF or increases ribosomal dissociation from the mRNA after termination of translation of the uORF, both resulting in the suppression of the translation of the X ORF (Figure 10, left arrow). We found that the structural alterations induced by the base-pair changes in a short stem-loop structure within the X/P UTR influence the translation of the X ORF (Figure S8), supporting this conclusion. On the other hand, in this model the structural change of the X/P UTR should occur in the cytoplasm. Considering DDX21 is mostly a nuclear protein [53], it is possible that the transient interaction of DDX21 with X/P mRNA in the nucleus is enough to maintain the structure of X/P UTR in the cytoplasm by introducing the UBPs (Figure 10, left arrow). Alternatively, DDX21 may be transported to the cytoplasm along with the mRNA in this case.
We revealed that phosphorylation of DDX21, as well as nucleolin, is inhibited by expression of P. Previous studies demonstrated that hyperphosphorylation of nucleolin increases its RNA binding affinity, whereas dephosphorylation reduces the affinity [42]. In this study, the RNA-binding activity of DDX21 was shown to be reduced significantly in cells expressing P. These data suggested that accumulation of P in infected cells blocks interaction of DDX21 with the X/P UTR, resulting in upregulation of translation of the X ORF by promotion of ribosomal reinitiation (Figure 10, right arrow). Note that in Figure 9D the PS26/28A did not fully recover the binding activity of DDX21 to the X/P UTR. This suggests that hypophosphorylation of DDX21 may be not exclusively involved in the promotion of the translation of X, although the in vitro binding assay based on the transfection may be insensitive for the detection of the binding activity of DDX21 to the 5′ UTR.
Previous studies showed that the intranuclear stoichiometry of N and P is important for BDV polymerase activity and that accumulation of P in the nucleus markedly disturbs both viral replication and persistent infection [21],[55],[56]. Interestingly, it has been demonstrated that X binds directly to P and promotes translocation of P to the cytoplasm from the nucleus [30],[33]. Therefore, P-dependent translational regulation of X must be a convenient and effective mechanism for ensuring an optimal level of P in the nucleus. The nuclear accumulation of P above the threshold level upregulates the translation of X, thereby leading to the translocation of P to the cytoplasm. This could keep the amount of P at the optimal level in the nucleus, which is unequivocally necessary for productive replication and/or persistent infection of BDV in the nucleus. A previous study, which demonstrated that the mutations in Ser26 and Ser28 of P aberrantly upregulate the viral polymerase complex activity, and that recombinant BDVs containing the phosphorylation mutations (rBDV-PS26/28A) reduce the expression of X in infected cells [43], may be consistent with our findings, although the possibility that two amino acid changes inevitably introduced in the X ORF of rBDV-PS26/28A affect the expression level of X has remained. In addition, a recent work using a mutant rBDV, which ectopically expresses X under the different transcriptional unit, demonstrated that the expression of X from the mutant virus is not as tightly linked to expression of P as in the wild type BDV, resulting in strong attenuation of the replication of the rBDV in rat brains [57]. This observation may also support our conclusion that the X/P UTR is not only specifically involved in the regulational expression of X but also essentially controls the balanced expression between X and P in infected cells. Furthermore, a recent work by Poenisch et al. [31] showed that recombinant BDVs containing either a premature stop codon in the uORF or mutations ablating the stop codons of the uORF express wild-type like X and P in cultured cells and efficiently replicate in the brains of adult rats. Although this observation may seem to conflict with our findings that the overlapped termination of uORF is critical for the translation reinitiation of X, the recombinant viruses may be able to recover the translation level of X by the expression of other transcription unit, such as a 1.9-kb mRNA, resulting in the efficient replication in infected cells. In fact, Poenisch et al. [31] have demonstrated that the 1.9-kb mRNA not only serves as a template for the synthesis of N but also might be used for the translation of the viral P protein and possibly X, suggesting that the 1.9-kb transcript is a multicistronic mRNA of BDV.
This is the first example, to our knowledge, of autogenous translational regulation of polycistronic mRNA mediated by its own encoding protein and host RNA helicases. The detailed description of the mechanism should provide novel insights into not only an ingenious strategy of virus replication but also the roles of RNA helicases in the translation of eukaryotic mRNAs. Further study remains to be done to discover cellular mRNAs using a similar translation strategy.
The COS-7 cell line was grown in Dulbecco's modified Eagle's medium (DMEM) supplemented with 5% heat-inactivated fetal calf serum (FCS) at 37°C in a humidified atmosphere of 95% air and 5% CO2. The OL cell line, derived from a human oligodendroglioma, was grown in high-glucose (4.5%) DMEM supplemented with 5% FCS. Cells were passaged every 3 days. The BDV strain huP2br [32],[58] was used for analyses in this study.
Construction of the expression plasmids for BDV X/P mRNA, N, P and P phosphorylation mutants has been described elsewhere [29],[30],[32],[33],[39],[43]. The mutant forms of the plasmids were generated using PCR-based site-directed mutagenesis. To generate X/P-Luciferase hybrid mRNAs, a luciferase gene was fused in frame with the X and P ORFs at the 148 nt and 149 nt positions of the coding sequences, respectively, and introduced into the pcDNA3 vector (Invitrogen) at the Kpn Ι-Not Ι sites. The first AUG codon of Luc was replaced by AAG. For expression of DDX21, DDX50, nucleolin, TOP1 and hnRNPQ1, corresponding cDNAs were amplified by RT-PCR from OL cells and inserted into pcXN2, pET32a or pET42a vectors (Novagen). Cells were transfected with equimolar ratios of plasmid DNAs using Lipofectamine™ 2000 (Invitrogen) or FuGENE6 (Roche Applied Science), according to the manufacturer's instructions, and cellular samples were collected at the desired times. The introduction of the correct sequences for the wild type and its mutant were confirmed by DNA sequencing and Western blotting analysis of protein production.
To generate glutathione S-transferase (GST)-tagged DDX21 and nucleolin recombinant proteins used in the Escherichia coli system, we cloned the amplified cDNAs into the pET42a vector (Novagen). The vectors were transformed into BL21 (DE3) (Novagen), and the expression of the GST-tagged proteins was induced by the addition of 0.3 mM IPTG. The cell pellets were resuspended in PBS(-) and then lysed by sonication. After centrifugation, the supernatants were loaded on glutathione sepharose 4B (Amersham Biosciences). Eluted proteins were concentrated using Centricon spin columns (Millipore Corporation) and dialyzed against a 20 mM HEPES (pH 7.5)-100 mM KCl buffer.
The His-tagged DDX21 was generated by the insertion of the PCR-based DDX21 cDNA into PET32a vector (Novagen) and the resultant plasmid was transformed into Rosetta-gami B(De3)pLysS competent cells (Novagen). Purification of the recombinant DDX21 using Ni-NTA agarose (QIAGEN) was performed according to the manufacturer's recommendations.
COS-7 and OL cells cultured in 12-well plates were transfected with plasmids expressing X/P-Luc hybrid mRNA. At 6 h post-transfection, cells were lysed and subjected to luciferase assay system (Promega Corporation), according to the manufacturer's recommendations. The relative levels of luciferase activity were calculated for each fusion plasmid.
For Western blotting, equal amounts of total lysate proteins of COS-7 or OL cells transfected with expression plasmids were subjected to SDS-PAGE and transferred onto polyvinylidene difluoride membrane (Millipore Corporation). Antibodies used in this study were as follows: anti-BDV P mouse monoclonal, anti-BDV X rabbit polyclonal antibodies [30],[33], mouse anti-Flag M2 (Sigma-Aldrich), mouse anti-HA 12CA5 (Roche Applied Science), rabbit DDX21 (Bethyl Laboratories), rabbit Nucleolin (Novus Biologicals), rabbit anti-Topo1 (TopGEN, Inc), mouse hnRNP-Q (ImmunoQuest Ltd), rabbit anti-phosphoserine (ZYMED Laboratories).
For immunoprecipitation (IP) assay, OL cells transfected with Flag-tagged plasmids were lysed with RIPA buffer [20 mM Tris-HCl (pH 7.4), 150 mM NaCl, 2 mM EDTA, 1% Nonidet P-40 (NP-40), 1% Na-deoxycholate with protease inhibitors]. After centrifugation at 15,000 rpm for 30 min, the supernatants were incubated with 40 µl of pre-equilibrated anti-Flag M2 agarose (Sigma-Aldrich) overnight at 4°C with gentle rotation. After incubation, beads were collected by centrifugation at 6,000 rpm for 40 s and washed four times with 1 ml of RIPA. The proteins immunoprecipitated with anti-Flag agarose were eluted with 3×Flag peptide (Sigma-Aldrich) in RIPA buffer and detected by Western blotting as described above. In IP for detection of phosphoserine, NaF and Na3VO4 were added in RIPA buffer, and the serine-phosphorylated proteins were detected by anti-phosphoserine antibody.
To detect the interaction of host factors with BDV X/P mRNA in vivo, BDV-infected OL cells were transfected with Flag-tagged targeted proteins and lysed with RIPA buffer with RNasin (Promega). After IP with anti-Flag M2, the co-immunoprecipitants were boiled in TE buffer and then treated with RNase-free DNase Ι for 20 min. Total RNAs were isolated from the aqueous solution and used as templates for RT-PCR using specific primers of X/P mRNA.
In vitro transcribed X/P-Luc mRNAs were prepared with Maxiscript Kits (Ambion). About 1.0 pmol of X/P-Luc mRNAs were pre-incubated with nuclear extracts of OL cells (total protein 1 to 4 µg) in a total 20 µl of binding mixture [10 mM HEPES (pH 7.6), 67 mM NaCl, 2 mM MgCl2, 1 mM DTT, 1 mM EDTA, 5% glycerol, 10U RNasin] for 30 min at room temperature. For competition, a serial dilutions of decoy RNAs were pre-incubated with the extracts prior to the reaction. Binding mixtures were then subjected to the in vitro translation system using 50 µl of rabbit reticulocyte lysate (Promega), according to the manufacturer's recommendations. After incubation period of 2 h at 30°C, 10 µl of mixture was subjected to luciferase assay as described above.
The 32P labeled-transcripts corresponding to the X/P and M/G UTRs were prepared with a mirVana miRNA Probe construction kit (Ambion), using PCR products or synthetic oligonucleotides as dsDNA templates. Transcription of the X/P and M/G UTRs was confirmed by their mobility in native PAGE. Unlabeled transcripts were prepared with MEGAshortscript™ T7 Kit (Ambion). The cell extracts were obtained from exponentially growing OL cells. The cells were lysed with buffer A [20 mM HEPES (pH 7.6), 10 mM NaCl, 1.5 mM MgCl2, 0.2 mM EDTA, 1 mM DTT, 0.1% NP-40, 20% glycerol and protease inhibitor cocktail] and then incubated on ice for 5 min. After collection of the cells, the lysate was incubated for a further 10 min. After centrifugation at 2,000 rpm for 5 min, the supernatant was collected as the cytoplasmic extract. The pellet was lysed with buffer B [20 mM HEPES (pH 7.6), 500 mM NaCl, 1.5 mM MgCl2, 0.2 mM EDTA, 1 mM DTT, 0.1% NP-40, 20% glycerol and protease inhibitor cocktail], incubated on ice for 30 min and separated by centrifugation at 15,000 rpm for 15 min. The soluble nuclear fraction was diluted in binding buffer [10 mM HEPES (pH 7.6), 100 mM NaCl, 1.5 mM MgCl2, 1 mM EDTA, 1 mM DTT, 0.1% NP-40, 10% glycerol]. About 1.0 pmol of 32P-labeled gel-purified probes was incubated with the nuclear extracts (4 µg) in a total of 30 µl of binding mixture [10 mM HEPES (pH 7.6), 67 mM NaCl, 2 mM MgCl2, 1 mM DTT, 1 mM EDTA, 5% glycerol, 20 µg tRNA, 10 U RNasin] for 20 min at room temperature. For competition, non-labeled probes were incubated with the nuclear extract for 20 min at room temperature prior to incubation with the labeled probes. For the assays using recombinant proteins, the probes were incubated with 5 pmol of GST-tagged DDX21 or nucleolin in a total of 20 µl of binding mixture [20 mM HEPES (pH7.5), 70 mM KCl, 2 mM MgCl2, 2 mM DTT, 0.2 mg/ml BSA, 20 U RNasin] for 10 min at 30°C and for 10 min at room temperature. The reaction mixtures were applied to 4% native polyacrylamide gels (40∶1 acrylamide-bisacrylamide) in TBE buffer. After electrophoresis, the gels were exposed to X-ray film overnight at −80°C.
Nuclear extracts of OL cells were prepared as described above. The nuclear extracts were passed through the RNA-negative coupled column and then loaded onto a consecutive RNA-positive column to remove nonspecific binding proteins. The extracts (total 2.5 mg of protein) were loaded on HiTrap Streptavidin HP column (1.0 ml bed volume; GE Healthcare) equilibrated with binding buffer three times (0.2 ml/min). The flow-through was incubated with 0.02 µmol of a 5′-biotinylated short (20 mer) RNA probe in binding buffer on ice for 30 min, and passed through a HiTrap column three times (0.2 ml/min). The column was washed with 30 ml binding buffer and then the proteins were eluted from the columns by the addition of binding buffer containing 600 mM NaCl. After dialysis with binding buffer, the sample was subjected to an X/P UTR- or M/G UTR-coupled column as a second step of RNA-affinity purification. After washing, the binding proteins were eluted from the column with the same as for the short RNA probe-coupled column.
Samples eluted from the RNA affinity columns were separated on 10% SDS-PAGE and visualized by silver-staining (Wako). The protein bands of interest were excised, digested in-gel with trypsin, and analyzed by nanocapillary reversed-phase LC-MS/MS using a C18 column (φ 75 µm) on a nanoLC system (Ultimate, LC Packing) coupled to a quadrupole time-of-flight mass spectrometer (QTOF Ultima, Waters). Direct injection data-dependent acquisition was performed using one MS channel for every three MS/MS channels and dynamic exclusion for selected ions. Proteins were identified by database searching using Mascot Server (Matrix Science).
For the protein pull-down assay, 200 pmol of recombinant His-DDX21 and approximately 100 pmol of truncated forms of GST-nucleolin were incubated with RIPA buffer for 1 h at 4°C. After the incubation, reaction mixtures were bound to glutathione-Sepharose 4B (Amersham Biosciences) in RIPA buffer overnight at 4°C. After washing with the same buffer three times, bound proteins were analyzed by immunoblotting with anti-DDX21 antibodies.
About 1.0 pmol of 32P-labeled probes were heated at 85°C for 5 min, quickly cooled on ice and equilibrated at 23°C for 20 min prior to the reaction, unless manipulated further. These RNAs were incubated with 5 pmol of GST-tagged DDX21 and GST-tagged truncated nucleolin, Nuc(1234R) in total 15 µl of binding mixture [20mM HEPES (pH 7.5), 70 mM KCl, 3 mM ATP, 0.2 mg/ml BSA, 20 U RNasin] at 23°C for 20 min. After the incubation, the reaction was terminated by the addition of 5×loading buffer [20 mM HEPES (pH 7.5), 70 mM KCl, 50% glycerol, 0.5% SDS, 0.2 mg/ml proteinase K, 0.01% BPB, 0.01% XC], which also inactivated the enzyme. A part of the reaction mixtures was then applied to 12% native polyacrylamide gel (40∶1 acrylamide-bisacrylamide) in TBE buffer. After electrophoresis, the gels were exposed to X-ray film overnight at −80°C.
The OL cells expressing Flag-tagged recombinant proteins were lysed with RIPA buffer including protease inhibitors and 40 µg/ml of RNase A, and IP were performed using anti-Flag M2 as described above. The precipitants were washed twice with washing buffer [20 mM Tris-HCl (pH 7.5), 70 mM NaCl, 70 mM KCl, 0.1% NP-40] and once with binding buffer [20 mM Tris-HCl (pH 7.5), 70 mM KCl, 0.1% NP-40] and subjected to in vitro binding assay. 10 pmol of 32P-labeled X/P UTR probe was added to 20 µl of 50% suspension of the protein-loaded beads. After adjusting the total volume to 250 µl with binding buffer, the reaction mixture was incubated at 4°C for 10 min with gentle agitation. Unbound probe was removed by washing three times with 1 ml of binding buffer. The amount of bound radio-activity was measured by scintillation counting and the specificity was achieved by eliminating background activity obtained from the bead with the mock-transfected cell extract.
|
10.1371/journal.pcbi.1002281 | Sparse Gamma Rhythms Arising through Clustering in Adapting Neuronal Networks | Gamma rhythms (30–100 Hz) are an extensively studied synchronous brain state responsible for a number of sensory, memory, and motor processes. Experimental evidence suggests that fast-spiking interneurons are responsible for carrying the high frequency components of the rhythm, while regular-spiking pyramidal neurons fire sparsely. We propose that a combination of spike frequency adaptation and global inhibition may be responsible for this behavior. Excitatory neurons form several clusters that fire every few cycles of the fast oscillation. This is first shown in a detailed biophysical network model and then analyzed thoroughly in an idealized model. We exploit the fact that the timescale of adaptation is much slower than that of the other variables. Singular perturbation theory is used to derive an approximate periodic solution for a single spiking unit. This is then used to predict the relationship between the number of clusters arising spontaneously in the network as it relates to the adaptation time constant. We compare this to a complementary analysis that employs a weak coupling assumption to predict the first Fourier mode to destabilize from the incoherent state of an associated phase model as the external noise is reduced. Both approaches predict the same scaling of cluster number with respect to the adaptation time constant, which is corroborated in numerical simulations of the full system. Thus, we develop several testable predictions regarding the formation and characteristics of gamma rhythms with sparsely firing excitatory neurons.
| Fast periodic synchronized neural spiking corresponds to a variety of functions in many different areas of the brain. Most theories and experiments suggest inhibitory neurons carry the regular rhythm while being driven by excitatory neurons that spike more sparsely in time. We suggest a simple mechanism for the low firing rate of excitatory cells – spike frequency adaptation. Combining this mechanism with strong global inhibition causes excitatory neurons to group their firing into several clusters and, thus, produce a high frequency global rhythm. We study this phenomenon in both a detailed biophysical and an idealized model that preserves these two basic mechanisms. Using analytical tools from dynamical systems theory, we examine why adaptation causes clustering. In fact, we show the number of clusters relates to a simple function of the adaptation time scale over a broad range of parameters. This allows us to develop several predictions regarding the formation of fast spiking rhythms in the brain.
| Synchronous rhythmic spiking is ubiquitous in networks of the brain [1]. Extensive experimental evidence suggests such activity is useful for coordinating spatially disparate locations in sensory [2], motor [3], attentional [4], and memory tasks [5]. In particular, network spiking in the gamma band (30–100 Hz) allows for efficient and flexible routing of neural activity [6]. Groups of neurons responding to a contiguous visual stimulus can synchronize such fast spiking to within milliseconds [7]. The processing of other senses like audition [8] and olfaction [9] has also been shown to employ synchronized gamma rhythms, suggesting this fast synchronous activity is indispensable in solving perceptual binding problems [10]. Aside from sensation, gamma band activity has been implicated in movement preparation in local field potential recordings of macaque motor cortex [3] and electroencephalogram recordings in humans [11]. Also, there is a boost in power of the gamma band in both sensory [12] and motor [13] cortices during an increase in attention to related stimuli, which may serve as a gain control mechanism for downstream processing [4]. Short term memory is another task shown to consistently use gamma rhythms in experiments where humans must recall visual stimuli [14]. Thus, there are a myriad of studies showing gamma band synchrony appears in signals of networks performing neural processing of a variety of tasks and information. This suggests an understanding of the ways in which such rhythms can be generated is incredibly important to understanding the link between single neuron activity and network level cognitive processing.
Many theoretical studies have used models to generate and study fast, synchronous, spiking rhythms in large neuronal networks [15]–[17]. One common paradigm known to generate fast rhythms is a large network of inhibitory neurons with strong global coupling [16]. Periodic, synchronized rhythms are stable because all cells must wait for global inhibition to fade before they may spike again. This observation lends itself to the theory that gamma rhythms can be generated solely by such mutual inhibition, the idea of interneuron network gamma (ING) oscillations [18]. Of course, this idea can be extended to large networks where excitatory neurons strongly drive inhibitory neurons that in turn feedback upon the excitatory population for a similar net effect [19], [20] (see also Fig. 5 of [21]), known as pyramidal–interneuron network gamma (PING) oscillations [18], [22]. Even when coupling is sparse and random, it is possible for large networks with some inhibitory coupling to spontaneously generate a globally synchronous state [19], [23]. The primary role of inhibitory neurons in gamma rhythms has been corroborated in vivo by [24], using optogenetic techniques. Light-driven activation of fast-spiking interneurons serves to boost gamma rhythms, whereas driving pyramidal neurons only increases the power of lower frequencies. Depolarization of interneurons by activating channelrhodopsin-2 channels has also been shown to increase gamma power in local field potentials [25]. Still, no conclusive evidence exists to distinguish between PING or ING being more likely, and [26] suggests that weak and aperiodic stimulation of interneurons is the best protocol to make this distinction. Nonetheless, it is clear that recent experiments have verified much of the extensive theory developed regarding the mechanism of gamma rhythms.
One particularly notable experimental observation of the PING mechanism for gamma rhythms is that constituent excitatory neurons fire sparsely and irregularly [12], [27], while inhibitory neurons receive enough excitatory input to fire regularly at each cycle. Due to their possessing slow hyperpolarizing currents, pyramidal neurons spike more slowly than interneurons [28], so this partially explains their sparse participation in a fast rhythm set by the interneurons. Modeling studies have accounted for the wide distribution of pyramidal neuron interspike intervals by presuming sparse random coupling in network connections [29] or by including some additive noise to the input drive of the population [30]. From this standpoint, the excitatory neurons are passive participants in the generation of fast rhythms, so their statistics have no relation cell to cell. The requirement, in these cases, is a high level of variability in the structure and drive to the network. However, an alternative explanation of sparse firing might suggest that excitatory neurons assemble into subpopulations, clusters, that fire in a more regular pattern for a transient period of time. This may be accomplished without the need for strong variability hardwired into a network.
One cellular mechanism that has been largely ignored in network models of fast synchronous spiking rhythms is spike frequency adaptation [30], [31]. Slowly activated hyperpolarizing currents known to generate spike frequency adaptation have been shown in many different populations of regular spiking cells within cortical areas where gamma rhythms arise. In particular, pyramidal neurons in visual cortex exhibit slow sodium and calcium activated afterhyperpolarizing current, proposed to play a major role in generating contrast adaptation [32]. Regular spiking cells in rat somatosensory cortex also have adaptive currents. Furthermore, they exhibit a type 1 threshold, where they can fire regularly at very low frequencies [33]. Also, recent experiments in primate dorsolateral prefrontal cortex reveal significant increases in interspike intervals due to spike frequency adaptation [34]. Synchronous spiking in the gamma range has been observed in visual [2], [12], somatosensory [35], [36], and prefrontal [14] cortex, all areas with neurons manifesting adaptation. Also, adaptation may promote a low resonant frequency in regular spiking neurons that participate in gamma rhythms, as revealed by optogenetic experiments [24]. Therefore, adaptation not only slows the spike rate of individual regular spiking neurons, but can play a role in setting the frequency of network level spiking rhythms.
Thus, we propose to study a paradigm for the generation of a network gamma rhythm in which excitatory neurons form clusters. This accounts for the key observation that excitatory cells do not fire on every cycle of the rhythm. The essential ingredients of the network are spike frequency adaptation and global inhibitory coupling. Spike frequency adaptation produces the slow firing of individual cells. The restrictions on the sparsity of coupling and the level of noise in the network are much looser than [30]. After identifying these properties of the network, we can extract several relationships between parameters of our model and attributes of the resulting clustered state of the network. One result of considerable interest is the relationship between the time constant of adaptation and the number of clusters that can arise in the network. Using two different methods of analysis, we can predict the cluster number to scale with adaptation time constant as .
The paper employs both a detailed biophysical model as well as an idealized model that we study for the formation of cluster states. Our results begin with a display of numerical simulations of cluster states in the detailed model. The main point of interest is that excitatory neurons possess a spike frequency adaptation current whose timescale appears to influence the number of clusters that can arise. To begin to understand how this happens, we analyze the periodic solution of a single adapting neuron, in the limit of large adaptation time constant, for an idealized model of adapting neurons. Using singular perturbation theory, we can derive an approximate formula for the period of a single neuron and thus an estimate of the number of clusters in a network of neurons. Then, an exact expression is derived for the periodic solution of an equivalent quadratic integrate and fire model with adaptation as well as its phase-resetting curve. Next, we employ a weak coupling assumption to predict the number of synchronized clusters that will emerge in the network as the amplitude of additive noise is decreased. The number of clusters in the predicted state is directly related to a Fourier decomposition of the phase-resetting curve. Our main result is that both the singular perturbation theory and weak coupling analysis predict the same power law relating cluster number to adaptation time constant. Finally, we compare our predictions made using singular perturbation theory and the weak coupling approach to numerical simulations of the idealized model and the detailed biophysical model.
For our initial numerical simulations, we use a biophysical model developed by Traub for a network of excitatory and inhibitory spiking neurons [37]. Parameters not listed here are given in figure captions. The membrane potentials of each excitatory neuron and each inhibitory neuron satisfy the dynamics:with synaptic currentswhere , , , and are random binary matrices such thatand the synaptic gating variables are givenwhereThe ionic currents of each excitatory and each inhibitory neuron are givenwhere gating variables evolve aswhere . The biophysical functions associated with the gating variables areCalcium concentration associated with the hyperpolarizing current responsible for spike frequency adaptation in excitatory neurons follows the dynamicsBias currents to both excitatory and inhibitory neurons have a mean and fluctuating partwhere fluctuations are given by a white noise process such thatFinally, the fixed parameters associated with the network model areRandom initial conditions are used for the simulations of the model, and we wait until the system has settled into a steady state to make calculations of the statistics. We evolve this model numerically, using the Euler-Maruyama method, with a time step of dt = 0.0001.
The majority of our analysis uses an idealized spiking neuron model to study the mechanism of clustering associated with a network of adapting neurons. The Traub model for a single neuron exhibits a saddle-node on an invariant circle (SNIC) bifurcation. It is possible to exploit this fact to reduce the Traub model to a theta neuron model with adaptation, if the system is close to the bifurcation and the adaptation is small and slow [38]. In [39], an alternative conductance based model with an afterhyperpolarizing (AHP) current was reduced using phase reduction type techniques, where the AHP gating variable was taken to evolve slowly. In particular, Fig. 3(c) of [39] shows that the associated phase-resetting curve has a characteristic skewed shape. We also eliminate the inhibitory cells from the idealization of this section by slaving their synaptic output to the total firing of the excitatory cells. To our knowledge, there is no rigorous network level reduction that would allow us to reduce the excitatory-inhibitory conductance based network to the idealized one we present here. We do not provide a meticulous reduction from the Traub network model to the network analyzed from here on. We do wish to preserve the essential aspects of the biophysical model described in the previous section, spike frequency adaptation and inhibitory feedback.
Therefore, we consider a system of spiking neurons, each with an associated adaptation current, globally coupled by a collective inhibition current(1a)(1b)(1c)for . Equation (1a) describes the evolution of a single spiking neuron with input , in the presence of spike frequency adaptation with strength and global inhibition with strength . Each neuron's input has the same constant component and a unique noisy component with amplitude where is a white noise process such that and for . The adaptation current associated with each neuron is discretely incremented with each spike and decays with time constant , according to equation (1b). Global inhibitory synaptic current is incremented by with each spike and decays with time constant . Notice, in the limit of pulsatile synapses (), the equation (1c) for inhibitory synaptic current becomesWe will make use of this reduction for some calculations relating cluster number to model parameters. The membrane time constant of neurons is usually approximated to be between 1–5 ms, so even though time has been nondimensionalized, its units could be deemed to be between 1–5 ms. In addition, experimental results suggest that the hyperpolarizing currents that generate spike frequency adaptation decay with time constants roughly 40–120 ms [40], [41], indicating that . This observation will be particularly helpful in calculating a number of results.
Note, we consider this model as an idealization of adapting excitatory spiking neurons coupled to a smaller population of inhibitory neurons that then collectively connect to the excitatory population. Our approximation is reasonable, considering inhibitory neurons evolve on a faster timescale than the adapting excitatory neurons, as they did in the more detailed biophysical Traub model. For our numerical simulations of this model, we employ the Euler-Maruyama method, with a time-step of dt = 0.0001.
To display the spikes from our simulations of the Traub model (see Figs. 1 and 2), we employ the following sorting technique. First, to better illustrate the formation of clusters, we sort the simulations displayed in Fig. 1 in order of increasing voltage at the end of the simulation using MATLAB's sort function. Similarly, we sort the neurons in Fig. 2a in decreasing order, according to their spike time closest to ms also using the sort function. We do not resort the neurons between the left and right panel, which displays the mixing effects of cycle skipping.
We use standard techniques for computing the interspike interval (ISI) and correlation coefficient (CC) for the population of spike trains. Calculations of the ISI take spike times of each neuron () and compute their difference (). Interspike intervals of all excitatory neurons are then combined into one vector and a histogram is then computed with MATLAB's hist function for a bin width of . We compute the CC for all possible pairs of excitatory neurons to ensure the best possible convergence. We first digitize two neurons' ( and ) spike trains into bins of and then use MATLAB's xcorr function to compute an unnormalized correlation function. This is then normalized by dividing by the geometric mean of both neuron's total firing and over the time interval. For the calculations displayed in Fig. 2, we use a total run time of .
The extensive singular perturbation theory analysis we carry out on the idealized network suggests that there is a clear cut scaling for the relationship between the number of clusters arising in a network and the adaptation time constant (We also use the following least squares method to fit data relating to attained from numerical simulations of the Traub model). To compare this result with the relations between and derived using a weak coupling assumption, we consider the function determined by (23). This gives the number of clusters associated with a particular and so must be an integer number. Since (6) is a continuous function, we wish to remove the stepwise nature of to make a comparison. Thus, we first generate the vector and matrixwhere and are the minimum and maximum number of clusters attained in the given range of . The function gives the minimal value of such that ; in other wordsNote that and . Now, we solve for the coefficients of the power function fit by solvingas an overdetermined least squares problem for the coefficient vector . We find the points are well fit by the specific case . To generate the inset plot, we simply compute the residualfor . This shows the global minimum is in very close proximity to .
As a means of comparison with our theory, we perform simulations of the idealized model by starting the system (1) at random initial conditionswhere are uniformly distributed random variables on , is given in Text S1, and . As suggested by our weak coupling analysis, we start the system with high amplitude noise (), where clusters are not well defined, and incrementally decrease as the system evolves until noise is relatively weak (). For low noise, each cluster is particularly well defined, especially when there are fewer clusters present.
We now describe the attainment of the data points corresponding to minimal ( for our calculations of the Traub model) to attain clusters for numerical simulations. These are computed by, first, simulating 20 realizations for each value of ( for the Traub model), starting with random initial conditions (2) and high noise, reducing noise and stopping after 20000 time units (20000 ms for the Traub model), and finally recording the number of clusters in the network for each realization. The points we then plot correspond to the first value of whose median cluster number is larger than the median for the previous () value. Increments in between neighboring () values are always no more than one.
Clustering of spiking activity in a network of neurons is the phenomenon in which only neurons belonging to the same cluster spike together, and two or more clusters spike each period of the population oscillation. The emergence of cluster states has been studied in globally coupled networks of phase oscillators with additive noise [42], where clusters can be identified using stability analysis of an associated continuity equation. Phase oscillator networks may also develop clustering in the presence of heterogeneous coupling [43] or time delays [44], [45]. Golomb and Rinzel extended early work in phase oscillators to show cluster states can arise in biologically-inspired networks of Wang–Rinzel spiking neurons [46]. They employed a stability analysis of periodic solutions to their network, using Floquet multipliers to identify which cluster state could arise for a particular set of parameters. Networks of leaky integrate-and-fire neurons can also exhibit clustering if coupled with fast inhibitory synapses [47] or there is sufficient heterogeneity in each neuron's intrinsic frequency [48]. In Hodgkin-Huxley type networks clustering has been witnessed due to a decrease in the amplitude of a delayed rectifier current [16] or by simply including a delay in synaptic coupling [44]. The addition of a voltage dependent potassium current to an excitatory-inhibitory network has also been shown to form two cluster states in detailed simulations [49].
In this section, we show clustering can arise in a detailed biophysical model network of spiking neurons developed by Traub (see Methods). The network consists of excitatory and inhibitory neurons, but only excitatory neurons possess a slow calcium activated hyperpolarizing current, representative of spike frequency adaptation. The connectivity structure is dense but random, where each pair of neurons has a set probability of being connected to one another, according to their type. Here, we present the results of numerical simulations of this model, showing the behavior of cluster states in the network. More specifically, we are interested in the way that spike frequency adaptation helps to generate these states. In later sections, we look at cluster states in an idealized network model in order to analytically study the role of adaptation in the onset of clustering.
We first present spike times of a model network of 200 excitatory and 40 inhibitory Traub neurons in Fig. 1 for two different time constants of the calcium-induced hyperpolarizing current. In particular, we find that, for slower adaptation, there is an increase in the number of clusters, but the overall frequency of the network decreases. This relationship persists over a wide range of model parameters, like network connectivity, synaptic strength, and input to neurons. To aid in the visualization of the clusters, we sort the neurons according to their voltage's value at the end of the simulation (see Methods).
Although the size of the clusters is fairly invariant over time, neurons do not remain in the same clusters indefinitely. In fact, by examining the state of neurons at times significantly before of after the time we sort them according to spike times (see Methods), we find that units of clusters begin to mix with one another, shown in Fig. 2(a). Neurons jump from one cluster to another. The mechanisms by which this can occur are that either a neuron fails to fire with its current cluster and fires with the next cluster or the neuron fires with the previous cluster. This is exemplified by the additional peaks in the interspike interval distribution shown in Fig. 2(c). The correlation coefficient is relatively low on short time scales and decreases significantly over long time scales since neurons skip cycles or spike early due to fluctuations in drive to the network (see Fig. 2(d)). As pictured in Fig. 2(e), on short timescales, excitatory neuron spike times are weakly correlated between clusters, before cycle hopping takes effect. We have found that higher amplitude noise leads to more frequent switching of neurons between clusters. In addition, as the number of clusters increases, each individual cluster appears to be less stable and neurons also hop from one cluster to the next more frequently. We have considered architectures for which the cross correlations between neurons decay more quickly due to sparser connectivity. The main goal of our study, though, is to examine clustering as a complementary mechanism to irregular input and random connectivity for generating sparse firing. This can be contrasted with the degradation of correlations between excitatory neurons on fast timescales in [30], due to strong fluctuations and sparse connectivity in their excitatory-inhibitory network.
Thus, the cluster state that arises in this biophysically based network of spiking neurons appears to be a stable state that exists over a large range of parameters. The essential ingredients are a slow adapting current and inhibitory neurons that only fire when driven by excitatory neurons.
The key feature of the detailed biophysical model that makes excitatory neurons susceptible to grouping into clusters is spike frequency adaptation. Few studies have examined the effects of adaptive mechanisms on the dynamics of synchronous states in spiking networks. In a study of two coupled adapting Hodgkin-Huxley neurons, their excitatory synapses transitioned from being desynchronizing to synchronizing as the strength of their spike frequency adaptation was increased [50]. In a related study, spike frequency adaptation was shown to shift the peak of an idealized neuron's phase-resetting curve, creating a nearly stable synchronous solution [51]. The effects of this on network level dynamics were not probed, and, in general, studies of the effects of adaptation on dynamics of large scale neuronal networks are fairly limited. A large excitatory network with adaptation can exhibit synchronized bursting, followed by long periods of quiescence set by the adaptation time constant [52]. Spike adaptation must build up slowly and be strong enough to keep neurons from spiking at all. More aperiodic rhythms were studied in populations of adapting neurons by [53], who showed the population frequency could be predicted by the preferred frequency of a single adapting cell. Adaptation has also been posed as a mechanism for disrupting synchronous rhythms in [54], where increasing the conductance of slow hyperpolarizing currents transitions a network to an asynchronous state. There remain many open questions as to how the strength and timescale of adaptive processes in neurons contribute to synchronous modes at the network level.
We therefore proceed by studying several characteristics of the cluster state as influenced by spike frequency adaptation. First, we study how the period of a single neuron relates to the strength and time scale of adaptation. Then, we find how these parameters bear upon the number of clusters arising in the network of adapting neurons with global inhibition. Approximate relations are derived analytically and then compared to the results of simulations of (1) as well as the Traub model.
We first present a calculation of the approximate period of a single adaptive neuron, uncoupled from the network. The singular perturbation theory we use relies upon the fact that the periodic solution is composed of three different regions in time: an initial inner boundary layer; an intermediate outer layer; and a terminal inner boundary layer. In this case, the initial and terminal boundary layers correspond to what would be the back and front of an action potential in a biophysical model of a spiking neuron, such as the Traub model. The intermediate layer corresponds to a refractory period imposed by the strong slow afterhyperpolarizing current. An asymptotic approximation to the periodic solution is pictured in Fig. 3, showing the fast evolution of in boundary layers and slow evolution in the outer layer. The slow timescale arises due to the fact that , so we shall use the small parameter in our perturbation theory. Key to our analysis is the fact that the end of the outer layer comes in the vicinity of a saddle-node bifurcation in the fast subsystem, determined by the equation (1a). It then turns out that, as a result, we must rescale time to be in the terminal boundary solution. Such an approach has been studied extensively by Guckenheimer in the Morris-Lecar and Hodgkin-Huxley neurons with adaptation, as well as general systems that support canards of this type [55], [56]. Nonetheless, we proceed by carrying out a similar calculation here and use it to derive an approximate formula for the period of the solution. We find that it matches the numerically computed solution remarkably well. In addition, we can use the expression for the period to explain why the number of clusters arising in the network (1), when compared to the adaptation time constant , will scale as .
To initially approximate the interspike interval for a deterministically–driven adaptive neuron, uncoupled from the network(2)we shall use singular perturbation theory. In particular, we exploit the fact that the adaptation time constant is large in comparison to the membrane time constant of a spiking neuron. Guckenheimer has carried out several other studies examining relaxation oscillations and canards in the vicinity of fold singularities [55], [56]. The usual approach is to decompose the full system into a fast and slow part and then use standard methods of bifurcation analysis to analyze constituent parts [57].
We are particularly interested in computing the approximate form of a periodic solution. The details of this calculation are carried out in Text S1. Our analysis exploits the fact that the fast subsystem, defined by the equation of the system (2), exhibits a saddle-node on an invariant cycle (SNIC) bifurcation. Thus, we have an approximate periodic solution that is split into two time regions, one before the subsystem reaches the SNIC at time and the other after, so(3)andwhere the parameters and are defined in Text S1 while and are Airy functions of the first and second kind. We plot this solution along with numerical simulations in Fig. 4. The location of the saddle-node bifurcation point of the fast subsystem correlates biophysically to the end of the refractory period imposed by the afterhyperpolarizing current. Notice that there is a cusp at the point where the outer and terminal boundary solution come together. In addition, the perturbative solution's phase arrives at zero before the actual solution's. This suggests that there are finer scaled dynamics arising from the phase variable being small in the vicinity of the saddle-node bifurcation of the fast subsystem. Such effects could potentially be explored with higher order asymptotics. For the purposes of this study, it suffices to truncate the expansion to two terms. The resulting formulae can be utilized extensively in the explanation of network dynamics.
In deriving our approximation to the periodic solution, we were able to calculate a relatively concise formula relating the period of the solution to the remainder of the parameters(4)where is the minimal solution to(5)such that (see Text S1). We illustrate the accuracy of this approximation over a wide range of adaptation time constants in Fig. 5. The approximation is fairly accurate for a substantial region of parameter space, but improves appreciably as and are increased.
We conclude our study of the periodic solution to (2) by using our formula for the period (4) to roughly calculate the number of clusters admitted by a network of adapting neurons with pulsatile inhibitory coupling. This also provides us with an estimate of the population spike frequency. Any inputs delivered to the neuron during the initial or the outer layer stage of the solution, equation (3), will have little or no effect on its firing time. During this interval, the adaptation variable constrains the phase so that it simply relaxes back to the same point on the trajectory following a perturbation. Once the terminal layer begins, the input is above a threshold such that the phase can increase at an accelerating rate. However, it is possible to hold the phase back with a negative perturbation. A neuron that has already begun its terminal phase when another cell spikes will always be forced to delay its own spike. As a result, over time, in a network, clusters of neurons would be forced apart to about the time length of the terminal layer. Therefore, the number of clusters will be roughly determined by the length of this terminal layer as compared with the total length of the period(6)Therefore, as the adaptation time constant increases, the number of clusters will scale as . While our main interest in this formula is its relationship to the adaptation time constant, there are also nonlinear relationships derived here between cluster number and other parameters. We shall compare this formula further with the predictions we calculate using weak coupling and the phase-resetting curve. Since the perturbative solution ceases its slow dynamics briefly before the numerical solution (see Fig. 4), we expect that this asymptotic formula (6) approximating cluster size may be a slight underestimate.
Nonetheless, it allows us to concisely approximate how the population frequency depends on the adaptation time constant as well as the cluster number . Since each neuron spikes with a period given by equation (4) and there are clusters of such neurons, the frequency of populations spikes in the network are given by(7)We plot this function versus as well as in Fig. 6. Notice, networks with neurons whose spike frequency adaptation have a longer time constant support synchronous spiking rhythms with lower frequencies, as in the Traub network (see Fig. 1). Also, by our mechanism, as more clusters are added, the population frequency decreases. This is due to the period of individual neuron spiking scaling more steeply with adaptation time constant than the cluster number.
We have identified general relationships between the adaptation time constant and two quantities of the idealized spiking network (1): the period of a single neuron and the cluster number of the network. These relationships help characterize the behavior of the cluster state in the adaptive network. In particular, the bifurcation structure of the fast-slow formulation of the single neuron system guides the identification of a timescale of the spike phase, which evidently guides network level dynamics. Singular perturbation theory is indispensable in making this observation.
As a means of studying the susceptibility of a single neuron to synchronizing to input from the network, we shall derive the phase-resetting curve of a neuron with adaptation. Biophysically, the phase-resetting curve corresponds to the amount that brief inputs to a tonically spiking neuron delay or advance the time of the next spike. First, we make a change of variables to the system (2), so the state of the neuron is now described by the quadratic integrate and fire (QIF) model with adaptation [58](8)We show in Text S1 that by using a sequence of further changes of variables, we are able to express the periodic solution to this system in terms of special functions. As has been shown previously, the solution to the adjoint equations of a system that supports a limit cycle is the infinitesimal phase-resetting curve (PRC) of the periodic orbit [59]. Therefore, with the function form of in hand, we can derive the adjoint equations by first linearizing the system (8) about the limit cycle solution so(9)The adjoint equations, under the inner product(10)will be(11)(12)Since is known, it is straightforward to integrate (11), to solve for the first term of the adjointBy plugging in (see Text S1), we find we can further specifywhere is given up to a scaling factor in Text S1.
It is now straightforward to plot the PRC of the QIF model with adaptation. To our knowledge, this is the first exposition of an analytic calculation of the PRC of the QIF model with adaptation. Although, the bifurcation structure of more general QIF models with adaptation has been analyzed in previous work by [60], [61]. The exact period can be computed using the right boundary condition given in Text S1, which can then be used to determine the initial condition for the adaptation variableWe then must plot a function which involves a Bessel function of imaginary order and imaginary argument(13)In Fig. 7, this is shown along with the numerically computed PRC, where pulsatile inputs are applied at discrete points in a simulation. Time is also normalized by the period to yield the phase variable . We find an excellent match between the two methods.
One can also derive a very accurate representation of the PRC by numerically solving the adjoint equations (11) and (12). This is also useful because Bessel functions with pure imaginary order and argument are particularly difficult to approximate as the magnitude of the order and argument become large. Accurate asymptotic approximations for this class of special functions are lacking, although [62] provides some useful formulae along these lines. Thus, we compute the PRC using numerical solution of the QIF system (8) and the adjoint equation (11), pictured in Fig. 8 for several different values. Time is normalized here, as in Fig. 7, so the phase variable goes between zero and one. This also eases comparison for different time constants . We find that, as we would suspect from our singular perturbation theory calculations, the region in which the neuron is susceptible to inputs shrinks as increases. This skewed shape to the PRC has been revealed previously in other studies of spiking models, where adaptation currents were treated in alternative ways [51], [63]. We also compute the PRC for the theta model numerically using the adjoint equations. To derive them, we linearize the system (2) about the limit cycle solution soThe adjoint equations, under the inner product (10) will be(14)(15)By solving (2) numerically, we can use the solution to then numerically integrate (14) to solve for the first term of the adjoint , which is the PRC of the theta model. We show this alongside the numerically calculated PRCs of the QIF model. Notice they are quite alike, save for the theta model's PRC being nonzero at . In the theta model's PRC, the change of variables creates a discontinuity.
Therefore, as revealed by an analytic formula and numerical method for computing the PRC, we find that spike frequency adaptation creates a lengthy time window during which the neuron is insensitive to inputs. As the time constant of adaptation is increased, this window occupies more of the solution period. With these formulations of the PRC in hand, we may carry out a weak coupling analysis of the network to quantitatively study predictions regarding solutions that emerge from instabilities of the incoherent state.
Due to large scale spiking network models usually being analytically intractable, a weak coupling assumption is commonly used to study their resulting activity patterns. This allows the reduction of each cell's set of equations to a single one for the phase [38]. Based on the averaging theorem, this reduction is valid as long as parameters of the model are such that each unit supports a limit cycle, their firing rates are not too heterogeneous, and coupling between units is not too strong [15], [59]. This also allows us to place our work in the context of previous studies of clustering in phase models [42]–[44].
Presuming the cells receive enough input to spontaneously oscillate and that they are weakly coupled, we can reduce the system to a collection of limit cycle oscillators [38]. Each oscillator will have some constant frequency , where we use the period computed using the exact solution (see Text S1) for a particular set of parameters. Thus, the network becomes(16)where is the coupling function attained by convolving the PRC with the synaptic timecourse(17)and is a white noise process such that and . To analyze the system (16), we consider the mean field limit . Mean field theory has been used extensively to study (16) when [64]–[66], but much less so when [67], [68]. Following such previous studies, we can employ a population density approach where oscillators are distributed in a continuum of phases so that denotes the fraction of oscillators between and at time . Thus, is nonnegative, -periodic in , and normalizedTherefore, evolves according to the Fokker-Planck equation [64], [65](18)where the instantaneous velocity of an oscillator isthe continuum limit of .
Now, in order to examine the effect that the phase-resetting curve has upon the solutions to (16), the weak coupling approximation to (1), we shall study instabilities of the uniform incoherent state of (18), given by . It is straightforward to check that this is indeed a solution by plugging it into (18). Since this is always a solution, for all parameters, we can examine the solutions that emerge when it destabilizes by studying its linear stability. We will show that for sufficiently large, the incoherent state is stable, but as is reduced, the solution destabilizes, usually at a unique Fourier eigenmode. We begin by lettingwhere . Expanding the continuity equation (18) to first order in , we arrive at an equation for the linear stability of the incoherent state(19)where . Expressing as a Fourier seriesand specifically taking , we can compute the eigenvalue of the th mode of using the spectral equation of the linear system (19), soApplying the change of coordinates , we have a general equation for the th eigenvalue(20)We can evaluate the integral term by considering the Fourier series expansion(21)so thatUpon plugging this into (20), we find the eigenvalue associated with the th mode of is related to the Fourier coefficients of by(22)Thus, as is reduced towards zero, the first eigenmode to destabilize will be the one whose eigenvalue crosses from the left to the right half of the complex plane first. Using equation (22), we can identify this mode as the first to have Re orThis corresponds to the for which is maximal. For the critical value at which the first eigenvalue has positive real part, we show plots of as a function of for several different parameters in Fig. 9. Notice that as the adaptation time constant is increased, and other parameters are held fixed, the critical increases. As the synaptic time constant is increased and other parameters are held fixed, the critical decreases. We contrast this with the case of excitatory coupling () in the system (1), where the PRC is nonnegative. In this case, the critical is fairly insensitive to changes in the time constants, virtually always predicting the mode becomes unstable first (not shown). Therefore, our weak coupling calculation approximates the number of clusters for a given set of parameters using the coupling function (17) with the Fourier expansion (21) so that(23)To compare with our singular perturbation theory results, we compute the approximate number of clusters using the weak coupling assumption for pulsatile synapses. In the limit , the coupling function becomes . Therefore, the Fourier coefficients are calculated directly from the PRC of the theta model. In Fig. 10, we plot the number of clusters as a function of , calculated using equation (23) along with the asymptotic approximation to the number of clusters (see equation (6)). Notice that the singular perturbation theory slightly underestimates as compared with weak coupling. This may be due to the fact that the singular perturbative solution reaches the saddle-node point slightly before the actual solution does, underestimating the length of the quiescent phase of the PRC. Nonetheless, both curves have a characteristic sublinear shape. We show in Fig. 11 that the weak coupling dependence upon scales as a power law, just as predicted by singular perturbative theory. Thus, even though our asymptotic approximation (6) is an underestimate, it provides us with the correct scaling for cluster number dependence upon adaptation time constant. The same power law scaling is reflected in networks with exponentially decaying synapses, as shown in Fig. 12. We plot predictions based on our weak coupling assumption for . As the synaptic time constant is increased, the number of clusters is diminished, since feedback inhibitory inputs relax more slowly. Therefore, we speculate an improved asymptotic approximation of cluster number that accounts for synaptic timescale might include an inverse dependence upon .
In this section, we present results of numerical simulations of the idealized network (1) of theta neurons with global inhibition and adaptation. In addition, we compare the scaling law predicted for the idealized model to the number of clusters arising in numerical simulations of the more detailed Traub model. We find that the qualitative predictions of our singular perturbation theory and weak coupling approximations are reflected in the dependence of the state of the network on model parameters. The quantitative relationship between adaptation time constant and cluster number is sensitive to the strength of global inhibitory feedback , holding for small values only. One would expect this, since approximations were made considering weak coupling.
In Fig. 13, we show the results of simulations for various adaptation time constants in the case of pulsatile synapses (). As predicted by the formulae of both our singular perturbation theory approximation (6) and weak coupling assumption (23), cluster number increases sublinearly with adaptation time constant. Notice in Fig. 13(c), when there are seven clusters, neurons of each cluster do not spike in as tight of a formation as can be found in simulations with four and six clusters. We conjecture that this is due to fewer neurons participating in each cluster and so less global inhibition is recruited each time a set of neurons fires. This smears the boundary between each cluster. In Fig. 14, we show the results of simulations in the case of exponentially decaying synapses with time constant . As predicted by our weak coupling analysis, the smoothing of the synaptic signal leads to there being fewer clusters on average for a particular value. Notice in both the pulsatile and exponential synapse cases, as the number of clusters increases, the interspike intervals are prolonged, as predicted by our approximation of the period (4). Therefore, the resulting frequency of population activity decreases, on average, with .
To quantitatively compare our theoretical predictions with numerical simulations of (1), we plot the minimal necessary to generate the number of clusters for each method. Theoretical calculations include both the singular perturbation approach (6) and the weak coupling approximation (23). The points we then plot in Fig. 15 correspond to the first value of whose median cluster number is larger than the median for the previous value (see Methods). Remarkably, the theoretical calculation using the weak coupling approach give a reasonable approximation to the behavior of the simulations. Comparing the result of pulsatile versus exponentially decaying synapses, the increase in with is clearly larger for the pulsatile synapse case. This can be contrasted with the results of van Vreeswijk, who found in simulations of inhibitory integrate and fire networks that median cluster number increased with synaptic timescale [47]. One particular aspect of simulations of the full model (1) that may escape our theoretical formulae (6) and (23) is the effect of different synaptic strengths. To produce fairly well resolved clusters, it was necessary to take , not very weak. Additionally, as the number of clusters increases, the strength of inhibitory impulses decreases. Both of these facts may bear upon potential cluster number and account for the nonlinear shape of the numerically developed relationship between and .
Finally, we return to the original detailed biophysical model to compare the predictions of cluster scaling made in the idealized model. Exchanging the idealized adaptation time constant for the time constant for calcium dynamics in the Traub model, , we examine how well the scaling holds in numerical simulations of the detailed model. We use the same method as that employed for the idealized model to identify the minimal at which a certain number of clusters appears (see Methods). Our results are summarized in Fig. 16 and show that, in fact, cluster number does approximately follow the adaptation time constant scaling predicted from the idealized model. This makes sense, since one can relate the Traub model to the idealized theta model using a normal form reduction, so their phase-resetting properties will be similar to a first approximation [38]. The quiescence invoked by strong adaptation will lead to sharp narrow peaks in the PRC for the Traub model (as shown for the idealized model in Fig. 8(b)). Therefore, our analysis of the theta model leads to an excellent prediction of the effects of adaptation upon the cluster state in the network of Traub neurons.
In this paper, we have studied the formation of cluster states in spiking network models with adaptation. We theorize clustering may be an alternative, or at least contributing, mechanism for the sparse firing of pyramidal cells during gamma rhythms [12]. Sparse gamma rhythms may, therefore, not rely solely upon the effects of input and connectivity heterogeneities [30]. Besides spike frequency adaptation, the other essential property for the formation of clusters in the network is feedback inhibition. Empirically, we observe the number of clusters increases with the time constant of adaptation in a detailed biophysical spiking network and a more idealized model. We can carry out a number of analytical calculations on the idealized model that help uncover the mechanisms of clustering. Results of a singular perturbative approximation of a single neuron's periodic spiking solution confirm that adaptation with longer timescales will shorten the relative length of time a neuron is susceptible to inputs. This is revealed in a compact expression (4) relating the period of the neuron to parameters. In particular, we can estimate the number of clusters generated in the network for a particular value of adaptation time constant and find they will scale as . We then compare this result to a formula that can be derived in the context of a phase model, where, incidentally, the phase-resetting curve can be computed exactly. In the weak coupling limit, the number of clusters is related to the Fourier modes of the phase-resetting curve. In fact, we can fit the number of clusters to a power law. These results are confirmed in simulations of the full idealized model (1) and are well matched to simulations of the detailed biophysical model.
Our results suggest a number of experimentally testable predictions. We have suggested that clustered states may be an organized synchronous state capable of generating sparse gamma rhythms [1]. Rather than a rhythm generated by a balanced network containing neurons with driven by high amplitude noise [30], gamma may be a rhythm generated by slow excitatory neurons that cluster into related groups temporarily but dissociate from one another after some length of time. This could be probed using multiunit recordings to look for clustering of pyramidal neurons on short timescales. Large networks that exhibit clustering may do so through this combination of adaptation and inhibition. This suggests that it may be possible to identify in vitro or in vivo clustering that depends upon spike frequency adaptation by examining the effects of curtailing calcium dependent potassium currents using cadmium, for example [69]. Our model suggests weakening spike frequency adaptation should lead to a decrease in cluster number. In addition, there are a growing number of ways to experimentally measure the PRC of single neurons [63], [70]. Since pyramidal cells are known to often possess adaptation currents, it may be possible to study the ways in which modulation of those currents' effects bears on a neuron's associated PRC. Our analysis indicates that stronger and slower spike frequency adaptation leads to PRCs with a steep peak at the end. Thus, different aspects of the cluster state shown here may be studied experimentally in several ways.
Clustering through intrinsic mechanisms may in fact be a way for networks to generate cell assemblies spontaneously [71]. If clustering is involved in the processing of inputs, shifting neurons from one cluster to another might disrupt the conveyance of some memory or sensation [10], [14]. In more specific networks, underlying heterogeneous network architecture may provide an additional bias for certain neurons to fire together. Alternatively, cell assemblies may be formed due to bias in the input strength to a recurrent excitatory-inhibitory network, as shown in [49]. They found that the inclusion of hyperpolarizing current could generate slow rhythms in the excitatory neurons with increased input. Our model does rely on a hyperpolarizing current but does not require a heterogeneity in the input. Also, each assembly possesses its own beta rhythm whereas the entire network possesses a gamma rhythm.
In the future, it would be interesting to pursue a variety of the theoretical directions suggested by our results. The singular perturbation calculation follows along the lines of a few previous studies of canards in the vicinity of fold singularities [55]–[57], [72]. Carrying out an even more detailed study of the bifurcation structure of the fast-slow system of the single neuron (2) may allow for a more exact calculation of how the period relates to the parameters. In particular, we may be able to compute the dynamics of relaxation time in the vicinity of the bottleneck near the saddle-node bifurcation of the fast system (see Fig. 3). We could also extend this calculation to other idealized spiking models with adaptation such as Morris-Lecar [55] or the quartic integrate and fire model [60]. In addition, we have considered examining the types of dynamics that may result in inhibitory leaky integrate and fire networks with adaptation. Excitatory integrate and fire networks have previously been shown to support synchronized bursting when possessing strong and slow enough adaptation [52]. It has also been shown that inhibitory integrate and fire networks without adaptation support clustering in the case of alpha function synapses [47]. In preliminary calculations, we find that a single integrate and fire neuron with strong and slow adaptation does not have the same steep peaked PRC as the theta model, due to there being no spike signature in the model. Therefore, it may not support clustered states through the same mechanism as the system we have studied. We have also mentioned that clustering arises in the network (1) through the application of a homogenous deterministic current with some additive noise. Therefore, applying an input with more temporal structure, for example at the frequency of the network or individual neurons, may lead to interesting variations of the clustered state. Finally, we seek to study other potential negative feedback mechanisms for generating clusters. In a large competitive spiking network, it may be possible for a subset of neurons to suppress the rest until synaptic depression exhausts inhibition. Multistable states supported with such mechanisms have been shown in small spiking networks [73], [74], but theory has yet to be extended to large scale synchronous states like clustering.
|
10.1371/journal.ppat.1008025 | KSHV-encoded LANA protects the cellular replication machinery from hypoxia induced degradation | Kaposi’s sarcoma associated herpesvirus (KSHV), like all herpesviruses maintains lifelong persistence with its host genome in latently infected cells with only a small fraction of cells showing signatures of productive lytic replication. Modulation of cellular signaling pathways by KSHV-encoded latent antigens, and microRNAs, as well as some level of spontaneous reactivation are important requirements for establishment of viral-associated diseases. Hypoxia, a prominent characteristic of the microenvironment of cancers, can exert specific effects on cell cycle control, and DNA replication through HIF1α-dependent pathways. Furthermore, hypoxia can induce lytic replication of KSHV. The mechanism by which KSHV-encoded RNAs and antigens regulate cellular and viral replication in the hypoxic microenvironment has yet to be fully elucidated. We investigated replication-associated events in the isogenic background of KSHV positive and negative cells grown under normoxic or hypoxic conditions and discovered an indispensable role of KSHV for sustained cellular and viral replication, through protection of critical components of the replication machinery from degradation at different stages of the process. These include proteins involved in origin recognition, pre-initiation, initiation and elongation of replicating genomes. Our results demonstrate that KSHV-encoded LANA inhibits hypoxia-mediated degradation of these proteins to sustain continued replication of both host and KSHV DNA. The present study provides a new dimension to our understanding of the role of KSHV in survival and growth of viral infected cells growing under hypoxic conditions and suggests potential new strategies for targeted treatment of KSHV-associated cancer.
| Hypoxia induces cell cycle arrest and DNA replication to minimize energy and macromolecular demands on the ATP stores of cells in this microenvironment. A select set of proteins functions as transcriptional activators in hypoxia. However, transcriptional and translational pathways are negatively regulated in response to hypoxia. This preserves ATP until the cell encounters more favorable conditions. In contrast, the genome of cancer cells replicates spontaneously under hypoxic conditions, and KSHV undergoes enhanced lytic replication. This unique feature by which KSHV genome is reactivated to induce lytic replication is important to elucidate the molecular mechanism by which cells can bypass hypoxia-mediated arrest of DNA replication in cancer cells. Here we provide data which shows that KSHV can manipulate the DNA replication machinery to support replication in hypoxia. We observed that KSHV can stabilize proteins involved in the pre-initiation, initiation and elongation steps of DNA replication. Specifically, KSHV-encoded LANA was responsible for this stabilization, and maintenance of endogenous HIF1α levels was required for stabilization of these proteins in hypoxia. Expression of LANA in KSHV negative cells confers protection of these replication proteins from hypoxia-dependent degradation, and knock-down of LANA or HIF1α showed a dramatic reduction in KSHV-dependent stabilization of replication-associated proteins in hypoxia. These data suggest a role for KSHV-encoded LANA in replication of infected cells, and provides a mechanism for sustained replication of both cellular and viral DNA in hypoxia.
| Kaposi Sarcoma associated herpesvirus (KSHV) or Human herpesvirus 8 (HHV8) infects human endothelial cells and B-lymphocytes and is strongly associated with Kaposi sarcoma (KS), Pleural Effusion Lymphoma (PEL) and Multicentric Castleman’s Disease (MCD) [1–4]. Like other herpesviruses, KSHV maintains the viral genome as extra-chromosomal episomes in latently infected cells with only a limited number of KSHV-encoded genes expressed [5–7]. Upon successful infection and establishment of latency, cellular transformation by KSHV relies upon its ability to degrade tumor suppressors or activating pro-oncogenic factors [8–11], though immune competency of infected individual plays a critical role in pathogenesis of KSHV infection [12].
KSHV-encoded latency associated nuclear antigen (LANA) is the major factor responsible for maintaining latency as well as tethering the viral episomal DNA to host chromatin [5, 13–15]. LANA binds directly to the terminal repeats, which contains the minimal replication unit through its carboxy-terminus while binding to cellular chromatin through its amino-terminus [16, 17]. For persistent replication of the KSHV genome, LANA also recruits the clamp loader proliferating cell nuclear antigen (PCNA) to the KSHV genome [18]. Epigenetic reprogramming of the KSHV genome is another key requirement for maintaining latent infection and escaping from host immune response by switching off expression of the majority of the genes [19, 20]. Recent studies demonstrated genome wide changes in methylation patterns, as well as histone modifications throughout the steps of infection for establishment of latency or lytic reactivation post-infection [21–23]. In normoxia, the KSHV genome replicates once per cell cycle to maintain the gross copy number, and its replication is dependent on the host cellular machinery. The inhibition of KSHV replication through Geminin, an inhibitor of Cdt1 and mammalian replication confirmed the involvement of host regulatory factors in latent replication of KSHV[17]. Additionally, expression of Cdt1, rescued the replication ability of plasmids containing the KSHV minimal replicator element [17].
LANA is involved in recruiting the DNA clamp loader PCNA to mediate efficient replication and persistence of KSHV [18]. We have also previously identified and characterized another latent origin, which supports replication of plasmids ex-vivo without LANA expression in trans and prompted our investigation using single molecule analysis of replicated DNA (SMARD)[24, 25]. The study resulted in identification of multiple replication initiation sites within the entire KSHV genome [25]. Chromatin immuno-precipitation assays performed using anti-origin recognition complex 2 (ORC2), and LANA antibodies from nuclear extracts of cells containing plasmids RE-LBS1/2, RE-LBS1, LBS1, or RE showed an association of ORC2 with the RE region[17]. Similarly, other host trans factors like MCMs was shown to be associated with the replication initiation complex [25].
Hypoxia and the hypoxia inducible factor HIF1α play a critical role in pathogenesis of KSHV by modulating expression of critical KSHV-encoded genes, as well as stabilizing several KSHV-encoded proteins. Although only a few hypoxia responsive elements (HREs) within promoters of KSHV-encoded genes have been validated, there are hundreds of uncharacterized HREs present across KSHV genome. The most important HREs characterized within the promoters of KSHV genes are those within the regulatory regions of LANA, the reactivation and transcriptional activator (RTA), and the viral G-Protein coupled receptor (vGPCR) [26–28].
LANA is involved in maintenance of KSHV latency, and it promotes tumorigenic properties through either activation of oncogenic pathways or repression of apoptotic pathways [14, 29]. RTA is involved in transcriptional activation of KSHV-encoded genes and lytic replication of the KSHV genome [30, 31]. The direct involvement of hypoxia in KSHV lytic replication have been demonstrated by a number of studies which showed that HIF1α facilitated KSHV-encoded RTA-mediated reactivation by binding to LANA to upregulate RTA expression [27]. Hypoxia is also reported to enhance the viral reactivation potential of the well-known reactivating compound 12-O-tetraecanoylphorbol-13-acetate [32]. Furthermore, the role of hypoxia in maintenance of latency is also crucial, where promoters of the key latent gene cluster coding for LANA, vFLIP and vCyclin harbor hypoxia responsive elements regulated by HIF1α [28]. Hypoxia-dependent expression of vGPCR is well known for modulating expression of several metabolic genes through ROS dependent epigenetic modifications [26]. It is also important to note that vGPCR up-regulated expression of HIF1α through activation of the MAPK kinase signaling pathway through the targeting of P38 [33]. This HIF1α-vGPCR positive feedback mechanism may explain in part the elevated levels of HIF1α in KSHV infected cells. The elevated levels of HIF1α in KSHV-infected cells was shown to modulate several pathways essential for cell proliferation, apoptosis, angiogenesis and metabolic reprogramming [34].
Hypoxia is a detrimental stress to aerobic cells and a consequence of restricted blood supply in the context of in-vivo conditions [35]. Cessation of cell cycle progression, and DNA replication are the main adaptive response of cells to minimize their energy and macromolecular demands [36, 37]. Additionally, transcription factors specifically stabilized in hypoxia (Hypoxia inducible factors; HIFs) regulate transcription of a number of genes responsible for reprogramming cell metabolism and promote survival [38, 39]. Stabilized HIF1α is also recognized as a negative regulator of cell division and replication through its non-transcriptional associated functions [40–42]. Furthermore, HIF1α knock-down abrogates hypoxia-mediated cell cycle arrest and promotes DNA replication in the subsequent synthesis phase [42]. It is well established that KSHV infection promotes HIF1α stabilization, and paradoxically further exposure of KSHV positive cells to hypoxia induces lytic replication. [27, 32]. The differential character of replication of the KSHV genome under hypoxic conditions begs the exploration of how KSHV manipulates the replication machinery to promote latent replication under hypoxia, a non-permissive and unfavorable condition. In this study we now show that KSHV not only allows an efficient transition to S-phase by stabilizing CyclinE/CDK2, but also protects critical replication-associated proteins involved in origin recognition, initiation and elongation from hypoxia-dependent degradation. In addition, KSHV-encoded LANA in conjunction with host-encoded HIF1α is necessary for efficient replication in the hypoxic microenvironment.
Cellular adaptive response towards hypoxia includes cessation of cell cycle progression and DNA replication to minimize energy demand and to ensure cell survival. In contrast to normal cells, the KSHV genome in infected cells is known to undergo reactivation and lytic replication [27, 32]. Interestingly, both the latency associated nuclear antigen (LANA) and replication and transcriptional activator (RTA), are up-regulated under hypoxic conditions. Where the former promotes cellular proliferation and oncogenesis, and the latter is the essential mediator of lytic replication[14, 30, 43]. BJAB-KSHV cells were used to investigate the role of KSHV in modulating cellular proliferation and DNA replication events under hypoxic conditions as compared with KSHV negative BJAB cells. The two cells lines were checked for their isogenic background by short tandem repeat (STR) profiling and after thawing a similar passage number of cells was used for the experiment [26]. As expression of transcript or protein levels of HIF1α does not represent a good marker of long-term hypoxic induction [44], PDK1 levels (a transcriptional target of HIF1α) was used to demonstrate induction of hypoxia (Fig 1A). Cell cycle analysis of BJAB and BJAB-KSHV cells grown in hypoxia for different time periods clearly indicated that the presence of KSHV can facilitate the G1/S transition under hypoxic conditions (Fig 1B and 1C and S1 Fig). Hypoxia induces arrest of G1/S transition [42], and bypassing this arrest is essential for entry and subsequent DNA replication and cellular proliferation. Therefore, we hypothesized that KSHV can manipulate the cellular machinery to bypass this arrest and promote DNA replication. To investigate this, we checked the status of Cyclins (Cyclin D1 & Cyclin E) and Cyclin dependent kinase (Cdk2), associated with G1/S transition (Fig 1D) [45]. We first investigated the status of Cyclin D1, Cyclin E and Cdk2 at the transcript level in BJAB and BJAB-KSHV cells growing under normoxic or hypoxic conditions at various time points (Fig 1E, 1F and 1G and S2 Fig). The results suggested that hypoxia exerts a similar effect on the transcription expression of these genes. Briefly, an almost 50% reduction in expression of Cyclin D1 and Cyclin E was observed at 24 hours of hypoxia treatment in both BJAB and BJAB-KSHV cells. Similarly, a 75% reduction in expression of Cyclin E was observed at 36 hours of hypoxia treatment in both BJAB and BJAB-KSHV cells (Fig 1E and 1F and S2 Fig). Interestingly, the expression of Cdk2 in these cells was observed to be independent of hypoxia with no significant differences seen in expression at any time periods (Fig 1G and S2 Fig). We, therefore, hypothesized that KSHV may affect the expression of these factors at the protein level. Based on the transcript analysis of the expression, we choose to analyze the expression of these proteins after 36 hours of hypoxic treatment. Cells were grown in normoxic or hypoxic conditions for 36 hours followed by analysis by Western blot to detect the differences in the protein levels. Interestingly, we observed that levels of both Cyclin D1 and Cyclin E as well as Cdk2 were significantly reduced in KSHV negative BJAB cells (Fig 1H). The presence of KSHV in BJAB-KSHV cells had a protective effect on these proteins from hypoxia-associated degradation (Fig 1H, compare lane 2 and 4). Furthermore, to corroborate these results, we performed the same experiment in HEK293T and HEK293T-BAC16-KSHV cells. The protection of Cyclins and Cdk2 protein levels in KSHV positive HEK293T-BAC16-KSHV compared to HEK293T confirmed that KSHV was able to block hypoxia-dependent degradation to allows S-phase entry and subsequent DNA replication (Fig 1I, compare lane 2 and 4).
Origin recognition by origin recognition complex (ORCs) proteins and formation of pre-initiation complex are the initial event of DNA replication[46]. Therefore, we investigated whether the presence of KSHV can differentially modulate origin recognition and pre-initiation steps under hypoxic conditions. A schematic showing the comprehensive list of proteins involved in origin recognition and pre-initiation of DNA replication are shown (Fig 2A). We investigated the levels of origin recognition complex proteins (ORC1-6) at both transcript and protein levels in BJAB and BJAB-KSHV cells growing under normoxic or hypoxic conditions. Additionally, levels of cell division cycle 6 (CDC6; essential component for assembly of pre-replication complex), Cdt1 (replication licensing protein essential for loading of minichromosomal maintenance proteins, MCMs), and a representative of MCMs (MCM3) were also measured at the transcript and protein levels (Fig 2A–2F). Among the ORCs, ORC1 and ORC4 appeared to be transcriptionally stable at the both 24 and 36 hours of hypoxic treatment (1%O2) with only a marginal down-regulation in both the BJAB and BJAB-KSHV cells (Fig 2B and S3A–S3D Fig). The expression of ORC2, ORC3 and ORC5 showed changes in expression profiles at transcript levels in both the BJAB and BJAB-KSHV cells and at both time points (24 and 36 hours). Briefly, ORC2 showed a down-regulation by nearly 20% in both BJAB and BJAB-KSHV at the end of 24 hours of hypoxic treatment. ORC2 expression was further down-regulated by nearly 50% at 36 hours of hypoxic treatment (Fig 2B). The expression of ORC3 at the transcript level showed an intermediate effect where the fold change difference was not significantly different in both BJAB and BJAB-KSHV cells at the end of 24 hours of hypoxic treatment. However, at 36 hours, expression of ORC3 was down regulated at nearly 50% in both BJAB and BJAB-KSHV cells (S3B Fig). Among the ORCs investigated, ORC5 showed the most drastic effects of hypoxia, where the fold change of ORC5 transcripts was observed to be nearly 50% down regulated at the end of 24 hours of hypoxic treatment, and by the end of 36 hours of hypoxic treatment, almost an 80% decrease was observed in both BJAB and BJAB-KSHV cells (S3D Fig). MCM3, critical for origin recognition was also dramatically down-regulated by greater than 80% in 24 hrs of hypoxia treatment and over 90% by 36 hrs (Fig 2E). The effects in BJAB-KSHV was similar but not substantially greater than in BJAB alone (Fig 2E). We hypothesized that similar to cyclins and CDK2, KSHV may influence the expression of ORCs and MCM3 at the protein level.
Western blot analyses were also performed to monitor these proteins using lysates from BJAB and BJAB-KSHV cells grown under normoxic or hypoxic conditions for 36 hours. Results clearly suggested that the presence of KSHV had a protective effect on these proteins from hypoxia-mediated degradation (Fig 2F and S3E Fig). We further corroborated these results in another KSHV positive cell lines, HEK293T-BAC16-KSHV cells compared to HEK293T cells. Similarly, protection of these proteins in HEK293T-BAC16-KSHV cells grown under hypoxic conditions confirmed that KSHV infection played a role in origin recognition by providing a level of protection for the origin recognition proteins from hypoxia-mediated degradation (S3F Fig).
Results showing KSHV-mediated protection of cell cycle, and DNA replication associated proteins from hypoxia-dependent degradation stimulated further investigation of key proteins involved in initiation of DNA replication. We investigated the levels of cell division cycle 45 (CDC45; essential for the loading of DNA polymerase 1 alpha on replication complex)[47], and DNA polymerase 1 alpha (DNA pol 1α; the rate limiting DNA polymerase with primase activity)[48] at the transcript, and protein levels in BJAB and BJAB-KSHV cells grown under normoxic or hypoxic conditions. Real-time expression analysis of these genes showed that the presence of KSHV did not provide a dramatic differential at the transcript level for CDC45 in normoxic or hypoxic conditions, but it up-regulated expression of DNAPol1 alpha to approximately 1.6-fold in normoxic condition (Fig 3A and 3B). Furthermore, at the 36 hour time point, hypoxia induced a similar level of down-regulation of transcripts by approximately 50–75% for the transcripts of both CDC45 and DNA pol 1α in BJAB and BJAB-KSHV cells grown under hypoxic conditions (Fig 3A and 3B).
Similar to the expression of genes involved in origin recognition and pe-initiation, we hypothesized that KSHV may influence the levels of these factors at the level of post-translation. Western blot analyses were performed to monitor CDC45 and DNA Pol 1α proteins in BJAB and BJAB-KSHV cells grown in normoxic or hypoxic conditions for 36 hours. The results indicated a clear protection of these replication proteins from hypoxia-mediated degradation by the presence of KSHV (Fig 3C compare lane 2 and 4). We further examined the role of KSHV as a major contributor to the inhibition of hypoxia-mediated degradation in HEK293T-BAC16-KSHV cells when compared to HEK293T cells (Fig 3D). Similarly, the levels of these proteins in HEK293T-BAC16-KSHV cells grown under hypoxic conditions were enhanced and supported a role for KSHV in origin recognition through protection of origin recognition proteins from hypoxia-mediated degradation.
To further rule out any possible role due to differences in the source of BJAB/BJAB-KSHV or HEK293T/ HEK293T-BAC16-KSHV cells or their passage numbers, we infected PBMCs with KSHV generated from HEK293T-BAC16-KSHV. We infected PBMCs with purified KSHV at a multiplicity of infection equal to 10. Initially, the infected cells were grown under normoxic conditions for 24 hours to allow for the expression of KSHV-encoded genes in the infected cells. KSHV infection of PBMCs was confirmed by GFP signals (for infection with KSHV generated from HEK293T-BAC16-KSHV) (Fig 4A). The mock control or infected PBMCs were then incubated under hypoxic conditions for another 24 hours. The induction of hypoxia was confirmed by western blot against PDK1. The status of the proteins examined above were monitored by western blot. Similar to the results seen in BJAB or HEK293T cells, a significant decrease in the level of all the proteins investigated (CCNE, CDK2, ORC2, MCM3, CDC6, Cdt1, CDC45 and DNAPol1A) was observed (Fig 4C). As expected, KSHV infected PBMCs grown under hypoxic conditions was able to rescue these proteins from hypoxia-mediated degradation (Fig 4C). These results strongly supported a role for KSHV in the rescue of DNA replication-associated proteins from hypoxia-mediated degradation. In another approach, levels of these proteins in KSHV negative BL41 were compared with KSHV positive BCBL1 cells. The cell lines were confirmed for the absence or presence of KSHV by immune staining against LANA protein (Fig 4B). The cells were grown under hypoxic conditions followed by analysis of PDK1 levels for the confirmation of induction of hypoxia. The comparative analysis of replication associated proteins further confirmed the protection potential of KSHV for replication associated proteins from hypoxia-mediated degradation (Fig 4D).
Latently infected KSHV positive cells predominantly express only a limited set of KSHV-encoded proteins. Furthermore, hypoxia is well known to induce expression of other KSHV-encoded genes. The latency associated nuclear antigen (LANA), replication and transcriptional activator (RTA), viral G-Protein coupled receptor (vGPCR) and viral cyclin (vCyclin), are well-established antigens being expressed either in latency or during hypoxic conditions [26–28]. We hypothesized that one, or a combination of these proteins will be responsible for rescuing the DNA replication-associated proteins from hypoxia-mediated degradation. To identify which of the KSHV-encoded protein was responsible for rescuing DNA replication-associated proteins from hypoxia-mediated degradation, we individually expressed these proteins in HEK293T cells along with mock transfection as control. Transfected cells were allowed to grow under normoxic or hypoxic conditions. The expression of KSHV-encoded antigens was confirmed using western blots with antibodies against the epitope tag fused to these proteins (Fig 5A–5D, top panel). The induction of hypoxia was confirmed by western blot analysis of PDK1 (Fig 5A–5D, second panel at the top).
Interestingly, analysis of the levels of proteins involved in DNA replication revealed that ectopic expression of KSHV-encoded LANA efficiently rescued these proteins from hypoxia-mediated degradation. Representative blots for the proteins that clearly showed that degradation occurred in hypoxia are shown (Fig 5A). Importantly, the expression of other KSHV-encoded antigens such as RTA, vGPCR or vCyclin showed little or no ability to protect these proteins from hypoxia-mediated degradation (Fig 5B–5D).
We then wanted to investigate the mechanism of how KSHV-encoded LANA protected these proteins from hypoxia-mediated degradation. KSHV-encoded LANA can interact with a number of replication-associated proteins such as ORC2 and MCM3 under normoxic conditions[49]. Therefore, we hypothesized that LANA may interact with these proteins to block their degradation in hypoxia by interfering with their ubiquitination. To demonstrate their interaction, FLAG tagged LANA was expressed in HEK293T cells (Fig 5E). Cells were grown under hypoxic conditions followed by immuno-precipitation assays using anti-FLAG antibodies. The results indicated that LANA associated with the select set of replication-associated proteins when grown under hypoxic conditions (Fig 5F). Further validation of these associations in cellular complexes was examined using immuno-fluorescence assays taking ORC2 and MCM3 as representative proteins. The results showed that LANA and ORC2 or LANA and MCM3 co-localized under hypoxic conditions which corroborated their association in cellular replication compartments (Fig 5G and 5H).
To further support the role of LANA in protecting DNA replication-associated proteins from hypoxia-mediated degradation, we investigated whether knock down of KSHV-encoded LANA in KSHV infected cells resulted in a loss of protection potential due to the presence of KSHV. The lentivirus-mediated knock down of KSHV-encoded LANA in BC3 cells was previously described[50]. BC3-ShControl or BC3-ShLANA cells were monitored for GFP fluorescence as well as the down-regulated expression of LANA at the protein levels (Fig 6A and 6B, top panel). BC3-ShControl or BC3-ShLANA cells were incubated under normoxic or hypoxic conditions for 36 hours. This was followed by investigation of the levels of a representative set of proteins involved in DNA replication. Western blot analysis for the proteins in these cells confirmed that LANA was a crucial viral antigen required for inhibition of degradation of these proteins under hypoxic conditions (Fig 6B). Briefly, the levels of CCNE, CDK2, ORC2, MCM3, CDC6, Cdt1, CDC45 and DNAPol1A were investigated. As expected, BC3-ShControl cells showed almost similar levels of these proteins independent of whether grown under normoxic or hypoxic conditions (Fig 6B; lane 1 compared to lane 2). However, BC3-ShLANA cells grown under hypoxic conditions showed a relatively lower level of these proteins under hypoxic conditions when compared to normoxic conditions (Fig 6B, lane 2 compared to lane 4). Further, the role of LANA knock-down was validated by single molecule analysis of replicated DNA (SMARD). Analysis for replicated vs non-replicated KSHV was performed by pulsing the BC3-ShControl or BC3-ShLANA cells, and visualizing KSHV DNA using KSHV specific probes while replicated DNA was visualized by immuno-staining against IdU/CldU (Fig 6C–6E). A significantly low level of replication was observed in BC3-ShLANA cells compared to BC3-ShControl cells grown under hypoxic conditions (see lower compartment, Fig 6E). Even under the normoxic conditions, knock down of LANA also had a negative effect on KSHV replication, but significantly less compared to hypoxic conditions (Fig 6D). These experiments clearly showed that the differences seen in the stabilization of replication associated proteins in hypoxic conditions was mainly at the protein level and not at the transcript level. We further showed that, that this occurred through inhibition of the ubiquitin-mediated proteosomal degradation system which is targeted by KSHV-encoded LANA. Cells grown in hypoxic condition with medium containing proteosomal inhibitor MG132 was compared with cells grown under normoxic or hypoxic conditions. The results strongly suggested that the presence of MG132 had a protective effect on these proteins from hypoxia-mediated degradation (S4A Fig). Also, a representative protein CDC6 was used to confirm role of LANA in inhibition of proteosomal degradation in hypoxic conditions. Cells expressing mock or LANA were grown under hypoxic conditions (with or without MG132) followed by immuno-precipitation and western blot with ubiquitin antibody. The results showed that presence of LANA significantly reduces ubiquitination under hypoxic conditions (S4B Fig, compare lane 2 and 4) and suggested that LANA is likely inhibiting the activity of one of the cellular E3-ubiquitin ligase.
HIF1α is a major cellular regulator which plays a critical role in KSHV-mediated oncogenesis and is known to interact at the transcriptional and post-transcriptional levels with KSHV factors to promote the cancer phenotype. Also, HIF1α is required for upregulation of LANA during growth of KSHV positive cells in hypoxia [28]. We wanted to investigate the role of HIF1α in hypoxia-mediated degradation of DNA replication-associated proteins in KSHV negative background or their protection in KSHV-positive background. To study the role of HIF1α, we transduced both BJAB and BJAB-KSHV cells with plasmid vectors containing ShControl or ShHIF1α. Knock down of HIF1α was confirmed by real-time PCR by monitoring the HIF1α transcripts in these cells grown under normoxic or hypoxic conditions (Fig 7A). The effect of HIF1α knock down was further confirmed by investigating expression of P4HA1, a known target of HIF1α after these cells were grown under normoxic or hypoxic conditions (Fig 7B). As expected, the fold change expression of P4HA1 in both BJAB and BJAB-KSHV cells transfected with a ShHIF1α construct was significantly less when compared to the cells containing the ShControl construct under similar conditions (Fig 7B). Upon, confirmation of HIF1α knock down in both BJAB and BJAB-KSHV cells, analysis of the levels of DNA replication- associated proteins was analyzed in cells grown under normoxic or hypoxic conditions. As expected, we observed an almost complete loss of all proteins investigated in BJAB cells transfected with either ShControl or ShHIF1α constructs. Notably, BJAB-KSHV cells showed rescue of these proteins in hypoxia when transfected with ShControl plasmid. Interestingly, knock down of HIF1α in BJAB-KSHV cells showed that the ability of KSHV to protect these proteins from hypoxia-mediated degradation was severely compromised (Fig 7C). These results suggested that HIF1α contributes to the effects of KSHV-encoded LANA-mediated rescue of the replication-associated proteins from hypoxia-mediated degradation. The observation of HIF1α knock down dependent degradation of replication associated proteins in hypoxia were further validated in naturally infected KSHV positive BC3 cells. Generation and characterization of BC3-ShControl and BC3-ShHIF1α were described earlier[26]. BC3-ShControl and BC3-ShHIF1α cells were grown under normoxic or hypoxic conditions followed by investigation of the replication-associated proteins. As expected, the BC3-ShControl cells showed protection from degradation for all the studied proteins in hypoxia while BC3-ShHIF1α cells were unable to protect these proteins under hypoxic conditions (Fig 7D). The results clearly suggest that HIF1α dependent transactivation of LANA is required for rescue of these proteins from degradation in hypoxic conditions as LANA levels were substantially reduced in the BC3 shHIF1α cells. Finally, we investigated whether KSHV reactivation was induced in the hypoxic conditions. We incubated BJAB-KSHV as well as naturally infected BC3 cells under normoxic and hypoxic conditions and estimated the relative yield of KSHV in the extracellular medium. The yield was compared with the cells grown in normoxic conditions for the maximum time period used for hypoxic induction. As expected, the results clearly showed that hypoxia induced viral reactivation (S4C Fig).
Oncogenic herpesviridae signatures are frequently found in blood and tissue samples of the world’s population with high representations in cancer patients [51, 52]. Pathogenesis due to herpesviridae infection is also a consequence of multi-factorial events, which depends on immune status, as well as genetic heterogeneity of infected individuals [12, 53, 54]. KSHV, a large double stranded DNA containing virus was identified in the late 20th century and its infection correlated strongly with the incidences of KS, PEL, MCD and KSHV inflammatory cytokine syndrome (KICS)[3, 55, 56]. During latent infection, epigenetic modification of the KSHV genome allows expression of only a limited number of KSHV encoded genes such as LANA, vFLIP, vCyclin and certain viral interferon regulatory factors [7, 22, 57].
LANA is a necessary factor for tethering of KSHV episomes to the host genome and also functions as a master regulator of latency [14, 58]. LANA can support latent replication through binding to the terminal repeats, as well as supporting hypoxia-mediated lytic replication by cooperating with HIF1α to up-regulate expression of the RTA[27]. Several viral gene products, including the KSHV LANA, vGPCR and vIRF-3 proteins, have been shown to influence HIF1α to function directly through protein-protein interaction or indirectly by enhancing transcriptional or post-transcriptional events [59–61]. LANA augmented HIF-1α stabilization by degrading VHL in the EC5S ubiquitin complex [9], and HIF-1α protein levels is higher in KSHV-positive PEL lines when compared to KSHV-negative cells. Additionally, expression studies showed that HIF-1α is enhanced by LANA, and that LANA stimulated the nuclear accumulation of HIF-1α[62, 63].
Notably, expression of vGPCR into NIH 3T3 mouse fibroblast cells resulted in activation of MEK and p38 signaling cascades, leading to direct phosphorylation of HIF-1α, and thus subsequent increases in HIF-1α transcriptional activity [33]. Furthermore, another latent nuclear antigen-2 (LANA-2) gene of KSHV, more commonly known as viral interferon regulatory factor 3 (vIRF-3), was implicated in the stabilization of HIF-1α, in addition to its oncogenic role of p53 inhibition [64]. Under normoxic conditions, vIRF-3 binds to the bHLH domain of HIF-1α and inhibits the breakdown of HIF-1α, which does not have an impact on its dimerization capability, but further enhanced the nuclear localization and transcriptional activity of HIF-1α [64]. The well characterized KSHV genetic loci influenced by hypoxia include open reading frames for LANA, RTA and vGPCR [26–28]. KSHV-encoded vGPCR is a constitutively active homolog of cellular GPCR and a bona-fide oncogene[65]. It can activate expression of HIF1α and in turn acts on MAP kinase pathways to enhance tumor development and angiogenesis [33]. The KSHV-encoded LANA can interact directly with HIF1α to activate expression of KSHV-encoded reactivation and transcriptional activator RTA [27]. We have recently observed that KSHV-encoded vGPCR is itself under the control of HIF1α, which activates its expression through its action at HREs within the vGPCR promoter to upregulate its expression [26]. Critically, hypoxia-dependent activation of vGPCR can potentially reprogram the metabolism of infected cells globally through generation of reactive oxygen species as well as targeting expression of DNA methyl transferases. In fact, hypoxia can work globally on the KSHV genome to modulate transcription of KSHV-encoded genes [26].
Hypoxia, in general exerts an arrest of cell cycle and DNA replication through activation of HIF1α, ATM, p53 and p21 dependent pathways as well as suppression of Myc dependent transcriptional cascade [41, 42, 66]. Furthermore, stabilized HIF1α binds efficiently with minichromosomal maintenance proteins (MCMs) to keep them inactive and keep the replication machinery in a dormant stage as a direct inhibition of DNA replication during hypoxic conditions [67]. Despite these negative regulations, KSHV infected cells bypass the G1/S transition to enter S-phase and allows productive replication (reactivation) of KSHV through yet undefined mechanisms [27, 32]. In this study, we investigated how KSHV manipulated hypoxia-mediated inhibition of DNA replication to drive replication when grown in this non-permissive and non-favorable condition. As the infection-based experiments using purified KSHV pose restrictions due to low and variable infection rates, as well as epigenetic reprogramming of the KSHV genome after entering the cells which lead to variable expression of KSHV-encoded genes, we compared the differential between BJAB and BJAB-KSHV cells, or HEK293T and HEK293T-BAC16-KSHV cells. Though, these cells are not naturally infected by KSHV, the presence of the complete KSHV genome confers to these cells the characteristics of latently infected cells. Comparative studies between these cells grown under normoxic or hypoxic conditions allowed for observation of the role of KSHV in protection of essential proteins required for G1/S transition (through stabilized Cyclins/CDK2), origin recognition (ORC1-5), Pre-initiation, initiation and elongation associated proteins (for example MCM, CDCs, and DNAPol1A). These results were replicated in PBMCs after infection with purified KSHV further supporting the role of KSHV in protecting cell cycle and DNA replication-associated proteins from hypoxia-mediated degradation. Interestingly, these differences were mainly observed at the level of proteins and not at the levels of transcript. The effect of hypoxia at the levels of transcript was similar in both KSHV negative and positive conditions suggesting that the stabilization was at the post translational stage.
The indispensable role of KSHV-encoded LANA in protecting these proteins from hypoxia-mediated degradation added a new function to the list of activities related to this multi-functional bonafide oncoprotein. The role was further confirmed by monitoring levels of ubiquitination of replication associated proteins in LANA expressing cells grown under hypoxic condition, which was significantly less compared to mock expressing cells grown under similar conditions. Importantly, the role of HIF1α in stabilization of these proteins provides additional clues as to why KSHV induced expression of this protein in infected cells. Combining the knowledge of HIF1α-mediated activation of KSHV-encoded antigens, and feedback regulation of HIF1α-vGPCR-HIF1α for sustained high levels of HIF1α as well as upregulation of LANA/RTA by HIF1α provides a more comprehensive strategy employed by KSHV to maintain continuous replication in hypoxia. It also provides additional information as to the mechanism by which KSHV is reactivated in non-permissive and non-favorable hypoxic conditions (Fig 8). Though, the slight stabilization of CDK2 protein under hypoxic conditions due to expression of RTA or vGPCR remain unclear, it is a matter for further investigation if these antigens do contribute by playing a role in transcriptional activation or stabilization through other strategies.
A number of questions remain unexplored in this study such as the regulatory proteins involved, mainly ubiquitin ligases that are likely targeted by LANA to mediate the stabilization of replication-associated proteins. LANA was shown to form complexes with replication-associated proteins in the replication compartments to regulate their activities by enhancing their stability in hypoxia, which also requires HIF1α activities. Also, the domain of LANA responsible for protecting cell cycle and replication-associated proteins from hypoxia dependent degradation would also be important to identify. Another factor associated with hypoxia, and responsible for G1/S arrest or replication stress is the shortage of energy in the form ATP, where molecular oxygen is an essential component for generation of energy through oxidative phosphorylation[68]. Shortage of ATP in hypoxic conditions pose a direct mechanism for termination of replication elongation. Identifying the mechanism by which these cells are able to manage this energy deficit for sustained replication in hypoxic condition would be another interesting topic to explore. Further, studies are ongoing to identify the different E3-ubiquitin ligases targeted by LANA in hypoxia to provide a more comprehensive picture of the molecular mechanism and to design targeted therapeutic strategies for intervention against KSHV-associated pathologies. Further, as sustained transcription of viral and host genes are also a pre-requisite for proper productive replication and maturation of virus upon reactivation, it would be interesting to investigate the role of KSHV infection on transcriptional stabilization under hypoxic conditions. Additionally, a comparative analysis of epigenetic reprogramming of KSHV genome under hypoxic conditions, which leads to release of repressors of lytic replication would be other area of exploration. These studies further elucidate the mechanism through which modulation of viral and host physiology under hypoxic conditions is regulated by KSHV.
Peripheral blood mononuclear cells (PBMCs) from undefined and healthy donors were obtained from the Human Immunology Core (HIC) of University of Pennsylvania. The Core maintains approved protocols of Institutional Review Board (IRB) in which a Declaration of Helsinki protocols were followed, and each donor/patient gave written, informed consent.
KSHV-negative BJAB cells were obtained from Elliot Kieff (Harvard Medical School, Boston, MA) and BJAB cells stably transfected with KSHV (BJAB-KSHV) were obtained from Michael Lagunoff (University of Washington, Seattle, WA). KSHV-positive body cavity lymphoma-derived BC3 and BCBL1 cells were obtained from the American type culture collection (ATCC) (Manassas, VA). BC3-ShControl and BC3-ShHIF1α cells were generated by lentivirus mediated transduction as described earlier [26]. BJAB, BJAB-KSHV, BC3, BC3-ShControl, BC3-ShHIF1α and BCBL1 cells were maintained in RPMI medium. Human Embryonic Kidney cell line (HEK293T) was obtained from Jon Aster (Brigham and Women’s Hospital, Boston, MA). HEK293T and HEK293-BAC16-KSHV cells were maintained in DMEM medium containing 7% bovine growth serum (BGS) and appropriate antibiotics at 37°C and 5% CO2. BC3-ShControl, BC3-ShHIF1α stable cells were maintained in selection media with puromycin (2μg/ml). HEK293T-BAC16-KSHV cells were also maintained in selection with hygromycin (100μg/ml). ShControl, ShHIF1α, pBS-puroA (KSHV terminal repeats), pBS-puroH (6kb KSHV genomic fragment; co-ordinate 26937–33194), pBS-puro-GA5 (10kb KSHV genomic fragment; co-ordinate 36883–47193) and Supercos1-GB22 (15kb KSHV genomic fragment; co-ordinate 85820–100784), pA3F-LANA and pEF-RTA plasmids were described in earlier publications [25]. Generation and maintenance of BC3ShControl, ShHIF1α and ShLANA cells was described earlier [26]. pLVX-ACGFP-vFLIP, pLVX-ACGFP-vCyclin and pLVX-ACGFP-vGPCR constructs were generated by PCR amplification and ligated into Xho1/Hind III, EcoR1/BamH1 and EcoR1/Apa1 site, respectively. For vFLIP and vCyclin, cDNA from KSHV positive BC3 cells was used as template while for vGPCR, pCEFL-vGPCR construct (a gift from Enrique A. Mesri; University of Miami Miller School of Medicine, Miami, FL) was used as template for PCR amplification. For Hypoxic induction, cells were grown in 1%O2 at 37°C for the indicated time periods. MG132 was procured from Sigma Aldrich (St. Louis, MO) and was used at a final concentration of 5μM.
RNA was isolated by standard phenol chloroform extraction. cDNA was synthesized from 2μg of RNA using superscript cDNA synthesis kit (Applied Biosystem Inc., Foster city, CA) according to manufacturer protocol. The sequence of primers used in the study is provided in the S1 Table. KSHV reactivation and purification was performed according to standard protocol described earlier [26]. To isolate KSHV virion DNA, KSHV virions were resuspended and lysed in 200μl of HMW buffer (10 mM Tris, 150 mM NaCl, 1 mM EDTA, 0.5% SDS and 0.5 mg/ml proteinase K. The virion DNA was further extracted using standard phenol chloroform extraction. Copy number calculation from purified KSHV preparation was performed using standard method. KSHV infection was performed at the multiplicity of infection equivalent to 10 in the presence of 20 μg/ml polybrene as described earlier[26]. The number of extracellular KSHV from the cell culture medium was estimated by real-time PCR through the standard curve method. In brief, equal number of cells in equal volume of cell culture medium were grown in normoxic or hypoxic conditions. Viral reactivation was measured by calculating viral copy number in extracellular culture medium. Viral particles from culture medium were concentrated through centrifugation followed by DNA isolation. DNA pellet were dissolved in an equal volume of water. The standard curve was generated using dilutions of plasmid vector containing KSHV genomic region (15kb KSHV genomic fragment; genome co-ordinates 85820–100784). The sequence of primers used for real-time PCR is given in S1 Table. Unit volume of DNA preparation from extracellular cell culture medium of individual samples were used to estimate the KSHV copy number using primers specific to the cloned KSHV DNA fragment.
Protein lysates were separated on 10% polyacrylamide gel followed by wet transfer to nitrocellulose membrane. Skimmed milk (5%) was used for blocking at room temperature for 1 hour with gentle shaking. Primary antibody against CDK2, Cyclin D1, Cyclin E, PDK1, ORC1, ORC2, ORC3, ORC4, ORC5, ORC6, MCM3, GFP, Ubiquitin and GAPDH (Santa Cruz Inc., Dallas, TX), Myc tag and LANA (purified ascites), DNAPOL1A and Cdt1 (Novus Inc., Centennial, CO), CDC6 and CDC45 (Cell Signaling Technology Inc. Danvers, MA), FLAG (Sigma Aldrich Inc., St. Louis, MO) were incubated overnight at 4°C with gentle shaking followed by washing with TBST. Probing with IR conjugated secondary antibody was performed at room temperature for 1 hour followed by washing with TBST. Membranes were scanned on an Odyssey scanner for detection of signals. A complete list and details of antibodies used are provided in S2 Table.
For confocal microscopy, 25,000 cells were semi-dried on 8-well glass slides followed by fixation in 4% paraformaldehyde. Combined permeabilization and blocking was performed in 1XPBS containing 0.3% Triton X-100 and 5% goat serum followed by washing (5 minutes each) with 1X PBS. Anti-LANA antibody (1:200 dilution) was diluted in 1X PBS containing 1% BSA and 0.3% Triton X-100 and was incubated overnight at 4°C. Slides were washed with 1X PBS followed by incubation with Alexa 448 conjugated anti-mouse secondary antibody. DAPI staining was performed for 15 minutes at room temperature followed by washing and mounting. Images were captured by confocal microscope.
Whole cell lysates were pre-cleared with Protein Agarose A/G beads and incubated overnight with indicated antibodies with gentle shaking. This was followed by antibody incubation, and Protein Agarose A/G beads were used to collect the immune complexes. The beads were washed 3 times with 1X PBS and resuspended in 60 μl 2X SDS loading dye. The immuno-precipitated complexes were run against 5% input sample lysate.
Pulse field gel electrophoresis (PFGE) and southern blot were described earlier [25]. Briefly, cells were pulsed with Chlorouridine (CldU, 10 μM) in cell culture medium for 4 hours and harvested by centrifugation. Cells were further resuspended in fresh medium containing Iodouridine (IdU, 10 μM) and again pulsed for 4 hours. At the end of pulsing, cells were pelleted, washed with 1XPBS and resuspended in 0.5ml 1XPBS. An equal volume of cell suspension and 1% InCert agarose were equilibrated at 45°C before mixing and molding in the cast for the preparation of cell embedded agarose plugs. The agarose plugs were digested with Proteinase K and followed by washing in 1XTE buffer pH8 with buffer changes after each 24 hours. Final washing of plugs were done in 1XTE pH8 containing 1mM PMSF. Pme1 was used to digest and linearize KSHV episomal DNA. Cell embedded plugs were fitted with agarose gel casting tray in 0.7% low melt agarose. DNA was separated by Pulse field gel electrophoresis for 36 hours on BioRad Chef DRII system (Bio-Rad Inc. Hercules, CA). Post PFGE, Pme1 digested DNA in agarose gel was depurinated and the gel was rinsed with double distilled water followed by denaturation. The gel was then rinsed with double distilled water followed by neutralization and further rinsed with distilled water and equilibrated in SSPE before alkaline transfer on Nylon membrane. Transferred DNA was subjected to UV cross-linking at 1400 Joules and prehybridization was performed at 42°C in prehybridization buffer. Hybridization probes were prepared by random priming method and hybridization was performed at 49°C in prehybridization buffer devoid of salmon sperm DNA. Membrane was washed with low stringency buffer followed by high stringency buffer. Membranes were wrapped in Saran Wrap and exposed to sensitive plates followed by imaging using a Phosphorimager.
Cells were harvested by centrifugation at 1000 rpm for 5 minutes and resuspended in 300μl 1X PBS. 700μL ice cold absolute ethanol (70% final) was added drop wise to the cell while gentle vortexing to avoid clumping of cells. Cells were fixed at 4°C on a rotating shaker for 30 minutes followed by washing in PBS. Cells were resuspended in 200μl 1X PBS and incubated with RNase A for 1 hour at 37°C. Following this, 250μl 1X PBS and 50μl Propidium Iodide (1mg/mL) was added to the cells, mixed and stained for 30 minutes at room temperature. The cells were washed with 1XPBS and resuspended then analyzed on FACS Calibur (Becton Dickinson Inc., San Jose, CA, USA). The acquired data were analyzed using FlowJo software (TreeStar Inc., San Carlos, CA, USA).
|
10.1371/journal.ppat.1004005 | A Highly Conserved Toxo1 Haplotype Directs Resistance to Toxoplasmosis and Its Associated Caspase-1 Dependent Killing of Parasite and Host Macrophage | Natural immunity or resistance to pathogens most often relies on the genetic make-up of the host. In a LEW rat model of refractoriness to toxoplasmosis, we previously identified on chromosome 10 the Toxo1 locus that directs toxoplasmosis outcome and controls parasite spreading by a macrophage-dependent mechanism. Now, we narrowed down Toxo1 to a 891 kb interval containing 29 genes syntenic to human 17p13 region. Strikingly, Toxo1 is included in a haplotype block strictly conserved among all refractory rat strains. The sequencing of Toxo1 in nine rat strains (5 refractory and 4 susceptible) revealed resistant-restricted conserved polymorphisms displaying a distribution gradient that peaks at the bottom border of Toxo1, and highlighting the NOD-like receptor, Nlrp1a, as a major candidate. The Nlrp1 inflammasome is known to trigger, upon pathogen intracellular sensing, pyroptosis programmed-cell death involving caspase-1 activation and cleavage of IL-1β. Functional studies demonstrated that the Toxo1-dependent refractoriness in vivo correlated with both the ability of macrophages to restrict T. gondii growth and a T. gondii-induced death of intracellular parasites and its host macrophages. The parasite-induced cell death of infected macrophages bearing the LEW-Toxo1 alleles was found to exhibit pyroptosis-like features with ROS production, the activation of caspase-1 and IL1-β secretion. The pharmacological inactivation of caspase-1 using YVAD and Z-VAD inhibitors prevented the death of both intravacuolar parasites and host non-permissive macrophages but failed to restore parasite proliferation. These findings demonstrated that the Toxo1-dependent response of rat macrophages to T. gondii infection may trigger two pathways leading to the control of parasite proliferation and the death of parasites and host macrophages. The NOD-like receptor NLRP1a/Caspase-1 pathway is the best candidate to mediate the parasite-induced cell death. These data represent new insights towards the identification of a major pathway of innate resistance to toxoplasmosis and the prediction of individual resistance.
| Toxoplasmosis is a ubiquitous parasitic infection causing a wide spectrum of diseases. It is usually asymptomatic but can lead to severe ocular and neurological disorders. The host factors that determine natural resistance to toxoplasmosis are yet poorly characterized. Among the animal models to study susceptibility to toxoplasmosis, rats develop like humans a subclinical chronic infection. The finding of a total resistance in the LEW rat strain has allowed genetic studies leading to the identification of Toxo1, a unique locus that controls the outcome of toxoplasmosis. In this report, a panel of recombinant inbred rat strains was used to genetically reduce the Toxo1 locus, on chromosome 10, to a limited region containing 29 genes. This locus is highly conserved among five resistant, by comparison to four susceptible, rat strains, indicating that refractoriness to toxoplasmosis could be predicted. The Toxo1-controlled refractoriness depends on the ability of macrophages to restrict parasite proliferation and the rapid death of both T. gondii and host macrophages in vitro. The NOD-like receptor NLRP1a/Caspase-1 pathway is the best candidate to mediate the parasite-induced cell death. Our data represent new insights towards the identification of a major pathway of innate immunity that protects from toxoplasmosis.
| Toxoplasma gondii is a widespread obligate intracellular protozoan parasite. One preeminent aspect of its life cycle is the establishment of a chronic infection in humans and many other vertebrate hosts [1]. Toxoplasmosis is most often asymptomatic depending on the parasite's ability to elicit host protective immunity [1]. A serious threat to human health can occur under congenital infection or reactivation of a latent infection in immunodeficient patients [2]. Epidemiological studies have indicated that the phenotypic expression of toxoplasmosis depends on the genetic make-up of both the host and the parasite [3], [4]. Variations in the outcome of Toxoplasma infection after exposure to similar risk factors [5], [6] and twin studies [7] support a significant role of the human host genetic background in the susceptibility to toxoplasmosis. Nevertheless, genetic studies in human are hampered by both population heterogeneity and environment variability. In experimental conditions, genetic and environmental factors are under control. Rats, like humans, usually develop subclinical toxoplasmosis. This contrasts with the severity of the disease developed in most strains of mice. Interestingly, an unexpected refractoriness to T. gondii infection was found in the LEW rat strain [8]. Compared to susceptible BN rats, infected LEW indeed displayed negative serology and lack of cyst burden in their brain [9]. Refractoriness of LEW rats was found to be a dominant trait dependent on hematopoietic cells [9]. It is associated with the ability of macrophages to restrict parasite proliferation in vitro [10].
Further genetic studies using LEW resistant and BN susceptible rats and derived reciprocal congenic strains have allowed the mapping of a locus named Toxo1, which fully controls the refractoriness of LEW rats to toxoplasmosis. Toxo1 has been confined to 7.6 megabases, on rat chromosome 10 (Rn10q.24) [10]. Recently, hNlrp1 a major candidate gene present in the orthologous region to Toxo1 in the human genome (Hs 17p32.2-p13.1) has been associated with human congenital toxoplasmosis [6].
In the present work, we used genetic dissection with a panel of BN and LEW sub-congenic rats and haplotype analysis of chromosome 10 on nine inbred rat strains either susceptible or resistant to define the localization of the gene or set of genes at work in Toxo1 and to analyze the mechanisms of toxoplasmosis refractoriness. We were able to localize the Toxo1 locus in a 891 kb region highly conserved in all resistant strains of rat. Sequencing of this locus in these nine strains revealed a high concentration of resistant-restricted conserved mutations at the bottom border of Toxo1 around Nlrp1. Functional studies in ex vivo infected peritoneal macrophages indicate that the Toxo1-mediated restriction of parasite proliferation is associated with the coordinate death of both parasites and host macrophages. The parasite-induced macrophage cell death involved a caspase-1 dependent mechanism and exhibited pyroptosis-like features. The parasite-induced killing of infected macrophages could be blocked by the capase-1 inhibitor without restoration of parasite proliferation. Therefore, we concluded that if at work, the NLRP1a/caspase-1 pathway is not indispensable to restrict parasite proliferation in macrophages.
We previously demonstrated that the Toxo1 7.6 Mb interval fully controls the outcome of T. gondii infection independently of the genetic background. The refractoriness to infection conferred by the LEW origin of Toxo1 is characterized by the early elimination of the pathogen resulting in a barely detectable specific immune response and in the absence of brain cysts [10]. In vitro, this Toxo1-LEW mediated refractoriness is associated with the control of parasite proliferation within macrophages [10]. To refine the localisation of the gene(s) that control(s) these in vivo and in vitro phenotypes, we generated a unique panel of congenic sub-lines. Results from the genetic dissection are shown on Figure 1. The parasites were found able to proliferate within the macrophages from the congenic BN.LEWc10-Ce, -Cf, -Cga, -Ci and LEW.BNc10-F sub-lines but not within the macrophages from the congenic BN.LEWc10-Cg and -Ch sub-lines (Figure 1A). Thus within the 7.6 Mb of the Toxo1 locus a 891 kb region controls the in vitro proliferation of parasites within macrophages. We further investigated refractoriness or susceptibility to T. gondii infection in vivo in rats from the seven congenic sub-lines used for these in vitro studies as well as in rats from the BN and LEW parental strains. The control of refractoriness to T. gondii defined by both the absence or low specific antibody response (Figure 1B) and the absence of cyst burden in the brain (Figure 1C), was directed by the same 891 kb region (Figure 1D). Thus, the interval located between the D10GF49 (57.26 Mb) and D10GF55 (58.15 Mb) microsatellite markers contains the gene or the set of genes that controls the toxoplasmosis outcome.
We hypothesized that genetic variation(s) underlying Toxo1-mediated innate refractoriness against T. gondii infection result(s) from an ancestral polymorphism instead of independent newly acquired mutations in the LEW strain and thus could be identified in various inbred rat strains. To challenge this hypothesis, we investigated toxoplasmosis outcome in seven other inbred rat strains (LOU, WF, WK, BDIX, OM, F344 and DA), in comparison to BN and LEW. Following infection, rats exhibited a dichotomous phenotype either developing high titers of anti-toxoplasma antibodies and cerebral cysts or remaining refractory to parasite infection (Figure 2). Three strains (OM, DA and F344) displayed, like the BN rat, phenotypes associated to chronic infection with high anti-T. gondii antibody responses (≥5000 u.a) and the detection of brain cysts. Of note, the number of brain cysts was significantly different among these rat strains (Figure 2B) reflecting other regulatory mechanism(s) for cyst formation. By contrast, the four other rat strains (LOU, BDIX, WK and WF) showed the refractory phenotype to T. gondii infection of the LEW strain neither developing specific antibodies nor cerebral cysts (Figure 2). As expected, parasite proliferation within peritoneal macrophages was observed only in the BN and the three other susceptible strains but neither in LEW nor in the four other refractory strains (Figure 2C). Thus, refractoriness against T. gondii infection is a common trait of several inbred rat strains.
To investigate the implication of the Toxo1 locus in resistance against T. gondii infection, we used a targeted method of genotyping to stratify (LOU x BN) F2 rat according to their genotype at Toxo1. Thirty five F2 rats were genotyped at the polymorphic marker D10GF41, located at the peak of the Toxo1 locus, and their anti-T. gondii Ab titers and brain cyst numbers were quantified following infection with T. gondii (Figure 3). All rats sharing the two LOU alleles (ll) showed no detectable or a weak anti-T. gondii Ab response (<10,000 a.u.). Conversely, all rats sharing the two BN alleles (nn) had high anti-T. gondii Ab titers (>20,000 a.u.). The apparent intermediate antibody response observed in the 15 heterozygous rats (nl) was not significantly different from that observed in the panel of rats sharing the LOU alleles (ll). In a similar way, brain cysts were observed in all the ten homozygous BN (nn) rats and in none of both the heterozygous BN/LOU (nl) and the nine homozygous LOU (ll). These results showed that Toxo1 directs in a dominant manner the toxoplasmosis outcome in LOU rats after T. gondii infection.
As Toxo1 controls resistance against T. gondii infection in the LEW and LOU rats, and given the similarities between the response of LEW, LOU, BDIX, WK and WF, we next examined if Toxo1- allelic similarities are conserved among all those resistant animals. Considering that as a result of common ancestry, patterns of allelic similarities and differences among strains can be discerned for every variable locus [11], we conducted a haplotype study on the nine strains previously described. This analysis, based on allele size data for microsatellite markers, consisted in identifying among the different strains the chromosomal regions with LEW genotype. For this purpose, 41 polymorphic microsatellite markers on chromosome 10 were investigated on these nine strains (Table S1). The genotype of the nine strains for these markers showed the conservation of an haplotype block in the five resistant strains (LEW, LOU, WF, WK and BDIX) between D10Arb7 and D1Rat297 as compared to the four susceptible strains (BN, OM, F344 and DA) (Figure 4). This highly conserved haplotype block extends on 2.8 Mb between D10Arb7 and D1Rat297 and overlaps the entire 891 kb Toxo1 locus (Figure 4). Thus, the data clearly suggested that conserved genetic variations within the Toxo1 locus on chromosome 10 is a common cause of resistance against T. gondii infection in inbred rat strains.
According to the genome database (www.ensembl.org, RGSC3.4 version), the Toxo1-891 kb interval contains 29 genes. None of these 29 genes could be retained as a candidate on the basis of a significant difference in their level of expression between macrophages from resistant vs. macrophages from susceptible congenic lines (Table S2, TextS1). Therefore, the entire Toxo1 locus of the nine rat strains studied in the haplotype analysis was sequenced to identify resistance-correlated variations in coding- and non-coding sequences. A total of 373 SNPs and 21 insertions/deletions were found strictly conserved among the five resistant strains as compared to susceptible strains. The distribution of these mutations along the locus displays a gradient with a densification at the bottom of Toxo1 (Figure 5). We identified 23 SNPs of which 16 are missense and one is a deletion in the coding sequences leading to the selection of four candidate genes: Inca1 (1 SNP), Kif1C (1 ins/del), Nlrp1a (13 SNPs) and Nlrp1b (2 SNPs). Given that Inca1 and Nlrp1b mRNAs are undetectable in peritoneal macrophages (Table S2) and according to the number of mutations in Nlrp1a vs Kif1C coding sequences, Nlrp1a appeared as the major candidate gene. It encodes the NOD-like receptor (NLR) NLRP1 that acts as an intracellular pattern recognition receptor (PRR) [12]. Both the mouse Nlrp1b and its ortholog rat Nlrp1a have been described as implicated in the control of a cell death process called pyroptosis that is induced by the lethal toxin (LT) from Bacillus anthracis [13].
Based on NLRP1 known function, we investigated the impact of the Toxo1 locus on parasite and host cell fate after ex vivo infection of peritoneal macrophages. Following host-cell invasion, T. gondii replicates within a newly formed non-fusogenic compartment, the parasitophorous vacuole (PV). Using fluorescent parasites and vacuole staining with anti-GRA5 or -GRA3 antibodies which both stain the PV membrane, we compared the fate of intravacuolar parasites by immunofluorescence microscopy after allowing them to invade either non-permissive LEW or permissive BN naive peritoneal macrophages (Figure 6). The analyses were performed at 2 and 8 hours following invasion using transgenic-YFP2 fluorescent parasites. According to our previous work [10] we found similar rates of parasite invasion in both permissive (Toxo1-BN) (15%±5%) and non-permissive (Toxo1-LEW) (17%±6%) naive peritoneal macrophages. At 2 hours post-infection, most of intracellular YFP2-parasites were found within a compartment positive for the GRA5 PV membrane marker, in both LEW and BN macrophages (Figure 6A) indicating that parasites are able to enter efficiently into resistant macrophages. However, we observed a slight decrease in the percentage of YFP- and GRA5- positive vacuoles within LEW (77%±4) as compared to BN (89%±6) macrophages (Figure 6B). At 8 hours post-infection, while YFP2-parasites started to divide within GRA5 positive vacuoles of permissive BN macrophages (Figure 6A), a dramatic drop of YFP staining was observed in GRA5-positive vacuoles of LEW (21%±4) as compared to BN (66%±6) macrophages (Figure 6A and B). The loss of YFP emission in GRA5-positive vacuoles was likely to reflect the death of parasites [14] although parasite egress could not be excluded. Interestingly, the difference was not so marked when the parasite surface was stained using anti-SAG1 antibodies. Indeed, in BN macrophages, the percentage of vacuoles containing SAG1-positive parasites (62.5%±1) was not statistically different from the percentage of vacuoles containing YFP2-parasites (66%±6). In contrast, in LEW macrophages the percentage of vacuoles containing SAG1-positive parasites (43%±3) was two-fold the percentage of vacuoles containing YFP2-parasites (21%±4), indicating that at least 2/3 of vacuoles in LEW macrophages still contained parasites (Figure 6C and D). Therefore, the decrease of YFP staining could be attributed, at least in majority, to parasite death resulting from the macrophage microbicidal activity, rather than to parasite egress. The parasite death within the PV was further examined using staining of small ubiquitin-related modifier (SUMO) as a read-out of the transcriptional activity of live parasites [15]. As shown in Figure 6E and 6F, 2 and 8 hours after infection, a dramatic drop in the percentage of SUMO positive vacuoles was observed in LEW macrophages while no difference was found in BN macrophages, thus supporting the early death of parasites within LEW macrophages. We finally examined whether this phenotype was under the Toxo1 control with our collection of sub-congenic rat strains. For this purpose, peritoneal macrophages from the four susceptible (BN.LEWc10-Ce, -Cf, -Cga, -Ci and LEW.BNc10-F) and the two refractory (BN.LEWc10-Cg and -Ch) sub-congenic lines were infected with T. gondii. In macrophages from the LEW and BN.LEWc10-Cg, -Ch lines in which Toxo1 is from LEW origin, the inhibition of parasite proliferation correlates with the induction of intracellular parasite death (LEW: 58±13%; BN.LEWc10-Cg: 69±2%; -Ch: 64±7%). Conversely, in macrophages from the BN.LEWc10-Ce, -Cf, -Cga and -Ci lines in which Toxo1 is from BN origin, the parasite proliferation was associated with a decrease of parasite death (BN.LEWc10-Ce: 17±11%; -Cf: 0±21%; -Cga: 3±7%; -Ci: 1±17%; LEW.BNc10-F: 0±16%) (Figure 6G). Altogether these results demonstrated that the Toxo1 locus from LEW origin mediates the T. gondii killing by infected peritoneal macrophages.
We next examined the fate of infected host cells in a comparative way depending on the Toxo1 genotype. At 8 hours post-infection, we observed that most of LEW macrophages infected with YFP-negative vacuoles presented a condensed nucleus (Figures 6A, 6C, 6E). Indeed, when this phenomenon was quantified we found that 65% of LEW macrophages with YFP-negative vacuoles, but only 14% of BN susceptible macrophages, showed a condensed nucleus (Figure 7A). This observation indicating a parasite-induced killing of LEW macrophages was further investigated using propidium iodide (PI) uptake. At 6 hours post-infection, about 40% of LEW macrophages (39%±12) had lost membrane integrity as compared to less than 10% of BN macrophages (6%±8) (Figure 7B, p<0.05). The lack of PI uptake under incubation with lysed fibroblasts ruled out the possible triggering by unrelated pathogen-associated molecular pattern (PAMP) (Figure 7B). Tracking of both PI uptake and YFP loss in kinetic experiments indicated that the parasite-induced cell death of LEW macrophages is concomitant with the death of intracellular parasites. Indeed, the increase of PI uptake by macrophages and the loss of YFP by intracellular parasites started from 1 hour and increased steadily until 6 hours after infection (Figure 7C). Further quantification of PI and GRA5 PV marker positive LEW macrophages demonstrated that 88% of dying cells were parasite-invaded cells (Figure 7D). Finally, the implication of Toxo1 in the host cell death phenotype was validated using our panel of sub-congenic animals. Specific induction of host cell death after infection of non-permissive peritoneal macrophages was found in all resistant congenic rats (LEW: 30±10%; BN.LEWc10-Cg: 35±3%; -Ch: 24±3%). By contrast, peritoneal macrophages from all animals permissive to T. gondii proliferation failed to initiate such a cell death process after infection (BN.LEWc10-Ce: 11±1%; -Cf: 8±3%; -Cga: 10±4%; -Ci: 13±7% and LEW.BNc10-F: 17±5%) (Figure 7E). Altogether, these data demonstrated that T. gondii invasion of resistant macrophages is rapidly followed by the combined deaths of intracellular parasites and infected host macrophages using a mechanism under the control of Toxo1 locus.
The rapid loss of membrane integrity of parasite-invaded LEW macrophages (Figure 7C) suggested that the T. gondii-induced cell death was not apoptotic. Accordingly, DNA from dying cells did not show the laddering resulting from chromatin fragmentation, observed in classical apoptotic cell death and pyroptosis (Figure 8A). Moreover, caspase-3 activation, which plays a central role in the executive phase of apoptosis, was not observed in infected LEW macrophages (Figure 8B). Additionally, when infected macrophages were incubated with FITC-labelled annexin V, phosphatidylserine exposure on the plasma membrane was not observed prior to the loss of membrane integrity as assessed by PI staining (Figure 8C). Finally, no difference in the number of acidic vacuoles was observed after lysotracker coloration between permissive BN and non-permissive infected LEW macrophages indicating that the death of infected LEW macrophages was not due to autophagy (Figure 8D).
Given the described role of NLRP1/Caspase-1 inflammasome pathway in the host response to pathogens [16], [17], we hypothesized that the T. gondii-induced cell death of infected resistant LEW macrophages could be associated with both ROS production and caspase-1 activation. The production of intracellular ROS by macrophages was monitored at two time points (15 min and 4 hours) by the dihydro-rhodamine 123. While T. gondii infection did not induce significant ROS production within permissive LEW.BNc10-F macrophages, a marked increase was recorded in infected resistant LEW macrophages (Figure 9A). We next examined caspase-1 activation within infected macrophages by using the fluorogenic activated caspase-1 specific staining (FLICA). At 4 hours post-infection, the percentage of parasite-induced FLICA positive cells was significantly higher in resistant LEW macrophages (29%±2) than in permissive LEW.BNc10-F macrophages (11%±1) (Figure 9B and C). The caspase-1 induction in infected LEW macrophages correlated with the increase of PI-positive cells indicating that caspase-1 was involved in the cell death induction process (Figure 9D). Consistent with these results, the processing of caspase-1 substrate IL-1β that could be prevented by the YVAD caspase-1 inhibitor was detected at 1 h and more evidently at 4 h post-infection in the culture supernatant of infected LEW macrophages and not in that of permissive LEW.BNc10-F macrophages (Figure 9E). The observed difference in the secretion of mature IL-1β was not due to a lack of pro-IL-1β expression since it was found to be induced in response to T. gondii infection within both resistant LEW and permissive LEW.BNc10-F macrophages (Figure 9E). By contrast, pro-IL-1β was not detectable in the cell lysate from LEW macrophages treated with YVAD (Figure 9E) indicating that the inhibition of caspase-1 activity resulted in the down-regulation of pro-IL-1β protein expression. Altogether, these results demonstrated that both ROS production and caspase-1/IL1β pathway are involved in the Toxo1-mediated cell death induction of resistant LEW macrophages.
The role of caspase-1 in the resistance of LEW macrophages was further investigated using pharmacological inhibitors. Both caspase-1 and pan-caspase inhibitors were able to protect the resistant LEW macrophages from the parasite-induced cell death (Figure 10A). Consistent with our above observations that both parasites and host macrophages killing processes are connected, the two inhibitors also prevented the death of intracellular parasites (Figure 10B). By contrast, these inhibitors failed to restore significant parasite proliferation in non-permissive macrophages (Figure 10C). Altogether these experiments revealed that the resistance of macrophages bearing Toxo1-LEW alleles relies on two pathways, the first one controlling parasite proliferation and the second one controlling the death of both intracellular parasites and host macrophages, via caspase-1 dependent inflammasome activation.
Forward genetics has proved to be a powerful tool to characterize novel biological pathways implicated in host resistance to infection [18], [19]. In rats, the Toxo1 locus located on chromosome 10 [10] controls the outcome of toxoplasmosis by a still poorly defined mechanism. In the present work, genetic dissection of Toxo1 with a panel of new congenic sub-lines together with haplotype mapping led us to identify a 891 kb interval of rat chromosome 10 that is highly conserved amongst resistant rat strains and that controls both T. gondii infection outcome in vivo and macrophage responses to infection in vitro. We further demonstrated that the Toxo1-mediated refractoriness of macrophages to T. gondii infection is associated with a caspase-1-dependent rapid T. gondii-induced death of both intracellular parasites and host macrophages.
The Toxo1-dependent death of infected macrophages displayed the chromatin condensation hallmark of apoptosis but neither caspase-3 activation nor DNA fragmentation were observed. In contrast to apoptosis which is usually a slow process characterized by membrane blebbing, the Toxo1-associated cell death was characterized by a rapid loss of plasma membrane integrity that occurs simultaneously with surface exposure of phosphatidylserine.
A non-apoptotic pathway triggered in T. gondii-infected macrophages has been described in mice [20]. It is mediated by IFN-γ-inducible immunity-related GTPases that trigger vacuole membrane disruption and the death of parasites, followed by the necrotic-like death of infected cells. While phenotypically, the IRG-dependent mouse T. gondii/macrophage deaths parallel the Toxo1-mediated T. gondii/macrophage deaths, major differences between the two in vitro models exist, providing strong evidences that mechanistically they are not identical. The mouse IRG-controlled mechanism requires IFN-γ stimulation and occurs in both fibroblasts and macrophages. In contrast, the Toxo1 effect in rats is strictly confined to hematopoeitic cells [9], [10] and does not depend on IFN-γ stimulation of macrophages in vitro. Moreover, the mouse IRG-controlled resistant system is genotypically restricted to non-virulent genotype II parasites, while in rats, both type I RH and type II PRU strains triggered the LEW T. gondii/macrophage deaths (Figure S1). In line with this, the effector of the LEW rat resistance is not the kinase parasite effector ROP 18 (data not shown) which, by phosphorylating IRGs, disrupts their association with the parasite-containing vacuole and thereby protects the parasite against elimination [21], [22]. Altogether, the rat Toxo1 locus-mediated resistance to toxoplasmosis does not operate via the IRG-dependent system.
Sequencing of the Toxo1 region in all five resistant rat strains (LEW, LOU, BDIX, WK and WF) and four susceptible strains (BN, OM, DA and F344) revealed a gradient in the conserved distribution of resistant-restricted mutations that peaks at the bottom of Toxo1 and particularly in the Nlrp1a coding sequence (23 SNPs). This highly conserved region has been previously demonstrated to be associated with the resistance of rats to the lethal Toxin (LT) of Bacillus anthracis [23]. Interestingly, while rat bearing the divergent Toxo1 alleles were susceptible to LT-mediated death, the highly conserved Toxo1-LEW alleles correlated with the total resistance of rats [23]. Moreover, similarly to what we found for the Toxo1 control of T. gondii infectivity, there was a perfect correlation between the in vivo phenotype (sensitivity vs resistance to LT) and the in vitro phenotype of macrophages, suggesting that the same gene or set of genes might be at work in the control of these two pathogens. In mice and rats, further genetic mapping associated genetic variants of both mNlrp1b and one of its rat ortholog rNlrp1a to macrophage sensitivity or resistance to LT by a mechanism dependent on a cell death process called pyroptosis, which is also triggered upon infection with intracellular pathogens such as Salmonella, Shigella or Listeria [24]–[26]. NLRP1 is part of the NLR family, which is known to act as an intracellular sensor for cytoplasmic danger signals [12]. After activation, the NLR form a multimeric protein complex called the inflammasome that provides a scaffold for the activation of caspase-1 and target death substrates by a still poorly understood mechanism [12]. The Toxo1 controlled-cell death induction that is triggered in peritoneal macrophages following T. gondii invasion, appeared to be also caspase-1 dependent and associated to IL-1β secretion and ROS production. Together, with the nuclear condensation, it thus features the hallmarks of the pyroptosis programmed-cell death which is uniquely dependent on caspase-1 and inherently proinflammatory. However, while pyroptosis is typically associated to DNA fragmentation [16], this later event was not observed in the T. gondii-induced cell death possibly due to a yet unexplained undetectable PARP (Poly ADP-Ribose Polymerase) expression in rat peritoneal macrophages (Figure S3, Text S1). Together these observations combined with the genetics studies tend to support that the NOD-like receptor NLRP1a/Caspase-1 pathway is the best candidate to mediate the Toxo1-dependent parasite-induced cell death.
Despite these remarkable genetic and biochemical similarities, LT- and T. gondii-induced phenotypes display striking major differences. First, while Toxo1-conserved LEW alleles are associated to the cell death induction of macrophages upon T. gondii infection, same allelic variants protect the macrophages from LT-mediated cell death [23]. Secondly, while the pharmacological inhibition of host cell death also prevents the concomitant death of parasites, it failed to restore the permissiveness of macrophages to parasite proliferation. Thus, although caspase-1 is induced following T. gondii infection, our data do not argue for its essential role in the macrophage control of parasite proliferation in vitro. They rather suggest that, like it is emerging in the case of intracellular bacteria, the Toxo1-directed resistance of macrophages to T. gondii proliferation may rather be a complex trait resulting from the combined activation of several pathways. For instance, the Naip5/Nlrc4-controlled restriction of Legionella pneumophila growth within non permissive murine macrophages is likely to involve both canonical pyroptosis and a yet undefined caspase-1 independent Naip5 pathway [27]. Very interestingly, in mouse macrophages infected by Shigella flexneri which display NLRC4- and NLRP3-dependent activation of caspase-1, IL-1β/IL-18 processing and cell death [17], [26], it has been demonstrated that inflammasome might negatively regulate pathogen-induced autophagy [26]. Given that other NLRs may affect autophagy [17], inflammasome inhibition of autophagy might be a generalized mechanism. Moreover, Harris J. and al. demonstrated that autophagy controls IL1-β secretion by targeting pro-IL1-β for degradation [28]. In line with this, the lack of detectable pro-IL1-β in resistant LEW macrophages treated with caspase-1 inhibitor, suggested that inactivation of caspase-1 results in the down-regulation of pro-IL1-β protein expression. It is therefore possible that in our model, the inhibition of caspase-1 could block the negative regulation of inflammasome promoting thus the effect of other pathway(s) capable to restrict parasite proliferation without pyroptosis. Altogether, our work provide evidence that natural resistance of rat macrophages to T. gondii is a complex trait relying on a mechanism involving the combined activation of at least two pathways: (i) the classical NLRP1a/Caspase-1 pathway, mediating host and parasite cell death and (ii) a yet undefined pathway that would directly control parasite proliferation within the parasitophorous vacuole. The parasite effector(s) eliciting these pathways and their possible interconnections remain to be investigated.
Toxoplasma infection is naturally acquired by the oral route. Following transcytosis across the intestinal barrier [29], the tachyzoite stage encounters leukocytes in which it replicates to further disseminate into the organism using the migratory properties of infected macrophages and dendritic cells [30]. In rats bearing the LEW-type Toxo1 locus, the refractoriness to toxoplasmosis is evidenced by the absence of both local parasite burden and specific antibody response [9]. Thus resistance likely results from the rapid clearance of the parasite following a vigorous killing response at the site of infection. Our data led us to propose a model where the early death of infected macrophages impairs further dissemination of the parasite, hence constituting an efficient barrier against successful infection. How the Toxo1 locus controls dendritic cell and monocyte responses after infection remains to be characterized.
In conclusion, this work highlighted several novel aspects of the host-parasite gene interaction. We unambiguously mapped the Toxo1 locus to an 891 kb region which directs the outcome of toxoplasmosis in the rat. This locus controls parasite infectivity in vivo and is critically associated to macrophage-dependent restriction of parasite intracellular growth and cell death induction of both intracellular parasites and infected macrophages in vitro. It is included within a haplotype block remarkably subjected to strong selection pressure for resistant strains, in contrast to susceptible strains (Figure S2, Text S1). The robustness of Toxo1-mediated resistance in controlling parasite infectivity indicates that this resistance might have been conserved among naturally resistant species [11]. Hence, Transmission Disequilibrium Test studies revealed that hNlrp1, the human ortholog of rat rNlrp1a, has alleles associated with the susceptibility to human congenital toxopolasmosis [6]. In the same way, recent work demonstrated that, in mice, NLRP1 is an innate immune sensor for Toxoplasma infection inducing a host-protective innate immune response to the parasite [31]. Altogether, these data identified a genetically-controlled major pathway of innate immunity to toxoplasmosis allowing predicting resistance of individuals. The results open the way to further investigations towards the gene(s) and the mechanisms at work, and could be applied to human toxoplasmosis in regards to the conserved synteny of Toxo1 region between rat and human.
Breeding and experimental procedures were carried out in accordance with national and international laws for laboratory animal welfare and experimentation (EEC Council Directive 2010/63/EU, September 2010). Experiments were performed under the supervision of M–F. C–D. (agreement 38 10 38) in the Plateforme de Haute Technologie Animale (PHTA) animal care facility (agreement n° A 38 516 10006 delivered by the Direction Départementale de la Protection des Populations) and were approved by the ethics committee of the PHTA (permits n° Toxo-PC-1 and n° Toxo-PC-2).
Production and genotype analyses of congenic lines were as described previously [10]. The congenic lineages were maintained by regular brother-sister mating. LEW/OrlRj (LEW), BN/OrlRj (BN) male rats were obtained from Janvier Laboratory (Le Genest-Saint-Isle, France). WF/N (WF), WKY/NHsd (WK), BDIX/Han (BDIX), F344/Nhsd (F344), Lou/CNimrOlaHsd (LOU) and DA/OlaHsd (DA) male rats were obtained from Harlan Laboratory (Gannat, France). OM/Han (OM) rats were kindly supplied by the Hannover Medical School (Germany). F2 (LOU×BN) progenies were produced in our animal facilities under specific pathogen-free conditions. Cysts from the recombinant T. gondii Prugniaud strain were used to test in vivo the susceptibility to toxoplasma infection. Two-month-old Swiss mice (Janvier laboratory) were infected orally with 10 Prugniaud cysts. Their brains were collected 3 months later and ground in a Potter. Cysts were counted in a Thoma's cell and diluted in PBS. Rats were infected orally with 20 cysts. One month later, blood was collected from the retro-orbital sinus for detection of anti-toxoplasma Ab response by ELISA. Rats were euthanized 2 months after infection and brains were collected to determine number of cysts.
Tachyzoites of Prugniaud type II strain, and RH, RH-YFP2 (kindly provided by B. Striepen, Athens) and RH-mcherry (kindly provided by A. Bougdour, Grenoble) type I T. gondii strains were maintained under standard procedures, by serial passage onto human foreskin fibroblast monolayers (HFFs) in D10 medium (DMEM supplemented with 10% heat-inactivated fetal bovine serum, 1 mM glutamine, 500 units.ml−1 penicillin and 50 µg.ml−1 streptomycin) at 37°C in a humidified atmosphere containing 5% CO2. The parasites were collected just before the experiment, centrifuged at 500× g for 7 min, suspended in serum-free medium (SFM, GIBCO) supplemented with 500 units.ml−1 penicillin and 50 µg.ml−1 streptomycin, and counted.
Rat resident peritoneal cells were obtained by injection of sterile PBS into the peritoneal cavity. Collected cells were centrifuged and resuspended in Serum Free Medium (SFM) (Life Technologies, Inc) and counted. Macrophages were obtained by adhering cells for 1 h at 37°C and 5% CO2. After 1 h, non-adherent cells were removed by gentle washing with SFM and parasites were added to macrophages settled on coverslips at a ratio of 3∶1. After incubation for 1 h at 37°C, wells were washed 3 times with SFM to remove extracellular parasites and cells fixed at different times post-infection in 4% formaldehyde.
Caspase-1 inhibitor VI (YVAD) and caspase inhibitor VI (pan-caspase, Z-VAD) were from Calbiochem (Merck Chemicals, France). Macrophages were incubated with 50 µM of Caspase-1 inhibitor or 100 µM of pan-caspase inhibitor for 2 h before infection and during all the infection.
Infected macrophages were permeabilized with 0.002% saponine or 0.1% triton-×100 to detect SAG1 (mAb Tg05-54), GRA5 (mAb Tg17-113) or GRA3 (mAb Tg2H1). The rabbit anti-Tg small ubiquitin-like modifier (TgSUMO) polyconal antibody was kindly provided by M.A. Hakimi (Grenoble, France) [15]. To detect acidic vacuoles, infected macrophages were stained with 50 nM LysoTracker Red for 30 min prior to fixation. Cell death was analyzed by visualizing the uptake of Propidium Iodide (PI) (Molecular Probes). Alexa488 and Alexa594 antibody conjugates (Molecular Probes) were used as secondary antibodies. Coverslips were mounted in mowiol and observed and counted with a Zeiss Axioplan 2 microscope equipped for epifluorescence and phase-contrast. Kinetic experiments were performed on the IX2 Olympus microscope and analyzed with ScanR software.
For analysis of caspase-1 activation, we used a fluorescent caspase-1 activity assay, FLICA, from Immunochemistry Technologies (Bloomington, MN). The assay was performed in 24-well plates, with 5.105 cells per well. Cells were incubated with T. gondii and stained with FLICA reagent (FAM-YVAD-FMK) as recommended by the manufacturer. Fluorescence was measured on the IX2 Olympus microscope and analyzed with ScanR software.
The level of apoptosis of infected LEW macrophages was assessed with Annexin V-propidium iodide staining. Rat peritoneal exudent cells were treated with etoposide (Sigma Aldrich, 50 µM) or incubated with parasites in SFM at a MOI of 1∶3 at 37°C, 5% CO2. After 1 h, cells were collected by centrifugation and incubated 4 h at 37° after addition of fresh medium. Macrophages were stained by the addition of Annexin V (Santa Cruz biotechnologies, FL-319) and Alexa488 secondary antibody (Molecular Probes) for 30 min and 1 µg/mL of PI for 5 min. Data acquisition was performed by flow cytometry on 4-colours FACSCalibur (BD Biosciences) equipped with 488 nm argon laser and CellQuest Software.
Rat peritoneal macrophages (∼106) were infected with 3×106 parasites for 1 h, washed and returned to 37°C for 6 h, 12 h or 18 h. For positive control of apoptosis, macrophages were treated with 50 µM etoposide (Sigma Aldrich). Macrophages were treated in lysis-buffer (100 mM Tris, pH 8.0, 100 mM EDTA, 4% Sodium dodecyl sulfate (SDS) with 1 µg/ml RNAse (Roche) (30 min, 37°C) followed by treatment with 100 µg/ml proteinase K (Euromedex) (30 min, 55°C). After extraction with phenol/chloroform, DNA was recovered by precipitation and analysed on 1.8% agarose gels.
Rat peritoneal exudent cells were incubated with parasites in SFM at a MOI of 1∶3 at 37°C, 5% CO2 for 15 min or 4 h and then incubated with 0.45 µM of dihydro-rhodamine 123 (DHR, Sigma Aldrich) 15 min at 37°C. At the end of incubation, FACS lysing buffer (BD Bioscience, Pont de Claix, France) was added to each sample and incubated for 15 min. The samples were then washed in PBS before analysis with a FACSCalibur flow cytometer (Becton Dickinson) and the CellQuest Pro software (Becton Dickinson).
Cells were lysed in cold RIPA (50 mM Tris-Hcl pH 7,4, 150 mM NaCl, 1% NP40, 0,25% Na-deoxycholate) buffer supplemented with protease inhibitors and centrifuged at 4°C and 13,000 g for 10 min. Supernatant of infected or uninfected macrophages (5.105 cells/well) were collected and precipitated with TCA (trihloroacetate) for 10 min at 4°C prior centrifugation at 16 000 g for 5 min. Pellets were washed two times in acetone then dried and resuspended in laemmli buffer. Protein extracts were subjected to electrophoresis on a 12% Tris-HCl SDS-PAGE and transferred to PVDF membranes (Amersham). Membranes were blocked for 1 h in TTBS (100 mM Tris-HCl, 0.9% NaCl, and 0.05% Tween 20) containing 5% skim milk before incubating overnight at 4°C with primary 1/500 anti-caspase 3 (Cell Signaling), 1/1000 anti-IL-1β (Millipore, AB1832P), 1/5000 anti-GAPDH (Santa Cruz) or 1/1000 anti-actin (Sigma Aldrich) antibodies followed by 1/10000 anti-rabbit secondary horseradish peroxidase (HRP)-linked antibodies (Jackson Immunoresearch). Visualization of signals was enhanced by luminol-based chemiluminescence (ECL, ThermoFisher Scientific).
The intracellular growth of T. gondii in rat peritoneal macrophages was monitored by selective incorporation of [3H]uracil as previously described [32]. Briefly, 5×105 macrophages were infected with 1.5×106 parasites for 1 h in SFM at 37°C and 5% CO2. After washing to eliminate extracellular parasites, cells were cultured for 20 h in the presence of [3H]uracil (5 µCi per well, Ci = 37 GBq). Monolayers were washed three times in PBS, disrupted with 500 µl of lysis/scintillation solution (Optiphase Supermix, Perkin Elmer) and radioactivity measured by liquid scintillation counting using a Wallac MicroBeta TriLux (Perkin Elmer).
Preparation of genomic DNA and genotyping were performed as described [33]. To genotype the 35 (LOU X BN) F2 rat progenies, six microsatellite markers (D10Rat49, D10Arb2, D10GF41, D10Rat27, D10Mgh4, D10Rat2) were selected to cover chromosome 10 with an average spacing of 20 Mb. For haplotype analysis, the polymorphism of 41 microsatellite sequences around the Toxo1 locus were analysed in the nine strains. Among these 41 markers, 17 newly identified microsatellite markers (Table S1) were used in addition to the 24 microsatellite markers selected from Rat Genome Database (RGD).
The anti-Toxoplasma IgG response was measured by specific enzyme-linked immunosorbent assay (ELISA). Total Toxoplasma antigens were prepared as previously described [34]. Immuno plates Maxisorp (Nunc) were coated overnight at 4°C with Toxoplasma antigens at 20 µg/ml. After washing, saturation was 1 h at 37°C with PBS containing 5% of skim milk. Sera were diluted at 1/20 and 1/1000 in PBS-0.01% Tween 20 and incubated 1h30 at 37°C. Plates were then washed with PBS-0.01% Tween 20, and peroxydase-conjugated anti-rat IgG (KPL) secondary antibody diluted at 1/5000 was incubated 1 h at 37°C. Finally, after six washes, 100 µl of substrate TMB-hydrogen peroxyde (TMB Ultra 1 Step, ThermoScientific) solution was added to the wells. The color reaction was stopped adding 50 µl of 3 N HCl. Optical densities at 492 nm and 630 nm were measured using an ELx800 absorbance microplate reader (Bio TeK Instruments). Results were expressed as arbitrary units.
Each rat brain was removed and homogenized in 16 ml of PBS. Brain suspensions were clarified by gentle incubation in proteinase K buffer (proteinase K 0.4 µg/ml, 10 mM Tris pH 8, EDTA 1 mM, sodium dodecyl sulfate 0.2%, sodium chloride 40 mM) for 15 min at 56°C. The reaction was stopped by incubation with PMSF 2 mM for 5 min at room temperature. Then, the suspension was washed with PBS and resuspended with FITC-Dolichos biflorus agglutinin (Vector laboratories, CA USA) 20 µg/ml for 30 min, at room temperature. After washing in PBS, rat brains were resuspended in 6 ml of PBS, distributed into six-well culture plates (1 ml per well) and cysts counted visually with an Axiovert 40 CFL inverted fluorescence microscope (Zeiss) [35].
A custom-made SureSelect oligonucleotide probe library was designed by IntegraGen (Evry, France) to capture the chr10 region containing Toxo1 locus (chr10: 57,200,000–58,200,000). The eArray web-based probe design tool was used for this purpose (https://earray.chem.agilent.com/earray). A total of 56,450 probes, covering a target of 6830692 bp, were synthesized by Agilent Technologies (Santa Clara, CA, USA). Library preparation, capture enrichment, sequencing, and variants detection and annotation, were performed by IntegraGen (Evry, France). Briefly, 3 µg of each genomic DNA were fragmented by sonication and purified to yield fragments of 150–200 bp. Paired-end adaptor oligonucleotides from Illumina were ligated on repaired DNA fragments, which were then purified and enriched by six PCR cycles. 500 ng of the purified libraries were hybridized to the SureSelect oligo probe capture library for 24 h. After hybridization, washing, and elution, the eluted fraction underwent 14 cycles of PCRamplification. This was followed by purification and quantification by qPCR to obtain sufficient DNA template for downstream applications. Each eluted-enriched DNA sample was then sequenced on an Illumina GAIIx as paired-end 75 bp reads. Image analysis and base calling was performed using Illumina Real Time Analysis (RTA) Pipeline version 1.10 with default parameters. Sequence reads were aligned to the reference rat genome (UCSC rn4) using commercially available software (CASAVA1.7, Illumina) and the ELANDv2 alignment algorithm. Sequence variation annotation was performed using the IntegraGen in-house pipeline, which consisted of gene annotation (RefSeq), detection of known polymorphisms (dbSNP 125) followed by mutation characterization (exonic, intronic, silent, nonsense etc.).
For in vivo experiments, data are expressed as means ± SEM and the significance of differences found between groups was initially derived from a Kruskal-Wallis H test and subsequently confirmed by the Mann-Whitney test. For in vitro experiments, data are expressed as means ± SD and the significance of differences found between groups was determined using two-tailed Student's t test.
|
10.1371/journal.pgen.1006941 | Dynactin binding to tyrosinated microtubules promotes centrosome centration in C. elegans by enhancing dynein-mediated organelle transport | The microtubule-based motor dynein generates pulling forces for centrosome centration and mitotic spindle positioning in animal cells. How the essential dynein activator dynactin regulates these functions of the motor is incompletely understood. Here, we dissect the role of dynactin's microtubule binding activity, located in the p150 CAP-Gly domain and an adjacent basic patch, in the C. elegans zygote. Analysis of p150 mutants engineered by genome editing suggests that microtubule tip tracking of dynein-dynactin is dispensable for targeting the motor to the cell cortex and for generating robust cortical pulling forces. Instead, mutations in p150's CAP-Gly domain inhibit cytoplasmic pulling forces responsible for centration of centrosomes and attached pronuclei. The centration defects are mimicked by mutations of α-tubulin's C-terminal tyrosine, and both p150 CAP-Gly and tubulin tyrosine mutants decrease the frequency of early endosome transport from the cell periphery towards centrosomes during centration. Our results suggest that p150 GAP-Gly domain binding to tyrosinated microtubules promotes initiation of dynein-mediated organelle transport in the dividing one-cell embryo, and that this function of p150 is critical for generating cytoplasmic pulling forces for centrosome centration.
| Animal cells rely on molecular motor proteins to distribute intracellular components and organize their cytoplasmic content. The motor cytoplasmic dynein 1 (dynein) uses microtubule filaments as tracks to transport cargo from the cell periphery to the cell center, where the microtubule minus ends are embedded at the centrosome. Conversely, when dynein is anchored at the cell cortex or on organelles in the cytoplasm, the motor can pull on microtubules to position centrosomes within the cell. The intracellular location of centrosomes determines cell geometry and cell fate, and studying the underlying mechanisms will help us understand polarized cell behaviors such as cell migration or neurite outgrowth, and how cleavage plane orientation is established during cell division. Here, we show in C. elegans embryos that dynactin, an essential dynein regulator, uses its microtubule binding activity to help dynein pull on microtubules for centrosome positioning during the first mitotic division. Our results with engineered dynactin and tubulin mutants suggest that microtubule binding by dynactin increases the efficiency with which dynein can initiate the transport of small organelles towards centrosomes. More organelles moving along microtubules through the viscous cytoplasm means that correspondingly larger pulling forces act on centrosomes. Thus, our work provides evidence for a novel functional link between dynactin's role in initiating transport of dynein cargo and the generation of cytoplasmic pulling forces critical for the positioning of centrosomes.
| Cytoplasmic dynein 1 (dynein) is the major microtubule (MT) minus-end directed motor in animals and transports various cargo from the cell periphery to the cell interior. The motor also moves and positions intracellular structures such as nuclei and centrosomes by pulling on the MTs to which they are connected. To generate pulling force, dynein is either attached to anchor proteins fixed at the cell cortex (cortical pulling) [1–4], or dynein is anchored on organelles in the cytoplasm (cytoplasmic pulling) [5,6]. In the latter instance, dynein generates MT length-dependent pulling forces by working against viscous drag as it transports organelles along MTs toward centrosomes.
Dynactin is an essential multi-subunit activator of dynein that forms a tripartite complex with the motor and cargo-specific adaptors proteins [7–11], but how dynactin supports the diverse functions of dynein remains incompletely understood. Dynactin is built around a short actin-like Arp1 filament and has its own MT binding activity, which resides at the end of a long projection formed by the largest subunit p150 [12]. p150 has a tandem arrangement of MT binding regions consisting of an N-terminal cytoskeleton-associated protein glycine-rich (CAP-Gly) domain and an adjacent patch rich in basic residues [13,14]. The CAP-Gly domain binds to MTs and to the MT plus-end tracking proteins (+TIPs) CLIP-170 and end-binding (EB) protein. In animal cells, +TIP binding of dynactin recruits dynein to growing MT ends [9,15–18].
The p150 CAP-Gly domain recognizes the C-terminal EEY/F motif present in α-tubulin and EB/CLIP-170 [19–22]. The C-terminal tyrosine of α-tubulin can be removed and re-ligated in a tyrosination-detyrosination cycle and is proposed to regulate the interactions with molecular motors and other MT binding proteins [23,24]. Tubulin tyrosination is required in mouse fibroblasts to localize CAP-Gly proteins, including p150, to MT plus ends [25], and recent work in vitro demonstrated that the interaction between p150's CAP-Gly domain and tyrosinated MTs enhances the initiation of processive dynein motility [26].
The functional significance of MT binding by p150 in animals is best understood in neurons. Single point mutations in the CAP-Gly domain cause the ALS-like motor neuron degenerative disease HMN7B and a form of parkinsonism known as Perry syndrome [27–29]. Cellular and in vivo studies addressing the underlying molecular defects revealed that p150 CAP-Gly domain-dependent binding of dynactin to dynamic MTs in the distal axon enhances the recruitment of dynein, which in turn facilitates efficient initiation of retrograde transport [30–32].
While the critical role of p150's CAP-Gly domain in neuronal trafficking is firmly established, little is known about how MT binding by dynactin regulates dynein functions in other cellular contexts. A study in D. melanogaster S2 cells reported multipolar spindles with a p150Glued construct lacking the CAP-Gly domain, suggesting a role in organizing MT arrays [33]. In budding yeast, introduction of the motor neuron disease mutation into p150Nip100 inhibited the initial movement of the spindle and nucleus into the bud neck during mitosis [34], suggesting that dynactin binding to MTs helps dynein generate pulling forces under load. In budding and fission yeast, dynein is off-loaded to cortical anchors via MTs for subsequent force production, and in budding yeast this requires MT tip tracking of dynein [35–39]. Whether MT tip tracking of dynein plays a role in delivering the motor to the cortex in animal cells remains to be determined.
MT binding of dynactin is significantly enhanced by electrostatic interactions between the p150 basic patch and the acidic tails of tubulins [14,40,41]. In the filamentous fungus A. nidulans, deletion of the basic patch in p150NUDM diminishes the accumulation of dynactin and dynein at MT tips and partially impairs nuclear migration and early endosome distribution [42]. Interestingly, humans express tissue-specific splice isoforms of p150 that lack the basic patch [43,44], but the implications for dynactin function are unclear.
In the C. elegans one-cell embryo, dynein and dynactin are essential for centrosome separation, migration of the maternal and paternal pronucleus, centration and rotation of the two pronuclei and the associated centrosomes (the nucleus-centrosome complex, NCC), assembly and asymmetric positioning of the mitotic spindle, chromosome congression, and transversal spindle oscillations in anaphase [45–49]. Here, we use a set of p150dnc-1 and α-tubulin mutants constructed by genome editing to define the role of dynactin's MT binding activity in this system. Our results uncover a functional link between the efficient initiation of dynein-mediated organelle transport, which requires dynactin binding to tyrosinated MTs, and the cytoplasmic pulling forces responsible for centration of centrosomes.
To investigate whether dynactin's MT binding activity contributes to dynein function in the early C. elegans embryo, we first asked whether dynactin is present at MT plus ends at this developmental stage. Live confocal imaging in the central plane of metaphase one-cell embryos co-expressing endogenous GFP::p50DNC-2 and transgene-encoded EBP-2::mKate2 revealed that dynactin travelled on growing MT tips from mitotic spindle poles to the cell cortex (Fig 1A, S1 Movie). Imaging of the cortical plane allowed end-on visualization of MT tips as they arrived at the cortex (Fig 1B and 1C), which facilitated quantification of dynactin levels at plus ends. Measurements of fluorescence intensity revealed the expected positive correlation between GFP::p50DNC-2 and EBP-2::mKate2 levels, but also showed that there is considerable variation in the amount of GFP::p50DNC-2 at MT plus ends (S1A Fig). Cortical residency times for EBP-2::mKate2 and GFP::p50DNC-2 were nearly identical (1.67 ± 0.03 s and 1.50 ± 0.05 s, respectively) and agreed with previously published measurements for cortical residency times of GFP::EBP-2 (S1C and S1D Fig) [50]. We also generated a dynein heavy chaindhc-1::gfp knock-in allele to assess the localization of endogenous dynein. DHC-1::GFP was readily detectable on growing MT plus ends in early embryos (S1B Fig, S2A Fig, S5A Fig, S2 Movie), although the signal appeared weaker than that of GFP::p50DNC-2. We conclude that a pool of dynein-dynactin tracks with growing MT plus ends in the early C. elegans embryo.
We used the quantitative cortical imaging assay to determine which +TIPs were required for MT tip targeting of dynactin and dynein. RNAi-mediated depletion of the three EB paralogs revealed that EBP-2 is required for GFP::p50DNC-2 targeting to MT tips, while EBP-1 and EBP-3 are dispensable (Fig 1B and 1D, S2C Fig). In mammalian cells, CLIP-170 acts as an essential linker between EB and dynactin [51–53]. To assess whether the CLIP-170-like protein CLIP-1 recruits dynactin to MT tips in C. elegans, we generated a null allele of clip-1 in the gfp::p50dnc-2 background (S2D Fig). This revealed that CLIP-1 is dispensable for MT tip localization of GFP::p50DNC-2 (Fig 1C and 1D), suggesting that dynactin is directly recruited by EBP-2. Next, we depleted dynein intermediate chainDYCI-1 and the dynein co-factor LIS-1. In both cases, GFP::p50DNC-2 levels at MT tips decreased substantially (Fig 1C and 1D). Conversely, depletion of p150DNC-1 showed that DHC-1::GFP targeting to MT tips was dependent on dynactin (S2A and S2B Fig).
We conclude that in the C. elegans early embryo, dynein and dynactin are interdependent for targeting to growing MT plus ends and require EBP-2 and LIS-1, but not EBP-1, EBP-3, or the CLIP-170 homolog CLIP-1 (Fig 1E).
Having established that dynein and dynactin require the EB homolog EBP-2 for targeting to MT tips, we next examined the role of the dynactin subunit p150DNC-1, whose N-terminal CAP-Gly domain (residues 1–69) mediates binding to EB and MTs (Fig 2A). In addition, p150DNC-1 contains a ~200-residue basic-serine rich region between the CAP-Gly domain and the first coiled-coil (CC1A), which has been proposed to regulate p150DNC-1 association with MTs [54,55] (Fig 2A). The highest density of basic residues is found between residues 140–169 (30% K or R, pI = 12.02). This region is encoded by exon 4 and part of exon 5, which are subject to alternative splicing (Fig 2B, S3A Fig). This is similar to human p150, which contains an alternatively-spliced basic patch of 28 residues (43% K or R, pI = 12.7) adjacent to the CAP-Gly domain [43]. We detected four splice isoforms of p150dnc-1 by reverse transcription PCR of RNA isolated from adult animals (S3B Fig): full-length p150dnc-1 including exons 4 and 5, p150dnc-1 without exon 4 (Δexon 4), p150dnc-1 without exon 5 (Δexon 5), and p150dnc-1 lacking exons 4 and 5 (Δexon 4–5). To define the function of individual splice isoforms, we edited the p150dnc-1 locus to generate animals in which p150dnc-1 expression was restricted to one of the four isoforms (Fig 2B, S3A Fig). Reverse transcription PCR confirmed that animals expressed single p150dnc-1 isoforms corresponding to full length, Δexon 4, Δexon 5, or Δexon 4–5 (S3B Fig). All mutant animals were homozygous viable and fertile (S3C Fig), demonstrating that none of the p150DNC-1 isoforms is essential. Despite differences in predicted molecular weight of only a few kDa (S3A Fig), single isoforms expressed in mutant animals were distinguishable by size on immunoblots with an antibody raised against a C-terminal region of p150DNC-1 (Fig 2C). Side-by-side comparison of isoform mutants and wild-type animals on the same immunoblot revealed that neither full-length p150DNC-1 nor p150DNC-1(Δexon 4–5) is prevalent in wild-type adults (Fig 2C). Instead, immunoblotting, together with reverse transcription PCR data (S3B and S3D Fig), suggested that p150DNC-1(Δexon 4) is the predominant isoform.
Humans express the neuron-specific splice variant p135, which lacks the entire N-terminal MT binding region [56]. In C. elegans hermaphrodite adults, 302 out of 959 somatic cells are neurons, yet we did not find evidence for a p135dnc-1 isoform at the mRNA level (S3D Fig), nor did our p150DNC-1 antibody detect any protein below ~150 kDa in wild-type animals (Fig 2C). We also generated a p150dnc-1::3xflag knock-in allele, and immunoblotting with antibody against 3xFLAG similarly failed to detect a p135 isoform (S3E Fig). We speculated that specifically suppressing the expression of p150DNC-1 isoforms might facilitate the detection of p135 and engineered a null allele of p150dnc-1 by inserting a stop codon in exon 1 immediately following the start codon (S3A Fig). The null mutation did not affect splicing of p150dnc-1 mRNA (S3B Fig) and therefore should permit expression of p135 from an alternative start codon, as is the case in humans [56]. However, immunoblotting produced no evidence of p135 expression in the absence of p150DNC-1 isoforms (Fig 2C). We conclude that C. elegans does not express significant amounts of a p135 isoform.
Next, we used cortical imaging of GFP::p50DNC-2 in one-cell embryos to determine the effect of p150DNC-1 isoforms on dynactin recruitment to MT tips. Full-length p150DNC-1 and p150DNC-1(Δexon 4) fully supported dynactin targeting to MT tips, and dynactin levels were even slightly increased (108 ± 4% of controls) for the Δexon 4 isoform (Fig 2D). By contrast, expression of p150DNC-1(Δexon 5) or p150DNC-1(Δexon 4–5) decreased dynactin levels at MT tips to 73 ± 4% and 86 ± 4% of controls, respectively (Fig 2D). Thus, surprisingly, the similarly basic regions encoded by exon 4 (27% K/R; pI = 11.2) and exon 5 (19% K/R; pI = 12) make differential contributions to dynactin targeting. We conclude that splice isoforms of p150DNC-1 regulate dynactin levels at MT tips.
Our analysis of p150DNC-1 isoforms suggested that the basic region had a relatively minor role in targeting dynactin to MT tips. To examine the role of the CAP-Gly domain, we used genome editing to separately introduce three point mutations into p150DNC-1 that compromise CAP-Gly domain function and cause neurodegenerative disease in humans (Fig 2E): G33S corresponds to human G59S, which causes motor neuropathy 7B [27]; G45R corresponds to human G71R, which causes Perry Syndrome [28]; and F26L corresponds to human F52L, which was recently identified in a patient with Perry Syndrome-like symptoms [29]. The F26L and G45R mutants could be propagated as homozygotes with high embryonic viability (99 ± 1% and 90 ± 2%, respectively), whereas the G33S mutant was lethal in the F2 generation (1 ± 1% embryonic viability) (S4A Fig). Immunoblotting of homozygous F1 adults showed that G33S animals had decreased levels of p150DNC-1, indicating that the mutation destabilized the protein (Fig 2F and 2G). By contrast, total levels of p150DNC-1 were not affected in the F26L or G45R mutant. Central plane imaging in one-cell embryos expressing GFP::p50DNC-2 showed that dynactin containing the F26L or G45R mutation was present on the mitotic spindle and prometaphase kinetochores but displaced from MT tips (Fig 2H, S3 Movie). Cortical imaging after introduction of the EBP-2::mKate2 marker revealed that GFP::p50DNC-2 levels at MT tips were reduced to 34 ± 4% and 27 ± 4% of controls in the F26L and G45R mutant, respectively (Fig 2I and 2J, S4 Movie, S5 Movie). Deletion of the basic patch encoded by exons 4 and 5 in the G45R mutant (G45R + Δexon 4–5) further reduced GFP::p50DNC-2 levels at MT tips to 15 ± 4% (Fig 2J) but had no additive effect on embryonic viability (90 ± 2%) (S4A Fig). Additional quantifications showed that in both the F26L and G45R mutant, GFP::p50DNC-2 still targeted to the nuclear envelope and kinetochores, while GFP::p50DNC-2 levels were reduced on spindle MTs (S4B Fig). We also introduced the mutations into animals expressing DHC-1::GFP, which confirmed that dynein levels were decreased at MT tips and on spindle MTs (S5B–S5D Fig). We conclude that point mutations in the p150DNC-1 CAP-Gly domain that cause human neurodegenerative disease reduce dynein-dynactin levels on MTs and greatly diminish the ability of dynein-dynactin to track with MT tips.
Next, we asked whether the p150dnc-1 mutants affected dynein-dynactin function in the one-cell embryo. We crossed the mutants with animals co-expressing GFP::histone H2B and GFP::γ-tubulin, which allowed precise tracking of pronuclei and centrosomes, respectively (Fig 3A). None of the mutants exhibited defects in centrosome separation, and pronuclear migration along the anterior-posterior axis proceeded with normal kinetics until pronuclear meeting, which occurred at the correct position in the posterior half of the embryo (Fig 3A and 3B, S6A Fig). However, subsequent centration of the nucleus-centrosome complex (NCC) slowed substantially in p150dnc-1 F26L, G45R, and G45R + Δexon 4–5 mutants, and NCC rotation was defective (Fig 3A–3D, S6 Movie). NCC centration was not significantly perturbed in the isoform mutants (S6A and S6B Fig), but the Δexon 4–5 mutant exhibited defects in NCC rotation (Fig 3D, S6C Fig). In all mutants, spindle orientation recovered during prometaphase, so that the spindle axis was largely aligned with the anterior-posterior axis of the embryo at the time of anaphase onset (Fig 3D, S6C Fig).
In controls, the mitotic spindle was displaced from the embryo center toward the posterior in preparation for asymmetric division (Fig 3C). By contrast, spindle assembly in p150dnc-1 CAP-Gly mutants already occurred in the posterior half of the embryo, and the spindle had to be moved only slightly to the posterior to be correctly positioned. In controls, the regular and vigorous oscillations of spindle rocking began at anaphase onset and lasted for approximately 100 s (Fig 3E, S7 Movie). By contrast, spindle rocking in p150dnc-1 CAP-Gly mutants was irregular and significantly dampened.
In addition to defects in NCC centration/rotation and spindle rocking, we observed a slight but consistent delay in chromosome congression in p150dnc-1 CAP-Gly mutants, indicating problems with the interactions between chromosomes and spindle MTs (S7A and S7B Fig, S6 Movie, S7 Movie). This did not result in obvious chromosome mis-segregation in the first embryonic division (S7A Fig). However, when the spindle assembly checkpoint (SAC) was inactivated by RNA-mediated depletion of Mad1MDF-1, embryonic viability decreased by 28% and 22% in the G45R and F26L mutant, respectively, whereas Mad1mdf-1(RNAi) in controls decreased embryonic viability by just 6% (S7C Fig). This suggests that SAC signaling is required during embryogenesis to prevent chromosome segregation errors when the p150DNC-1 CAP-Gly domain is compromised.
We conclude that mutations in the p150DNC-1 CAP-Gly domain perturb a specific subset of dynein-dynactin functions in the one-cell embryo.
Anaphase spindle rocking requires cortical dynein pulling on astral MTs [2,57]. Since spindle rocking was affected in p150dnc-1 CAP-Gly mutants, we sought to assess the extent of phenotypic overlap between p150dnc-1 CAP-Gly mutants and inhibition of dynein-dependent cortical pulling. We tracked centrosomes and pronuclei after co-depleting GPR-1 and GPR-2, which are required for cortical anchoring of dynein-dynactin [2,57]. In contrast to p150dnc-1 CAP-Gly mutants, gpr-1/2(RNAi) delayed the initial separation of centrosomes and the onset of pronuclear migration (Fig 4A–4C). Pronuclear migration and NCC centration subsequently occurred at a slightly faster rate than in controls, so that the NCC achieved near-normal centration by nuclear envelope breakdown (NEBD) (Fig 4A and 4C). These results are consistent with slowed centrosome separation and faster centering reported after co-depletion of GOA-1 and GPA-16, the Gα proteins acting upstream of GPR-1/2 [58,59]. Thus, the kinetics of pronuclear migration and NCC centration differ between gpr-1/2(RNAi) and p150dnc-1 CAP-Gly mutants. NCC rotation, by contrast, was affected in both perturbations (Fig 4D). Importantly, gpr-1/2(RNAi) in the p150dnc-1(G45R) mutant enhanced the rotation defect of gpr-1/2(RNAi) on its own, arguing that GPR-1/2 and the p150DNC-1 CAP-Gly domain contribute to NCC rotation through parallel pathways. After NEBD, depletion of GPR-1/2 prevented posterior displacement of the spindle and the lack of cortical pulling was especially evident in the track of the posterior centrosome (Fig 4C). In addition, the mitotic spindle was shorter than controls during metaphase and failed to elongate properly in anaphase (Fig 4B). By contrast, posterior centrosome movement towards the cortex in p150dnc-1 CAP-Gly mutants was indistinguishable from controls (Fig 3C, Fig 4C), and spindle length was normal throughout metaphase and anaphase (Fig 4B). These results argue that, although spindle rocking is compromised in p150dnc-1 CAP-Gly mutants, cortical dynein is still able to generate robust pulling forces on astral MTs. We also tracked centrosomes after depletion of EBP-2, which, just like p150dnc-1 CAP-Gly mutants, delocalized dynein-dynactin from MT tips (Fig 1B and 1D). Strikingly, posterior spindle displacement was exaggerated in ebp-2(RNAi) embryos compared with controls (Fig 4C). These results suggest that cortical pulling forces used for asymmetric spindle positioning can be generated in the absence of MT tip-localized dynein-dynactin.
The robust dynein-dependent cortical pulling observed in one-cell embryos depleted of EBP-2 and in embryos of p150dnc-1 CAP-Gly mutants implied that the motor was able to target to the cortex under these conditions. To test this directly, we measured the intensity of the DHC-1::GFP signal in line scans drawn across the cortex in one-cell embryos at anaphase. This revealed a cortically enriched pool of DHC-1::GFP that was dependent on GPR-1/2, as expected (Fig 4E). Cortical dynein in the one-cell embryo was unaffected in the p150dnc-1(G45R) mutant and after ebp-2(RNAi). We also imaged the 4-cell embryo, in which dynactin and dynein become prominently enriched at the EMS-P2 cell border prior to EMS and P2 spindle rotation [60,61]. Quantification of GFP::p50DNC-2 and DHC-1::GFP levels at the EMS-P2 cell border revealed that cortical levels of dynein-dynactin were unchanged in the p150dnc-1(G45R) mutant and after ebp-2(RNAi) (Fig 4F). We conclude that MT tip tracking of dynein-dynactin is dispensable for cortical targeting of the motor in the early embryo.
CAP-Gly domains bind the C-terminal EEY/F motif of α-tubulin, and the tyrosine residue is critical for the interaction (Fig 2A) [22]. We therefore asked whether decreased affinity of dynactin for tyrosinated MTs could be contributing to the defects observed in p150dnc-1 CAP-Gly mutants. Of the 9 α-tubulin isoforms in C. elegans, TBA-1 and TBA-2 are the major α-tubulin isotypes expressed during early embryogenesis [62]. We mutated the C-terminal tyrosine of TBA-1 and TBA-2 to alanine and obtained animals homozygous for either mutation alone (YA) or both mutations combined (YA/YA) (Fig 5A). Immunoblotting of adult animals with the monoclonal antibody YL1/2, which is specific for tyrosinated tubulin, revealed that levels of tubulin tyrosination were decreased in tba-1(YA) and tba-2(YA) single mutants, with tba-2(YA) having a more pronounced effect (Fig 5B). Combining the two mutations dramatically decreased total levels of tubulin tyrosination. Importantly, immunoblotting with an antibody insensitive to tubulin tyrosination confirmed that total α-tubulin levels were not affected in the three mutants (Fig 5B). We then used immunofluorescence to directly assess tyrosinated tubulin levels in the early embryo. In controls, the mitotic spindle of the one-cell embryo was prominently stained with the antibody against tyrosinated tubulin (Fig 5C). By contrast, the tubulin tyrosination signal was undetectable in the tba-1/2(YA/YA) double mutant, despite normal spindle assembly. Thus, we were able to generate animals without detectable tubulin tyrosination in early embryos.
We next addressed the functional significance of tubulin tyrosination in the one-cell embryo. Strikingly, we found that the tba-1/2(YA/YA) mutant exhibited NCC centration/rotation defects reminiscent of those observed in p150dnc-1 CAP-Gly mutants (Fig 5D–5G, S8 Movie). Combining the tba-1/2(YA/YA) mutant with the p150dnc-1(G45R) mutant did not significantly exacerbate the centration/rotation defects of the either mutant on its own, indicating that both mutants act in the same pathway (Fig 5D–5G, S8 Movie). Interestingly, in contrast to p150dnc-1 CAP-Gly mutants, anaphase spindle rocking was not affected in the tba-1/2(YA/YA) mutant (Fig 5H).
We also examined the effect of the tba-1/2(YA/YA) mutant on GFP::p50DNC-2 localization and found that dynactin levels at MT tips were identical to controls (Fig 5I). Thus, in contrast to mouse fibroblasts [25], tubulin tyrosination in the C. elegans embryo is not required to target dynactin to MT tips.
Dynein-mediated transport of small organelles along MTs towards centrosomes is proposed to generate the cytoplasmic pulling forces for centration (the centrosome-organelle mutual pulling model) [5,63,64]. To ask whether the centration defects in our mutants correlate with defects in MT minus end-directed organelle transport, we monitored the movement of early endosomes, labelled with mCherry::RAB-5, from pronuclear meeting until NEBD. Time-lapse sequences recorded in a focal plane that included the NCC were used for semi-automated tracking of early endosomes that moved from the cell periphery towards centrosomes (Fig 6A). In control embryos, we counted 16.3 ± 3.5 tracks/min during the ~6 min centration interval (Fig 6B). This was reduced to 0.8 ± 0.6 tracks/min in embryos depleted of p150DNC-1 by RNAi, confirming that dynactin is required for early endosome movement directed towards centrosomes. The p150dnc-1(G45R + Δexon 4–5) mutant also strongly reduced the number of observed tracks to 5.3 ± 1.1 tracks/min (Fig 6B, S9 Movie). The tba-1/2(YA/YA) mutant had a less severe effect but still substantially reduced the number of tracks to 10.6 ± 2.2 per min. We also determined the maximal velocity in each track (determining the mean speed was complicated by frequent pausing of particles) and the total track displacement. This revealed only minor differences between controls and either mutant (Fig 6B). We conclude that p150dnc-1 CAP-Gly and α-tubulin tyrosine mutants reduce the frequency with which early endosomes move towards centrosomes during the centration phase. These results are consistent with the idea that dynactin binding to tyrosinated MTs enhances the efficiency of transport initiation by dynein, as recently documented in vitro [26].
In the distal axon of neuronal cells, EB-dependent recruitment of dynactin to dynamic MT plus ends is proposed to ensure efficient initiation of retrograde transport by dynein [32]. To test whether EBP-2 plays a role in the initiation of centrosome-directed organelle transport in the one-cell C. elegans embryo, we tracked early endosomes after ebp-2(RNAi). The number of early endosome tracks was reduced from 16.3 ± 3.5 to 6.9 ± 1.7 per minute after EBP-2 depletion (Fig 6B). Strikingly, ebp-2(RNAi) in the tba-1/2(YA/YA) mutant further reduced the number of early endosome tracks to 2.7 ± 1.0 per minute (Fig 6B) and enhanced the NCC centration defect compared to the individual perturbations (Fig 6C, S8 Movie). This suggests that EBP-2 is able to promote the initiation of dynein-mediated transport from MT tips even in the absence of tubulin tyrosination, consistent with the observation that dynactin targeting to MT tips is unaffected in the tubulin tyrosine mutant (Fig 5I). We conclude that EBP-2 and tubulin tyrosination independently contribute to the initiation of dynein-mediated organelle transport and NCC centration.
Dynactin's MT binding activity is crucial in neurons, as illustrated by single point mutations that compromise the function of the p150 CAP-Gly domain and cause neurodegenerative disease [30,31,65,66]. Here, we introduced these CAP-Gly mutations into C. elegans p150DNC-1 to investigate how dynactin's interaction with MTs and +TIPs contributes to dynein function in early embryogenesis. Together with the analysis of engineered p150dnc-1 splice isoform and tubulin tyrosine mutants, our work provides insight into the regulation and function of MT tip tracking by dynein-dynactin in animals and uncovers a link between dynactin's role in initiating dynein-mediated transport of small organelles and the generation of cytoplasmic pulling forces.
Dynein accumulates at and tracks with growing MT plus ends in species ranging from fungi to mammals, but requirements for MT tip tracking differ. In the C. elegans early embryo, MT tip recruitment of dynein-dynactin shares similarity with the pathway in budding yeast (dynactin depends on dynein and LIS-1) and mammalian cells/filamentous fungi (dynein depends on dynactin). Surprisingly, similar to what was reported for the fungus U. maydis [65], accumulation of dynein-dynactin at MT tips does not require a CLIP-170-like protein in C. elegans. Instead, dynactin is likely directly recruited by EBP-2, one of the three EB homologs. Work in mouse fibroblasts knocked out for tubulin tyrosine ligase showed that decreased tyrosinated tubulin levels displaced CLIP-170 and p150 from MT tips [25]. By contrast, we show that MT tip targeting of C. elegans dynactin is independent of tubulin tyrosination, possibly because there is no requirement for a CLIP-170 homolog. Overall, our analysis of dynein-dynactin targeting to MT tips in C. elegans highlights the evolutionary plasticity of +TIP networks.
Our characterization of engineered p150dnc-1 mutants establishes the functional hierarchy among p150DNC-1's tandem arrangement of MT binding regions: the CAP-Gly domain clearly provides the main activity, while the adjacent basic region plays an auxiliary role. Together with previous work in human cells [43], our results support the idea that alternative splicing of p150's basic region constitutes a conserved mechanism in animals for fine-tuning dynactin's affinity for MTs.
In budding yeast, dynein must first be targeted to MT tips prior to associating with cortical anchors [35,67]. Our results indicate that this pathway may not be used in C. elegans, as cortical accumulation of dynein-dynactin in the early embryo was unaffected in the p150dnc-1(G45R) mutant and after depletion of EBP-2, which displaced the majority of dynactin and dynein from MT tips. In agreement with normal cortical targeting of the motor, dynein-dependent cortical pulling forces remained robust in p150dnc-1 CAP-Gly mutants, although defects in spindle rocking indicate that the p150DNC-1 CAP-Gly domain does contribute to proper cortical force generation in anaphase. Importantly, depletion of EBP-2 even appeared to enhance cortical pulling during posterior spindle displacement. Thus, our results argue that dynein is recruited by its cortical anchors directly from the cytoplasm, and that dynein-dependent cortical pulling is therefore mechanistically uncoupled from prior MT tip tracking of the motor (Fig 6D). Surprisingly, even the p150dnc-1(G45R + Δexon 4–5) mutant, which shows the most severe reduction in dynactin levels at MT tips (15 ± 4% of controls), is viable and fertile, suggesting that MT tip tracking of dynein-dynactin is by and large dispensable for development.
If not delivery of dynein to the cell cortex via MTs, what is the purpose of dynactin's MT binding activity? We found that p150dnc-1 CAP-Gly mutants have defects in the centration and rotation of the NCC, which consists of the two centrosomes and the associated female and male pronucleus. Experimental work and biophysical modelling support the idea that centration forces in the one-cell embryo are generated by dynein-mediated cytoplasmic pulling [5,63,64], although a centration/rotation model based on cortical pulling has also been proposed [68]. In the cytoplasmic pulling model, dynein works against viscous drag as it transports small organelles (e.g. endosomes, lysosomes, yolk granules) along MTs towards centrosomes, which generates pulling forces on MTs that move the NCC. Prior work showed that movements of early endosomes and centrosomes are correlated, and RNAi-mediated depletion of adaptor proteins that tether dynein to early endosomes and lysosomes inhibited centration, indicating that there is a functional link between organelle transport and cytoplasmic pulling forces [5]. In agreement with this idea, the p150dnc-1(G45R + Δexon 4–5) mutant not only inhibited centration but also significantly decreased the number of early endosomes that displayed directed movement toward centrosomes. This effect on early endosome transport is consistent with the p150 CAP-Gly domain's role in initiating dynein-mediated transport, which is well-established in the context of retrograde axonal transport in neurons [30–32]. Compromising the efficiency with which organelle transport is initiated is predicted to decrease cytoplasmic pulling forces, because the magnitude of the net pulling force acting on centrosomes is proportional to the number of organelles travelling along MTs.
The frequency of centrosome-directed early endosome movement was also decreased in the tba-1/2(YA/YA) mutant, which severely reduced the levels of tubulin tyrosination in the early embryo. This fits well with recent work in vitro demonstrating that the interaction between the p150 CAP-Gly domain and tyrosinated MTs enhances the efficiency with which processive motility of dynein-dynactin is initiated [26]. Furthermore, a recent study in neurons provided evidence that initiation of retrograde transport in the distal axon is regulated by tubulin tyrosination [69]. Interestingly, depletion of EBP-2, both on its own and in the tba-1/2(YA/YA) mutant, also decreased early endosome transport. This suggests that EBP-2 promotes dynein-mediated transport initiation from MT tips, presumably through its interaction with the p150DNC-1 CAP-Gly domain, and that it can do so even in the absence of tyrosinated tubulin. In agreement with this idea, we observed that dynactin was still recruited to MT tips in the tubulin tyrosine mutant. Importantly, in addition to lowering the frequency of early endosome transport, the tba-1/2(YA/YA) mutant also affected centration of the NCC, and depletion of EBP-2 in the tba-1/2(YA/YA) mutant exacerbated the centration defect, as predicted by the centrosome-organelle mutual pulling model (Fig 6D).
Why do the p150dnc-1 CAP-Gly and tubulin tyrosine mutants affect centration of the NCC, but not pronuclear migration until pronuclear meeting? One plausible explanation is that during pronuclear migration the male and female pronuclei, which are large (~10 μm diameter) and equal in size, assist each other's movement as dyneins anchored on the female pronucleus walk along MTs nucleated by the centrosomes attached to the male pronucleus [70]. By contrast, during centration, the two pronuclei must be moved in the same direction, which might render cytoplasmic pulling forces more sensitive to changes in centrosome-directed transport of small (~1 μm diameter) organelles.
Finally, our data suggest that MT binding by dynactin contributes to chromosome congression. The effect is unlikely an indirect consequence of the delay in spindle orientation along the A-P axis, as chromosome congression problems were not observed after gpr-1/2(RNAi), which also causes spindle orientation defects. Likewise, normal chromosome congression after ebp-2(RNAi) suggests that the defect in p150dnc-1 CAP-Gly mutants is not due to delocalization of dynactin from MT tips. Therefore, it is likely that the contribution to chromosome congression comes from the p150DNC-1 CAP-Gly domain pool at kinetochores, where it could aid in the capture of MTs. The decrease in embryonic viability in p150dnc-1 CAP-Gly mutants after inhibition of the SAC indicates that the chromosome congression defects persist in later embryonic divisions.
In summary, our work demonstrates that dynactin's MT binding activity is functionally relevant in the context of embryonic cell division. Unlike previous work that addressed p150 CAP-Gly domain function in D. melanogaster S2 cells [33], we do not observe defects in bipolar spindle formation in p150dnc-1 CAP-Gly mutants. Instead, the most striking consequence of inhibiting p150DNC-1 CAP-Gly function or tubulin tyrosination is defective centrosome centration, which we propose is a consequence of defective initiation of dynein-mediated organelle transport, in agreement with the centrosome-organelle mutual pulling model [5]. The transport initiation function of p150's CAP-Gly domain is likely generally relevant in circumstances where positioning of subcellular structures depends on dynein-mediated cytoplasmic pulling, for example the centration of sperm asters in the large eggs of amphibians and sea urchins [6,71–73].
Worm strains used in this study are listed in S1 Table. Worms were maintained at 16, 20 or 25°C on standard NGM plates seeded with OP50 bacteria. A Mos1 transposon-based strategy (MosSCI) was used to generate strains stably expressing EBP-2::mKate2 and mKate2::EBP-1 [74]. Transgenes were cloned into pCFJ151 for insertion on chromosome II (ttTi5605 locus), and transgene integration was confirmed by PCR. The following alleles were generated by marker-free CRISPR-Cas9-based genome editing, as described previously [75,76]: gfp::p50dnc-2, dynein heavy chaindhc-1::gfp, p150dnc-1(F26L), p150dnc-1(G33S), p150dnc-1(G45R), p150dnc-1(exon 4-5-6 fusion), p150dnc-1(Δexon 4 + exon 5–6 fusion), p150dnc-1(Δexon 5 + exon 3–4 fusion), p150dnc-1(Δexon 4–5), p150dnc-1 null, α-tubulintba-1(Y454A), α-tubulintba-2(Y448A), CLIP-170clip-1 null, and p150dnc-1::3xflag. Genomic sequences targeted by sgRNAs are listed in S2 Table. The modifications were confirmed by sequencing and strains were outcrossed 6 times with the wild-type N2 strain. Other fluorescent markers were subsequently introduced by mating. The p150dnc-1(G33S) allele and the p150dnc-1 null allele were maintained using the GFP-marked genetic balancer nT1 [qIs51]. Homozygous F1 progeny from balanced heterozygous mothers were identified by the absence of GFP fluorescence. None of the homozygous F1 p150dnc-1 null progeny reached adulthood, and homozygous F2 p150dnc-1(G33S) progeny died during embryogenesis.
For production of double-stranded RNA (dsRNA), oligos with tails containing T3 and T7 promoters were used to amplify regions from N2 genomic DNA or cDNA. Primers used for dsRNA production are listed in S3 Table. PCR reactions were cleaned (NucleoSpin Clean-up, Macherey-Nagel) and used as templates for T3 and T7 transcription reactions (MEGAscript, Invitrogen). Transcription reactions were cleaned (NucleoSpin RNA Clean-up, Macherey-Nagel) and complementary single-stranded RNAs were annealed in soaking buffer (3x soaking buffer is 32.7 mM Na2HPO4, 16.5 mM KH2PO4, 6.3 mM NaCl, 14.1 mM NH4Cl). dsRNAs were delivered by injecting L4 hermaphrodites, and animals were processed for live imaging after incubation at 20°C for 24 h or 48 h for partial and penetrant depletions, respectively.
An affinity-purified rabbit polyclonal antibody against the N-terminal region of dynein intermediate chainDYCI-1 (residues 1–177) was generated as described previously [77]. In brief, a GST fusion was expressed in E. coli, purified, and injected into rabbits. Serum was affinity purified on a HiTrap N-hydroxysuccinimide column (GE Healthcare) against covalently coupled DYCI-11−177. Antibodies against p150DNC-1 (GC2) and p50DNC-2 (GC5) were described previously [78].
For immunofluorescence of C. elegans embryos, 10–12 adult worms were dissected into 3 μL of M9 buffer (86 mM NaCl, 42 mM Na2HPO4, 22 mM KH2PO4, 1 mM MgSO4) on a poly-L-lysine-coated slide. A 13 mm2 round coverslip was placed on the 3 μl drop, and slides were plunged into liquid nitrogen. After rapid removal of the coverslip ("freeze-cracking"), embryos were fixed in −20°C methanol for 20 min. Embryos were re-hydrated for 2 x 5 min in PBS (137 mM NaCl, 2.7 mM KCl, 8.1 mM Na2HPO4, and 1.47 mM KH2PO4), blocked with AbDil (PBS with 2% BSA, 0.1% Triton X-100) in a humid chamber at room temperature for 30 minutes, and incubated with primary antibodies [mouse monoclonal anti-α-tubulin DM1A (1:1000) and rat monoclonal anti-tyrosinated α-tubulin YL1/2 (1:500)] for 2 h at room temperature. After washing for 4 x 5 min in PBS, embryos were incubated with secondary antibodies conjugated with fluorescent dyes [Alexa Fluor 488 goat anti-rat IgG (1:1000) and Alexa Fluor 568 goat anti-mouse IgG (1:1000); Life Technologies—Molecular Probes] for 1h at room temperature. Embryos were washed for 4 x 5 min in PBS and mounted in Prolong Gold with DAPI stain (Invitrogen).
Images were recorded on an inverted Zeiss Axio Observer microscope at 1 x 1 binning with a 100x NA 1.46 Plan-Apochromat objective and an Orca Flash 4.0 camera (Hamamatsu). Image files were imported into Fiji for further processing.
For each condition, 100 worms were collected into 1 mL M9 buffer and washed 3 x with M9 buffer and once with M9 / 0.05% Triton X-100. To 100 μL of worm suspension, 33 μL 4x SDS-PAGE sample buffer [250 mM Tris-HCl, pH 6.8, 30% (v/v) glycerol, 8% (w/v) SDS, 200 mM DTT and 0.04% (w/v) bromophenol blue] and ~20 μL of glass beads were added. Samples were incubated for 3 min at 95°C and vortexed for 2 x 5 min. After centrifugation at 20000 x g for 1 min at room temperature, supernatants were collected. Proteins were resolved by 7.5% or 10% SDS-PAGE and transferred to 0.2 μm nitrocellulose membranes (Hybond ECL, Amersham Pharmacia Biotech). Membranes were rinsed 3 x with TBS (50 mM Tris-HCl, pH 7.6, 145 mM NaCl), blocked with 5% non-fat dry milk in TBST (TBS / 0.1% Tween 20) and probed at 4°C overnight with the following primary antibodies: mouse monoclonal anti-FLAG M2 (Sigma, 1:1000), mouse monoclonal anti-α-tubulin B512 (Sigma, 1:5000), rat monoclonal anti-tyrosinated α-tubulin YL1/2 (Bio-Rad Laboratories, 1:5000), rabbit polyclonal anti-DYCI-1 (GC1, 1:1000), rabbit polyclonal anti-DNC-1 (GC2, 1:1000), and rabbit polyclonal anti-DNC-2 (GC5, 1:5000). Membranes were washed 5 x with TBST, incubated with goat secondary antibodies coupled to HRP (JacksonImmunoResearch, 1:5000) for 1 hour at room temperature, and washed again 3 x with TBST. Proteins were detected by chemiluminescence using Pierce ECL Western Blotting Substrate (Thermo Scientific) and X-ray film (Fuji).
Total RNA was isolated from adult hermaphrodites using the TRIzol Plus RNA Purification Kit (Invitrogen). After 3 washes with M9, pelleted worms were homogenized in 200 μL of TRIzol reagent with a pellet pestle homogenizer and incubated at room temperature for 5 min. After addition of 40 μL chloroform, samples were shaken vigorously by hand, incubated at room temperature for 3 min, and centrifuged at 12000 x g for 15 min at 4°C. The upper phase containing the RNA was transferred to an RNase-free tube and an equal volume of 70% ethanol was added. Further RNA purification steps were performed according to the manufacturer's instructions. Purified RNA was treated with DNase I (Thermo Scientific), and cDNA was synthesized with the iScript Select cDNA Synthesis Kit (Bio-Rad Laboratories). The following oligos were used for the PCR reactions in S3B Fig: forward oligo on p150dnc-1 exon 3 (GAATGTCACCTGCTGCTT); forward oligo on p150dnc-1 exon 4 (AAAGCGGTCTACAACTCC); reverse oligo on p150dnc-1 exon 5 (GATTGCGATAAGTTGGAGA); reverse oligo on p150dnc-1 exon 6 (AGTAGTCGTGGACGCTTT). For the SL1 PCR shown in S3D Fig, the following oligos were used: forward oligo on SL1 (GGTTTAATTACCCAAGTTTGA); reverse oligo on p150dnc-1 exon 6 (TCCAGTATCATCAATCTTCTT).
Embryonic viability tests were performed at 20°C. L4 hermaphrodites were grown on NGM plates with OP50 bacteria for 40 h at 20°C, then singled-out to mating plates (NGM plates with a small amount of OP50 bacteria). After 8 h, mothers were removed and the number of hatched and unhatched embryos on each plate was determined 16 h later.
Gravid hermaphrodite worms were dissected in a watch glass filled with Egg Salts medium (118mM KCl, 3.4 mM MgCl2, 3.4 mM CaCl2, 5 mM HEPES, pH 7.4), and embryos were mounted onto a fresh 2% agarose pad. Imaging was performed in rooms kept at 20°C. Embryos co-expressing GFP::histone H2B and GFP::γ-tubulin were imaged on an Axio Observer microscope (Zeiss) equipped with an Orca Flash 4.0 camera (Hamamatsu), a Colibri.2 light source, and controlled by ZEN software (Zeiss). Embryos expressing GFP::p50DNC-2, dynein heavy chainDHC-1::GFP, EBP-2::mKate2, and mCherry::RAB-5 were imaged on a Nikon Eclipse Ti microscope coupled to an Andor Revolution XD spinning disk confocal system composed of an iXon Ultra 897 CCD camera (Andor Technology), a solid-state laser combiner (ALC-UVP 350i, Andor Technology), and a CSU-X1 confocal scanner (Yokogawa Electric Corporation), controlled by Andor IQ3 software (Andor Technology).
All imaging was performed in one-cell embryos unless otherwise indicated. Image analysis was performed using Fiji software (Image J version 2.0.0-rc-56/1.51h).
Values in figures and text are reported as mean ± SEM with a 95% confidence interval. Statistical analyses was performed with GraphPad Prism 7.0 software. The type of statistical analysis performed is indicated in the figure legends. Differences were considered significant at P values below 0.05.
|
10.1371/journal.ppat.1006655 | An acquired mechanism of antifungal drug resistance simultaneously enables Candida albicans to escape from intrinsic host defenses | The opportunistic fungal pathogen Candida albicans frequently produces genetically altered variants to adapt to environmental changes and new host niches in the course of its life-long association with the human host. Gain-of-function mutations in zinc cluster transcription factors, which result in the constitutive upregulation of their target genes, are a common cause of acquired resistance to the widely used antifungal drug fluconazole, especially during long-term therapy of oropharyngeal candidiasis. In this study, we investigated if C. albicans also can develop resistance to the antimicrobial peptide histatin 5, which is secreted in the saliva of humans to protect the oral mucosa from pathogenic microbes. As histatin 5 has been shown to be transported out of C. albicans cells by the Flu1 efflux pump, we screened a library of C. albicans strains that contain artificially activated forms of all zinc cluster transcription factors of this fungus for increased FLU1 expression. We found that a hyperactive Mrr1, which confers fluconazole resistance by upregulating the multidrug efflux pump MDR1 and other genes, also causes FLU1 overexpression. Similarly to the artificially activated Mrr1, naturally occurring gain-of-function mutations in this transcription factor also caused FLU1 upregulation and increased histatin 5 resistance. Surprisingly, however, Mrr1-mediated histatin 5 resistance was mainly caused by the upregulation of MDR1 instead of FLU1, revealing a previously unrecognized function of the Mdr1 efflux pump. Fluconazole-resistant clinical C. albicans isolates with different Mrr1 gain-of-function mutations were less efficiently killed by histatin 5, and this phenotype was reverted when MRR1 was deleted. Therefore, antimycotic therapy can promote the evolution of strains that, as a consequence of drug resistance mutations, simultaneously have acquired increased resistance against an innate host defense mechanism and are thereby better adapted to certain host niches.
| The yeast Candida albicans is part of the normal microflora of most healthy persons, but it can also cause symptomatic infections when host defenses are compromised. C. albicans frequently generates genetically altered variants that are better adapted to changes in its environment during colonization and infection. We investigated if C. albicans can evolve resistance to histatin 5 (Hst 5), an antimicrobial peptide that is produced in the saliva of humans and protects the oral cavity against this pathogen. We found that activated forms of the transcription factor Mrr1 reduce the susceptibility of C. albicans to killing by Hst 5, a phenotype that was partially caused by Mrr1-mediated overexpression of the multidrug efflux pump MDR1. Gain-of-function (GOF) mutations in Mrr1 are a frequent cause of resistance to the antifungal drug fluconazole, especially during long-term treatment of oropharyngeal candidiasis in AIDS patients, but they may also reduce the fitness of the fungus in the absence of the drug. Fluconazole-resistant clinical C. albicans isolates containing GOF mutations in Mrr1 displayed enhanced Hst 5 resistance, demonstrating that antimycotic therapy can promote the evolution of strains that simultaneously have acquired increased resistance against an innate host defense mechanism and are thereby better adapted to specific host niches.
| The yeast Candida albicans is a member of the microbiota of the oral cavity and the gastrointestinal and genitourinary tracts in most healthy persons. When host defenses are compromised, C. albicans can also cause symptomatic infections, which range from superficial skin or mucosal infections to life-threatening, disseminated infections. During both colonization and infection, C. albicans must adapt to environmental changes and stressful conditions encountered in its various host niches. To a large extent this is achieved by reversibly regulating gene expression and biochemical activities according to cellular needs [1, 2]. In addition, C. albicans also produces genetically altered variants that are better adapted than the originally colonizing strain to long-lasting changes in its habitat or a new host niche [3–5]. The generation of such variants is facilitated by the high genomic plasticity of this diploid fungus, which often leads to the amplification or loss of partial or whole chromosomes, especially in response to stress [6]. The increased or decreased copy number of genes that are located on the affected chromosomes of the resulting aneuploid cells may confer a selective advantage under certain adverse conditions [7, 8]. Aneuploidies are unstable, and cells can revert to the normal diploid state by chromosome loss or reduplication in the absence of selection pressure [9]. Genetic variants may also arise by simple point mutations that enable the cells to better tolerate harmful conditions. The acquisition of advantageous mutations is frequently followed by loss of heterozygosity for the mutated allele, which can occur by mitotic recombination or loss of the chromosome containing the wild-type allele [10–18]. These events are promoted in a stressful environment and further enhance the effect of the mutations [19, 20].
A well-documented example of such microevolution within the host is the development of antifungal drug resistance during therapy [21]. Infections by C. albicans are commonly treated with fluconazole, which inhibits the biosynthesis of ergosterol, the main sterol in fungal cell membranes. Mutations in the drug target enzyme sterol 14α-demethylase, encoded by ERG11, result in amino acid exchanges that reduce the affinity of the enzyme for the drug [22]. Similarly, mutations in FKS1, encoding β-1,3-glucan synthase, the target of echinocandin drugs, also result in reduced drug binding [23]. Other point mutations that cause increased drug resistance affect transcription factors and result in permanently changed gene expression programs. The transcription factor Upc2 regulates the expression of ERG11 and other ergosterol biosynthesis genes [24, 25]. Gain-of-function (GOF) mutations in Upc2 that result in hyperactivity of the transcription factor cause constitutive upregulation of its target genes and increased fluconazole resistance [15, 26–28]. Similarly, GOF mutations in the transcription factors Mrr1 and Tac1, which regulate the expression of the multidrug efflux pumps MDR1 and CDR1/CDR2, respectively, result in constitutive overexpression of their target genes and are responsible for fluconazole resistance in many clinical C. albicans isolates [11–13, 16, 29–31].
Mrr1, Tac1, and Upc2 belong to the zinc cluster transcription factor family, which is unique to the fungal kingdom and characterized by a well-conserved DNA-binding motif containing six cysteine residues that coordinate two zinc atoms [32]. C. albicans possesses 82 predicted zinc cluster transcription factors, which are involved in the regulation of diverse cellular processes, although the functions of many of them have not yet been studied in detail [33, 34]. It is conceivable that GOF mutations like those found in Mrr1, Tac1, and Upc2 may also occur in other members of the family and confer new phenotypes that are advantageous under adverse conditions encountered in some host niches. As many transcription factors are activated in response to specific signals and are often not active under standard growth conditions, deletion of the corresponding genes does not necessarily result in an obvious phenotype when the conditions in which they are required are not known. This was the case for Mrr1 and Tac1, whose ability to confer drug resistance was only uncovered by the identification of GOF mutations in fluconazole-resistant clinical isolates [16, 30]. The availability of hyperactive alleles of zinc cluster transcription factors may therefore reveal their biological function and also predict the potential of C. albicans to generate variants with novel phenotypes by acquiring GOF mutations in these transcriptional regulators. Since it cannot be generally predicted which mutations would render a wild-type transcription factor hyperactive, we recently established a method for the artificial activation of zinc cluster proteins by C-terminal fusion with the heterologous Gal4 activation domain and generated a library of C. albicans strains expressing all zinc cluster transcription factors of this fungus in a potentially hyperactive form [34]. Screening of this library showed that one of the artificially activated transcription factors, which was thereafter termed Mrr2, conferred fluconazole resistance by upregulation of the major C. albicans multidrug efflux pump CDR1 [34]. Based on these findings, other investigators searched for naturally occurring gain-of-function mutations in the MRR2 gene in a collection of fluconazole-resistant C. albicans isolates. Indeed, three epidemiologically related isolates with elevated CDR1 expression levels contained a mutated MRR2 allele that caused CDR1 overexpression and increased fluconazole resistance when introduced into a drug-susceptible strain, demonstrating the clinical relevance of the predicted resistance mechanism [35].
In addition to antifungal drugs, which are introduced by medical treatment, C. albicans encounters many other harmful molecules within its host, which may be taken up with the diet, generated by other members of the microbiota, or produced by the host as a defense mechanism against invading pathogens. Humans secrete saliva containing different antimicrobial peptides, including histatins, in order to protect the oral mucosa from pathogenic microbes. Histatins have strong antifungal activity, with histatin 5 (Hst 5) exhibiting the most potent fungicidal activity against C. albicans and other Candida species [36]. Hst 5 is an unusual antimicrobial peptide in that it is not membrane-lytic but rather acts intracellularly to cause cell death [36]. C. albicans possesses several mechanisms to evade killing by Hst 5 and thereby can tolerate the presence of low levels of this antimicrobial peptide. The extracellular glycodomain of the plasma membrane protein Msb2 is shed into the environment and binds Hst 5 as well as other antimicrobial peptides, thereby protecting C. albicans from their action, and extracellular Hst 5 is also proteolytically degraded by secreted aspartic proteases of the fungus [37–39]. Furthermore, C. albicans can recover from stresses generated by intracellular Hst 5 by mechanisms that are mediated by MAP kinases [40]. As Hst 5 acts within the cells, preventing its intracellular accumulation is another potential resistance mechanism. It was recently shown that the Flu1 efflux pump transports Hst 5 out of the cells and that mutants lacking FLU1 are hypersusceptible to killing by Hst 5 [41]. We reasoned that FLU1 expression, like that of the drug efflux pumps MDR1, CDR1, and CDR2, might also be regulated by a zinc cluster protein and that C. albicans might acquire Hst 5 resistance by GOF mutations in this transcription factor. We, therefore, set out to identify transcription factors that regulate FLU1 expression and investigate whether C. albicans can develop increased Hst 5 resistance by this mechanism.
If FLU1, like other known C. albicans efflux pumps, is regulated by a zinc cluster protein, a hyperactive form of this transcription factor should cause constitutive FLU1 overexpression and, consequently, increased Hst 5 resistance. flu1∆ mutants not only exhibit increased susceptibility to Hst 5 but are also hypersensitive to mycophenolic acid (MPA), a phenotype that is easily recognizable on agar plates containing a suitable concentration of the drug [42]. We therefore screened our library of C. albicans strains expressing artificially activated forms of all zinc cluster transcription factors for increased MPA resistance. In a preliminary test, the strains were directly transferred from our stock collection in microtiter plates onto agar plates with and without MPA. Candidate strains that grew better than the wild-type control in the presence of the inhibitor were then retested in a more sensitive dilution spot assay. This resulted in the identification of four hyperactive transcription factors (Mrr1, Mrr2, War1, and Zcf35) that caused clearly increased MPA resistance (Fig 1A).
MPA resistance might be brought about by different mechanisms, for example by upregulation of the IMH3 gene, which encodes inosine monophosphate dehydrogenase, the target enzyme of MPA [43]. To investigate if the increased MPA resistance of our strains was caused by FLU1 overexpression, we introduced the artificially activated transcription factors into reporter strains containing GFP under the control of the FLU1 promoter. As can be seen in Fig 1B, basal FLU1 expression levels were detectable with GFP as a reporter gene, because the fluorescence of the wild-type reporter strains was well above background fluorescence. Intriguingly, FLU1 promoter activity was increased by all four hyperactive transcription factors (3- to 6-fold), indicating that Mrr1, Mrr2, War1, and Zcf35 regulate FLU1 expression and that their activation results in overexpression of this efflux pump.
We next tested if the increased MPA resistance conferred by hyperactive forms of Mrr1, Mrr2, War1, and Zcf35 depended on Flu1. For this purpose, we generated two independent series of flu1∆ mutants and complemented strains of the wild-type strain SC5314. In accord with a previous report [42], homozygous flu1∆ mutants were hypersusceptible to MPA, and reintroduction of a functional FLU1 copy into these mutants increased their MPA resistance to the level observed for heterozygous mutants (S1 Fig). The genes encoding the hyperactive transcription factors were then introduced into the homozygous flu1∆ mutants and the MPA susceptibilities of the resulting strains compared with those of the corresponding wild-type strains. Fig 2 shows that none of the hyperactive transcription factors caused a noticeable increase in MPA resistance in the absence of Flu1, indicating that this effect on the phenotype of the cells was entirely Flu1-dependent.
Although FLU1 overexpression in the heterologous host Saccharomyces cerevisiae had previously been found to result in elevated fluconazole resistance (hence the name given to the gene), deletion of FLU1 in a C. albicans laboratory strain had little effect on fluconazole susceptibility [42]. In line with these findings, we did not observe a detectable increase in fluconazole susceptibility of the flu1∆ mutants derived from the wild-type strain SC5314, neither in dilution spot assays on agar plates nor when we determined the minimal inhibitory concentration of fluconazole in a broth microdilution assay (Fig 2). Hyperactive forms of Mrr1, Mrr2, and Zcf35 cause increased fluconazole resistance [34]. For Mrr1 and Mrr2, this was also evident in the dilution spot assays on agar plates containing a defined concentration of fluconazole, whereas the strains with the hyperactive Zcf35 showed even slightly reduced growth on the fluconazole plates, despite the elevated MIC for these strains (Fig 2). The strains containing the hyperactive War1 also showed slightly reduced growth on the fluconazole agar plates, but exhibited increased fluconazole resistance when tested in the MIC assays, a phenotype that was not observed for these strains in our previous study, presumably because a different assay medium was used. The increased fluconazole resistance conferred by the four hyperactive transcription factors was also observed when they were expressed in flu1∆ mutants (Fig 2), demonstrating that FLU1 overexpression did not contribute to the fluconazole-resistant phenotype.
Although FLU1 mediates Hst 5 resistance, its expression is not induced by Hst 5 [41]. Therefore, we investigated if basal FLU1 expression levels depend on any of the identified transcription factors. For this purpose, the PFLU1-GFP reporter fusion was introduced into mrr1∆, mrr2∆, war1∆, and zcf35∆ mutants. S2 Fig shows that all mutants displayed wild-type FLU1 promoter activity, indicating that none of these transcription factors is required for basal FLU1 expression.
As hyperactive forms of Mrr1, Mrr2, War1, and Zcf35 cause FLU1 upregulation, we reasoned that they should also mediate increased resistance to Hst 5. Therefore, we compared the percent killing of the wild-type parental strain SC5314 and derivatives containing the artificially activated transcription factors after 60 min of incubation in the presence of various Hst 5 concentrations (Fig 3). The hyperactive Mrr1 indeed conferred increased resistance to the antimicrobial peptide; in the presence of 30 μM Hst 5, killing was reduced from 80% for the wild type to ca. 54% for strains containing the artificially activated Mrr1. In contrast, strains with the hyperactive Mrr2 and War1 were as efficiently killed by Hst 5 as the wild type, and the strains with the hyperactive Zcf35 showed even enhanced sensitivity. Therefore, despite the increased FLU1 expression levels, artificially activated Mrr2, War1, and Zcf35 were unable to confer Hst 5 resistance. We speculated that these hyperactive transcription factors have additional effects on the cells that increase their susceptibility to Hst 5 and thereby abrogate any advantage conferred by FLU1 overexpression. However, subsequent experiments provided a different explanation for this result (see below).
Our finding that an artificially activated form of Mrr1 causes increased Hst 5 resistance was particularly intriguing, since many fluconazole-resistant clinical C. albicans isolates contain GOF mutations in Mrr1 [13, 16, 44]. Such mutations are selected under antifungal therapy, because the hyperactive transcription factor mediates overexpression of the multidrug efflux pump MDR1 and other genes, and thereby confers fluconazole resistance. We investigated if such naturally occurring Mrr1 GOF mutations also cause FLU1 overexpression and enhanced Hst 5 resistance. Mrr1 GOF mutations have a much stronger effect on MDR1 expression and fluconazole resistance after loss of heterozygosity for the mutated MRR1 allele [16, 45], and we reasoned that the same would be true for any effect on FLU1 expression and Hst 5 resistance. To assess the effect of such mutations in an isogenic background, we replaced both endogenous MRR1 alleles of strain SC5314 by alleles with GOF mutations that were originally discovered in fluconazole-resistant clinical isolates and resulted in the amino acid exchanges P683S, G997V, G878E, Q350L, N803D, T360I, K335N, and T896I. The clinical isolate 6692 contains different GOF mutations (T360I and K335N) in its two MRR1 alleles [13]. To reproduce this scenario, we also constructed strains containing a combination of these hyperactive alleles. In addition, strains in which the endogenous MRR1 alleles were replaced in the same fashion by an unmutated wild-type allele were included in the experiments to make sure that the sequential allele replacement strategy alone did not have a phenotypic effect.
We first tested the fluconazole susceptibilities of the strains (Table 1). As previously reported [20], replacement of the endogenous MRR1 alleles of strain SC5314 by a nonmutated MRR1 copy did not affect drug susceptibility, whereas the P683S mutation caused a 32-fold increase in the MIC of fluconazole (from 0.5 μg/ml to 16 μg/ml). A similar effect was observed for all other GOF mutations, which raised the MIC to 16 or 32 μg/ml. In each case, identical results were obtained for two independently constructed strains. We then introduced the PFLU1-GFP reporter fusion into the strains with the different MRR1 GOF mutations to determine their effect on FLU1 expression. Fig 4A shows that all hyperactive MRR1 alleles caused elevated FLU1 expression levels, albeit to various degrees. The strongest upregulation (ca. 7-fold) was mediated by the N803D and T360I mutations, and the weakest effect (2- to 3-fold upregulation) was observed for the P683S and T896I mutations. The differences were reproducible, because in all cases the two independently constructed reporter strains exhibited similar fluorescence values. These results demonstrated that the overexpression of FLU1 by a hyperactive Mrr1 is not a peculiar effect of the artificially generated Mrr1-GAD fusion but a general characteristic of naturally occurring activated forms of Mrr1.
We selected six different Mrr1 GOF mutations (G997V, G878E, Q350L, N803D, T360I, T896I) to investigate if they also resulted in increased Hst 5 resistance. As can be seen in Fig 4B, all strains containing hyperactive MRR1 alleles were less efficiently killed by Hst 5 than the parental wild-type strain SC5314 (killing decreased from ca. 90% to ca. 60%), demonstrating that naturally occurring Mrr1 GOF mutations conferred increased Hst 5 resistance. In a further control experiment we confirmed that the strains in which the endogenous MRR1 alleles were replaced by a nonmutated wild-type copy exhibited wild-type Hst 5 susceptibility, as expected (S3 Fig).
Despite the fact that hyperactive forms of Mrr2, War1, and Zcf35 caused even stronger FLU1 upregulation than the artificially activated Mrr1 (see Fig 1B), only the latter enhanced the resistance of the cells to Hst 5 (Fig 3). We therefore investigated if the increased Hst 5 resistance mediated by a hyperactive Mrr1 was indeed caused by FLU1 overexpression. For this purpose, we compared the percent killing by Hst 5 of wild-type and flu1∆ cells with and without the artificially activated MRR1 allele (Fig 5). Unexpectedly, deletion of FLU1 in the prototrophic wild-type strain SC5314 did not result in hypersensitivity to Hst 5, in contrast to previous observations with the auxotrophic laboratory strain CAF4-2 [41]. Even more surprisingly, the hyperactive Mrr1 reduced the susceptibility of the cells to killing by Hst 5 both in the presence and absence of FLU1 with comparable efficiency (at 30 μM Hst 5, the flu1∆ mutants were even slightly more resistant in this experiment). Therefore, Mrr1-mediated Hst 5 resistance must be caused by other mechanisms, which might also mask any additional contribution of FLU1 overexpression in this strain background.
As hyperactive forms of Mrr1 cause overexpression of many genes [16, 45], we searched the known Mrr1 target genes for candidates that might promote Hst 5 resistance. TPO2 encodes another putative efflux pump that is closely related to Flu1 [41], and it is bound and upregulated by a hyperactive Mrr1 [45]. Mdr1 also has high similarity to Flu1 [42], and MDR1 is one of the most strongly upregulated genes in strains with Mrr1 GOF mutations [16, 45]. Although no role of TPO2 and MDR1 in Hst 5 resistance had been found in a previous study [41], it seemed possible that these Mrr1 target genes could mediate Hst 5 resistance when they are overexpressed. Northern hybridization experiments showed that both MDR1 and TPO2 are expressed only at low levels in the wild-type strain SC5314 (Fig 6, left lanes). Interestingly, only Mrr1, but not the other artificially activated transcription factors, Mrr2, War1, and Zcf35, caused upregulation of MDR1 and, more weakly, TPO2 (Fig 6A), providing a potential explanation why the latter did not cause Hst 5 resistance. Most naturally occurring GOF mutations in Mrr1 had an even stronger effect on MDR1 and TPO2 expression than the artificially activated Mrr1, although expression levels strongly depended on the specific allele (Fig 6B).
As FLU1, MDR1, and TPO2 all are upregulated by a hyperactive Mrr1, it was conceivable that each of these three related transporters contributes to Mrr1-mediated Hst 5 resistance and the importance of a single efflux pump would not be easily detectable. We therefore constructed flu1∆ mdr1∆ tpo2∆ triple mutants and introduced the G997V GOF mutation, which caused an efficient upregulation of all three genes (see Fig 4A and Fig 6B), into both MRR1 alleles of the mutants. The deletion of the transporters in the wild-type parental strain, in which they were not or only weakly expressed, did not result in hypersusceptibility to Hst 5 (S4 Fig). However, the increased Hst 5 resistance conferred by the hyperactive Mrr1 was reduced, albeit not abolished, in the triple mutants, indicating that one or more of the transporters, but also additional Mrr1 target genes, promote Hst 5 resistance (Fig 7A). To assess the relative contribution of each individual transporter, we introduced the G997V mutation also into both MRR1 alleles of flu1∆, mdr1∆, and tpo2∆ single mutants. Deletion of FLU1 or TPO2 did not or only slightly reduce the Hst 5 resistance of cells containing the MRR1G997V mutation (Fig 7B and 7D). In contrast, deletion of MDR1 increased the Hst 5 susceptibility of cells with the hyperactive Mrr1 (killing increased from ca. 60% to ca. 80% in the presence of 30 μM Hst 5), although the cells were still more resistant than the wild type (ca. 90% killing) (Fig 7C). These results demonstrated that overexpression of MDR1 is one mechanism that contributes to the increased Hst 5 resistance of cells with Mrr1 GOF mutations.
We next investigated if fluconazole-resistant clinical C. albicans isolates that have acquired MRR1 GOF mutations during antimycotic therapy also exhibit increased Hst 5 resistance. For this purpose, we selected five matched pairs of fluconazole-susceptible and -resistant isolates from AIDS patients with different MRR1 GOF mutations. We had previously generated mrr1∆ mutants from the five resistant isolates [13, 16], which allowed us to assess the contribution of Mrr1 to their phenotypes. To compare FLU1 expression levels, we introduced the PFLU1-GFP reporter fusion into all these strains. Two independent reporter strains were derived from each clinical isolate and one from both independently constructed mrr1∆ mutants of each resistant isolate to ensure the reproducibility of the results (one of the mrr1∆ mutants of isolate 2271 exhibited a high unspecific autofluorescence and could not be used for these experiments). As can be seen in S5 Fig, FLU1 expression was moderately increased in all five resistant isolates compared to the matched susceptible isolates (2- to 5-fold), and the elevated expression levels returned to those observed in the susceptible isolates when MRR1 was deleted. These and previous results [13, 16] demonstrate that Mrr1 GOF mutations cause concomitant upregulation of both efflux pumps, MDR1 and FLU1, in fluconazole-resistant clinical C. albicans isolates.
We then assessed if the Mrr1 GOF mutations also resulted in enhanced Hst 5 resistance of the clinical isolates. Fig 8 shows that killing by Hst 5 was reduced in all five fluconazole-resistant isolates with Mrr1 GOF mutations compared to their matched susceptible isolates, but to different degrees. A relatively minor but significant reduction in killing (from 94% to 84%) was observed for isolate B4 compared to isolate B3 (Fig 8B), whereas isolate DSY2286 was highly resistant, as hardly any killing of this isolate was observed at the tested Hst 5 concentration (Fig 8C). In four of the five cases, Hst 5 resistance was fully or largely mediated by the hyperactive Mrr1, because the percent killing was elevated to the levels of the matched susceptible isolates when MRR1 was deleted. Isolate 6692 was an exception, because the mrr1∆ mutants derived from it retained the increased Hst 5 resistance. This strain apparently possesses other mechanisms of Hst 5 resistance that override the contribution of the hyperactive Mrr1. Collectively, these results demonstrate that fluconazole-resistant C. albicans strains with Mrr1 GOF mutations have simultaneously acquired increased Hst 5 resistance, because the overexpression of MDR1 and other Mrr1 target genes not only mediates fluconazole resistance but also increased resistance to the antimicrobial peptide.
The acquisition of GOF mutations in zinc cluster transcription factors is a frequent cause of fluconazole resistance in clinical C. albicans strains. GOF mutations in Mrr1, Tac1, and Upc2 enable the cells to continue to grow in the presence of the drug and outcompete wild-type cells in the population. However, the hyperactive transcription factors also decrease the fitness of the cells in the absence of drug selection, at least under some environmental conditions [20, 44, 46]. As Mrr1, Tac1, and Upc2 regulate different sets of genes [31, 45, 47], the fitness defect must have a different basis in each case; consequently, it is aggravated in highly resistant strains containing several hyperactive transcription factors [20]. The deregulated gene expression probably causes an unnecessary waste of energy and reduces the ability of the cells to appropriately adapt to specific host niches. Here, we have uncovered a previously unrecognized effect of hyperactive forms of Mrr1 in addition to their ability to mediate fluconazole resistance. The overexpression of the multidrug efflux pump MDR1 (and possibly to some extent also FLU1) and other Mrr1 target genes confers increased resistance to the antifungal peptide Hst 5, which is present in the saliva of humans. Fluconazole-resistant strains with GOF mutations in Mrr1 have mainly been isolated from the oral cavity of HIV-infected patients suffering from oropharyngeal candidiasis [13, 14, 16, 42, 44, 48–55]. Increased Hst 5 resistance may therefore represent an additional advantage that counterbalances the fitness costs of Mrr1 hyperactivity and helps the mutants to establish themselves in this host niche. Nevertheless, Mrr1 mutations have been found only after fluconazole treatment of oropharyngeal candidiasis and have not been detected in pretreatment isolates from the same patients. This suggests that exposure to Hst 5 is not sufficient to select for Mrr1 GOF mutations when C. albicans colonizes the oral cavity as a harmless commensal without causing disease symptoms and does not have to cope with antimycotic therapy.
It is difficult to estimate how strongly a hyperactive Mrr1 contributes to the survival and competitive fitness of C. albicans in the oral cavity of humans. After short-term exposure to Hst 5, as in our in vitro experiments, strains with a hyperactive Mrr1 exhibited significantly reduced killing compared to wild-type controls. In the natural habitat, Hst 5 is constantly produced, resulting in continued exposure of C. albicans to the antifungal peptide, and the ability to avoid its toxic effects may present an even more significant advantage. The fitness gain due to enhanced Hst 5 resistance will likely also depend on the Hst 5 concentration, which varies among individuals and is generally lower in HIV-infected persons [56]. As Hst 5 is produced only by humans and other primates, the relevance of increased Hst 5 resistance cannot be meaningfully studied in mouse models of experimental candidiasis.
Although we initially searched for regulators of FLU1 expression in order to identify transcription factors that might confer Hst 5 resistance, our subsequent experiments revealed that Flu1 contributed little, if at all, to Hst 5 resistance in derivatives of strain SC5314 carrying hyperactive forms of MRR1. Instead, we unexpectedly found that overexpression of the efflux pump MDR1 was one mechanism by which such strains became less susceptible to killing by the antimicrobial peptide. MDR1 overexpression, therefore, confers increased resistance not only to fluconazole but also to Hst 5, which is a previously unrecognized function of this efflux pump that is of potential clinical relevance. It is well possible that FLU1 makes a more important contribution in some clinical isolates with MRR1 GOF mutations, because all such isolates that were tested concomitantly overexpressed both MDR1 and FLU1, and some of them may contain FLU1 alleles that encode more efficient Hst 5 transporters. Allelic differences that affect the efficiency of drug transport have been previously reported for the drug efflux pump CDR2 [57]. Nevertheless, a hyperactive Mrr1 still conferred increased Hst 5 resistance, albeit less efficiently, even in strains lacking both MDR1 and FLU1 as well as TPO2, encoding a related putative transporter, indicating that additional Mrr1 target genes are involved in this phenotype. This is similar to Mrr1-mediated fluconazole resistance, which is only partially mediated by Mdr1 and involves other Mrr1 target genes that remain to be identified [45]. The effect of specific GOF mutations on the expression of individual Mrr1 target genes varies. For example, the T360I and T896I mutations caused comparable MDR1 expression levels, but only the T360I mutation also resulted in strong TPO2 overexpression (see Fig 6B). Therefore, the degree of resistance to fluconazole and Hst 5 conferred by a particular Mrr1 GOF mutation will depend not only on the expression levels of MDR1 but also on those of the other target genes that contribute to these phenotypes.
The successful establishment and expansion of genetically altered variants within a population require that the beneficial effects of a mutation outweigh its negative consequences. As noted above, the increased Hst 5 resistance may not sufficiently offset the fitness costs caused by a hyperactive Mrr1 to select for MRR1 GOF mutations in strains colonizing the oral cavity of healthy persons. In contrast, the increased fluconazole resistance of such strains supports their emergence in the presence of the drug. It is intriguing that antimycotic therapy promotes the evolution of strains which, as a consequence of a drug resistance mutation, have acquired increased resistance to an innate host defense mechanism and are better adapted to some stressful conditions encountered in the host even in the absence of antifungal drug treatment.
Candida albicans strains used in this study are listed in S1 Table. The complete library of C. albicans strains containing artificially activated zinc cluster transcription factors has been described previously [34]. All strains were stored as frozen stocks with 17.2% glycerol at -80°C and subcultured on YPD agar plates (10 g yeast extract, 20 g peptone, 20 g glucose, 15 g agar per litre) at 30°C. Strains were routinely grown in YPD liquid medium at 30°C in a shaking incubator. For selection of nourseothricin-resistant transformants, 200 μg/ml nourseothricin (Werner Bioagents, Jena, Germany) was added to YPD agar plates. To obtain nourseothricin-sensitive derivatives in which the SAT1 flipper cassette was excised by FLP-mediated recombination, transformants were grown overnight in YCB-BSA-YE medium (23.4 g yeast carbon base, 4 g bovine serum albumin, 2 g yeast extract per litre, pH 4.0) without selective pressure to induce the SAP2 promoter controlling caFLP expression. Alternatively, strains containing a SAT1 flipper cassette in which the caFLP gene is expressed from the MAL2 promoter (as in plasmids pMRR1R4 to pMRR1R10) were grown overnight in YPM medium (10 g yeast extract, 20 g peptone, 20 g maltose per liter) instead of YCB-BSA-YE to induce the MAL2 promoter. Appropriate dilutions of the cultures were plated on YPD agar plates and grown for 2 days at 30°C to obtain single colonies. Nourseothricin-sensitive clones were identified by restreaking on YPD plates and on YPD plates containing 100 μg/ml nourseothricin.
A FLU1 deletion construct was obtained by amplification of the FLU1 upstream and downstream regions from genomic DNA of strain SC5314 with the primer pairs FLU1.01/FLU1.02 and FLU1.03/FLU1.04, respectively (all oligonucleotide primers used in this study are listed in S2 Table). The PCR products were digested with SacI/SacII and XhoI/ApaI, respectively, and cloned on both sides of the modified SAT1 flipper cassette contained in plasmid pSFS5 [58] to result in pFLU1M1. To reintroduce an intact FLU1 copy into flu1∆ mutants, the FLU1 coding region plus upstream and downstream sequences was amplified with primers FLU1.01 and FLU1_compleR. The PCR product was digested with SacI/SacII and substituted for the FLU1 upstream region of plasmid pFLU1M1 to generate pFLU1K1. A PFLU1-GFP reporter fusion was constructed as follows. The FLU1 upstream region was amplified with primers FLU1.01 and FLU1rev_upstream, and a GFP-TACT1 fragment was amplified from plasmid pNIM1 [59] with primers FLU1forw_GFP and ACT19. The gel-purified PCR products were then used as templates for a fusion PCR with primers FLU1.01 and ACT19. The PCR product was digested with SacI/SacII and inserted instead of the FLU1 upstream region in plasmid pFLU1M1 to generate pFLU1G1.
A WAR1 deletion construct was generated by amplifying the WAR1 flanking sequences with the primer pairs WAR1.01/WAR1.02 and WAR1.03/WAR1.04, digesting the PCR products with SacI/SacII and XhoI/ApaI, respectively, and inserting the fragments on both sides of the SAT1 flipper cassette of pSFS5, generating pWAR1M1. Similarly, a ZCF35 deletion cassette was obtained by amplifying the ZCF35 flanking sequences with the primer pairs ZCF35-7/ZCF35-8 and ZCF35-9/ZCF35-10 and cloning the SacI/SacII- and XhoI/ApaI-digested PCR products in pSFS5 to generate pZCF35M3. A TPO2 deletion cassette was generated by amplifying the TPO2 flanking sequences with the primer pairs TPO2.01/TPO2.02 and TPO2.03/TPO2.06 and cloning the SacI/SacII- and XhoI/KpnI-digested PCR products in pSFS5 to generate pTPO2M2. A new MDR1 deletion cassette was generated by substituting the SAT1 flipper cassette from pSFS5 for the old SAT1 flipper cassette in the previously described plasmid pMDR1M2 [45], yielding pMDR1M3.
Plasmids pMRR1R2 and pMRR1R3, which contain the wild-type MRR1 gene and a mutated allele with the P683S GOF mutation, respectively, have been previously described [20, 45]. These plasmids also contain the recyclable SAT1 flipper cassette to allow the sequential replacement of both endogenous MRR1 alleles in strain SC5314. To obtain analogous constructs with other MRR1 GOF mutations, the SacI-BglII fragments from plasmids pZCF36K5, pZCF36K12, pZCF36K14, pZCF36K15, pZCF36K17, pZCF36K18, and pZCF36K19, containing G997V, G878E, Q350L, N803D, T360I, K335N, and T896I GOF mutations, respectively, in MRR1 [13], were inserted instead of the wild-type MRR1, generating plasmids pMRR1R4 to pMRR1R10.
C. albicans strains were transformed by electroporation [43] with the following gel-purified linear DNA fragments. The insert from pFLU1M1 was used to sequentially delete the two wild-type alleles of FLU1 in strain SC5314, and the insert from pFLU1K1 was used to reintroduce an intact FLU1 copy into the homozygous mutants. The insert from pFLU1G1 was used to express the GFP reporter gene under the control of the endogenous FLU1 promoter in various strain backgrounds. The cassettes from plasmids pMRR1GAD1, pZCF34GAD1, pWAR1GAD1, and pZCF35GAD1 [34] were used to introduce artificially activated forms of MRR1, MRR2, WAR1, and ZCF35, respectively, into PFLU1-GFP reporter strains and flu1∆ mutants. The deletion cassettes from pMDR1M3, pTPO2M2, pWAR1M1, and pZCF35M3 were used to generate mdr1∆, tpo2∆, war1∆, and zcf35∆ mutants of strain SC5314. The cassettes from pMDR1M3 and pTPO2M2 were also used for the construction of flu1∆ mdr1∆ tpo2∆ triple mutants. The inserts from plasmids pMRR1R4 to pMRR1R10 were used to replace the wild-type MRR1 alleles of strain SC5314 by mutated alleles with different GOF mutations. The insert from pMRR1R4 was also used to introduce the G997V GOF mutation into both MRR1 alleles of flu1∆, mdr1∆, and tpo2∆ single mutants and flu1∆ mdr1∆ tpo2∆ triple mutants. The correct integration of each construct as well as recycling of the SAT1 flipper cassette were confirmed by Southern hybridization using the flanking sequences as probes. Introduction of the MRR1 GOF mutations was confirmed by reamplification of the genes from heterozygous and homozygous mutants and sequencing of the PCR products. In each case, two independent series of strains were generated and used for further analysis.
Genomic DNA from C. albicans strains was isolated as described previously [60]. The DNA was digested with appropriate restriction enzymes, separated on a 1% agarose gel, transferred by vacuum blotting onto a nylon membrane, and fixed by UV crosslinking. Southern hybridization with enhanced chemiluminescence-labeled probes was performed with the Amersham ECL Direct Nucleic Acid Labelling and Detection System (GE Healthcare UK Limited, Little Chalfont Buckinghamshire, UK) according to the instructions of the manufacturer.
Overnight cultures of the strains were diluted 10−2 in fresh YPD medium and grown for 4 h at 30°C. Total RNA was extracted by the hot acidic phenol method [61] combined with a purification step with the RNeasy Mini Kit (Qiagen, Hilden, Germany). RNA samples were separated on a 1.2% agarose gel, transferred by capillary blotting onto a nylon membrane, fixed by UV crosslinking, and hybridized with digoxigenin-labeled MDR1 (positions 647 to 1688 in the MDR1 coding sequence, amplified with primers NB-MDR1_FW and NB-MDR1-RV), TPO2 (positions 115 to 1154 in the TPO2 coding sequence, amplified with primers TPO2_forN3 and TPO3_revN3), and ACT1 (positions 1512 to 1677 in the ACT1 coding sequence, amplified with primers ACT_RT and ACT2_RT) probes. The bound probes were detected with a peroxidase-labeled anti-digoxigenin AP-conjugate (Roche, Basel, Switzerland).
The fluconazole susceptibilities of the strains were determined by a previously described broth microdilution method [62], with slight modifications. A 2-day-old colony from a YPD agar plate was suspended in 2 ml of a 0.9% NaCl solution, and 4 μl of the suspension was mixed with 2 ml 2x SD-CSM medium (13.4 g yeast nitrogen base without amino acids [YNB; BIO 101, Vista, Calif.], 40 g glucose, 1.54 g complete supplement medium [CSM, BIO101]). A twofold dilution series of fluconazole (Sigma GmbH, Deisenhofen, Germany) was prepared in water, starting from an initial concentration of 512 μg/ml. One hundred microliters of each fluconazole solution was then mixed with 100 μl of the cell suspension in a 96-well microtiter plate and the plates were incubated for 48 h at 37°C. The MIC of fluconazole was defined as the drug concentration that abolished or drastically reduced visible growth compared to a drug-free control.
Overnight cultures of the strains in SD-CSM medium were diluted to an optical density at 600 nm of 2.0. Ten-fold dilutions from 100 to 10−5 were prepared in a 96-well microtiter plate and ca. 5 μl of the cell suspensions transferred with a replicator onto SD-CSM agar plates without or with 0.5 or 1.0 μg/ml MPA or 5 μg/ml fluconazole. Plates were incubated for 2 days at 30°C and photographed.
Overnight cultures of the GFP reporter and control strains were diluted 10−2 in fresh YPD medium and grown for 3 h at 30°C. The cultures were tenfold diluted in 1 ml cold phosphate-buffered saline (PBS) and flow cytometry was performed using the MACSQuantAnalyzer (Miltenyi Biotec; Bergisch Gladbach, Germany) equipped with an argon laser emitting at 488 nm. Fluorescence was detected using the B1 fluorescence channel equipped with a 525 nm band-pass filter (bandwidth 50 nm). Twenty thousand cells were analyzed per sample and counted at a flow rate of approx. 500 cells per second. Fluorescence data were collected by using logarithmic amplifiers. The mean fluorescence (arbitrary values) was determined with MACSQuantify (Version 2.4, Miltenyi Biotec) software.
The susceptibility of C. albicans cells to Hst 5 was measured using microdilution plate assays as previously described [63]. Briefly, 10 ml of YPD medium was inoculated with single colonies of each strain. Cells were grown overnight at room temperature. Overnight cultures were diluted to an A600 of 0.3 to 0.4 and then incubated at 30°C with shaking (220 rpm) until an A600 of ∼1.0 was reached. Cells were washed twice with 10 mM sodium phosphate buffer (NaPB), pH 7.4. The cells (1 × 106) were then mixed with different concentrations of Hst 5 at 30°C for 60 min and diluted in 10 mM NaPB. Aliquots of 500 cells were spread onto YPD agar plates and incubated for 24 to 48 h until colonies became visible. The percent killing was calculated as [1 − (number of colonies from Hst 5-treated cells/number of colonies from control cells)] × 100%. Assays were performed in quadruplicate for each strain. The reference strain SC5314 was included in each set of experiments to validate assay conditions.
Since the FLU1 expression and Hst 5 sensitivity assays were performed under controlled conditions in vitro, we assumed that the collected data fit a standard normal distribution. The one-way ANOVA test was used when three or more groups were compared with each other, and the two-tailed t-test was used when only two groups were compared, as indicated in the figures. Values for the two independently constructed strains A and B were combined in each case, except for the experiments shown in Fig 4B, which were performed on separate occasions. All statistical tests were conducted using GraphPad Prism version 7.03 software.
|
10.1371/journal.pntd.0002898 | Deciphering the Origin of the 2012 Cholera Epidemic in Guinea by Integrating Epidemiological and Molecular Analyses | Cholera is typically considered endemic in West Africa, especially in the Republic of Guinea. However, a three-year lull period was observed from 2009 to 2011, before a new epidemic struck the country in 2012, which was officially responsible for 7,350 suspected cases and 133 deaths. To determine whether cholera re-emerged from the aquatic environment or was rather imported due to human migration, a comprehensive epidemiological and molecular survey was conducted. A spatiotemporal analysis of the national case databases established Kaback Island, located off the southern coast of Guinea, as the initial focus of the epidemic in early February. According to the field investigations, the index case was found to be a fisherman who had recently arrived from a coastal district of neighboring Sierra Leone, where a cholera outbreak had recently occurred. MLVA-based genotype mapping of 38 clinical Vibrio cholerae O1 El Tor isolates sampled throughout the epidemic demonstrated a progressive genetic diversification of the strains from a single genotype isolated on Kaback Island in February, which correlated with spatial epidemic spread. Whole-genome sequencing characterized this strain as an “atypical” El Tor variant. Furthermore, genome-wide SNP-based phylogeny analysis grouped the Guinean strain into a new clade of the third wave of the seventh pandemic, distinct from previously analyzed African strains and directly related to a Bangladeshi isolate. Overall, these results highly suggest that the Guinean 2012 epidemic was caused by a V. cholerae clone that was likely imported from Sierra Leone by an infected individual. These results indicate the importance of promoting the cross-border identification and surveillance of mobile and vulnerable populations, including fishermen, to prevent, detect and control future epidemics in the region. Comprehensive epidemiological investigations should be expanded to better understand cholera dynamics and improve disease control strategies throughout the African continent.
| Cholera is a potentially deadly diarrheic disease caused by the toxin-secreting bacterium Vibrio cholerae. In many poor countries, this prototypical waterborne disease is considered endemic and linked to the climate-driven proliferation of environmental reservoirs of the pathogen. Although such a statement implies radical public health consequences, it has never been proven in Africa. The present study aimed to elucidate the origin of the cholera epidemic that struck the Republic of Guinea in 2012 following a three-year lull period. This investigation integrated a spatiotemporal analysis of the national case databases, field investigations and thorough genetic analyses of 38 clinical bacterial isolates sampled throughout the Guinean epidemic. The Guinean V. cholerae DNA sequence results were aligned and compared with the sequences of nearly 200 strains isolated throughout the world over the past 60 years. Overall, these results suggest that the 2012 cholera epidemic strain was likely imported from Sierra Leone to Guinea by traveling fishermen. The emergence of cholera epidemics due to human-driven activity may be widespread throughout Africa. This highlights the importance of transborder collaborative public health strategies targeting highly mobile and high-risk populations. Similar integrated studies should be conducted in other countries impacted by the disease to better understand the spread of recent epidemics and thus better intercept future outbreaks.
| Cholera is generally considered endemic in West Africa [1], especially in countries such as Nigeria, Benin, Togo, Ghana, Liberia and the Republic of Guinea [2]. In 2004–2008, Guinea was struck by a succession of regional cholera outbreaks responsible for 17,638 reported cases and 786 deaths [3]. In 2009, the country established an early cholera alert system including cholera microbiological surveillance to quickly detect emerging epidemics [4]. However, the following years in Guinea were marked by a lull in cholera transmission until new cases were reported between February and April 2012 in several maritime prefectures spanning 200 km [5]. Between April and June, a reactive oral cholera vaccination campaign was implemented by the Guinean Ministry of Health and Médecins Sans Frontières (Doctors Without Borders) in two prefectures, Forecariah and Boffa [6]. However, during the rainy season in July and August, the epidemic exploded in the capital Conakry and then spread to inland areas. By the time the end of the epidemic was declared in December 2012, 7,350 cases and 133 deaths had been officially reported to the World Health Organization (WHO), from 11 out of 34 prefectures [7].
To provide a scientific foundation for the control and prevention of future outbreaks, it is critical to understand the origin of cholera epidemics in coastal areas, which has remained subject to debate. In Peru and Bangladesh, a similar near simultaneous appearance of cholera at different locales along coastal or estuarine areas has been considered a key argument in favor of the “cholera paradigm” [8]. According to this general model for cholera transmission, coastal waters in these regions represent reservoirs of multiclonal epidemic-provoking Vibrio cholerae strains whose growth is directly associated with plankton blooms driven by climatic and environmental conditions [8]–[12]. Conversely, whole-genome-based phylogenetic analyses of Peruvian and other South American isolates from the 1990s have found that the strains form a clonal and independent lineage within the seventh pandemic [13], [14]. Such molecular approaches have recently highlighted the function of human-to-human transmission of the disease [12], which could be the main driver of clonal outbreak diffusion, even along coastal areas.
To assess whether the 2012 Guinean cholera epidemic was caused by local environment-to-human transmission or was rather initiated by the human-driven importation of a single toxigenic clone we used a multidisciplinary approach involving spatiotemporal analyses, field investigations and several complementary V. cholerae genotyping methods.
The Republic of Guinea spans 245,857 km2 and is administratively divided into 33 prefectures plus the capital Conakry. In 2012, the country had an estimated population of 12 million inhabitants. At that time, the Guinean national health surveillance system prospectively reported all suspected cholera cases based on the WHO definition of the disease [15]. Each Prefectural Health Directorate (DPS – Direction Préfectorale de la Santé) tallied new cases recorded at the various health structures of the prefecture on a weekly basis. Aggregated morbidity and mortality cholera data were then transmitted to the Directorate of Prevention and Disease Control (DPLM – Direction de la Prévention et de la Lutte contre la Maladie), which compiled the information in a national database of 7,350 cases. DPLM also retrospectively compiled a line list of 6,568 patients, which included the date of consultation and geographical origin down to the village level and was anonymized prior to analysis. To limit notification bias, both databases were subsequently compared and merged, which enabled the retrieval of 393 additional cases. The use of these data for epidemiological, research and publication purposes was approved by the Guinean Ministry of Health (Ministère de la Santé Publique et de l'Hygiène Publique). Daily-accumulated rainfall data were obtained from satellite estimates (TMPA-RT 3B42RT derived) provided by the National Aeronautics and Space Administration (available at: http://disc2.nascom.nasa.gov/Giovanni/tovas/realtime.3B42RT_daily.2.shtml). As most cases were recorded in Maritime Guinea and, to a lesser extent, Middle Guinea, daily rainfall data were averaged on the position 9.00N-12.00N/15.00-11.75W, which excluded the eastern two-thirds of the country where precipitation levels were lower and much fewer cholera cases were reported. Population estimates for 2012 were obtained from the Guinean Expanded Program for Immunization at both the prefectural and sub-prefectural levels. Their estimates were based on the general population census of 1996 considering prefecture-specific annual population growth rates, which were provided by the Guinean Statistics National Institute (INS – Institut National de la Statistique de Guinée) and ranged from 0.71% to 6.51%.
Field investigations of index cases and local conditions that supported cholera emergence and transmission were prospectively conducted in affected areas throughout the epidemic by epidemiologists of the Guinean Health Ministry and the country team of the African Cholera Surveillance Network (Africhol; http://www.africhol.org) to organize the public health response. They included basic interviews among affected communities identified by the hospital- and community-based surveillance system (including rumors) and followed routine procedures of the Integrated Disease Surveillance and Response System of the Guinean Ministry of Health. Retrospective field investigations were also conducted in August and September 2012 mainly to review the register books of treatment facilities, but also to interview local health authorities and staff regarding the 2012 outbreak as well as to observe ecological, social, water and sanitation conditions in affected areas.
With the support of the Africhol Consortium and following standard procedures [16], the reference laboratory of the Public Health National Institute (INSP – Institut National de Santé Publique) tested 236 clinical samples positive for V. cholerae O1 throughout the duration of the 2012 epidemic, out of which 212 isolates were prospectively stored in a biobank created for that purpose. In September 2012, 50 of these isolates were selected for genotyping, subcultured and then transported in glycerol tubes at room temperature to Marseille, France. Isolates were selected in a manner in which the samples were temporally and spatially representative of outbreak diffusion during the first 8 months of the epidemic and included early and later isolates from all 7 prefectures available in the biobank. Upon arrival in Marseille, the strains were recultivated on non-selective trypticase soy agar (TSA) medium (Difco Laboratories/BD) for 24 hours at 37°C. Suspected V. cholerae colonies were identified via Gram-staining, oxidase reaction and agglutination assessment with V. cholerae O1 polyvalent antisera (Bio-Rad). For DNA extraction, an aliquot of cultured cells was suspended in 500 µL deionized water, incubated for 10 min at 100°C and centrifuged for 10 min at 1500× g. The pellet was then resuspended in 250 µL deionized water and incubated for 5 min at 100°C. The supernatant (containing DNA) was subsequently stored at −20°C. DNA was directly extracted from the glycerol transport tubes for the isolates that failed to grow upon culture.
Genotyping of the V. cholerae strains was performed via MLVA (Multiple Loci VNTR (Variable Number Tandem Repeat) Analysis) of 6 VNTRs (Table 1), including 4 previously described assays [17], [18] and 2 assays specifically designed for this study to improve the discriminating power of the analysis. The novel VNTR assays were designed based on the reference strain El Tor N16961 (GenBank accession numbers AE003852.1 and AE003853.1) using Perfect Microsatellite Repeat Finder webserver (currently unavailable). Specific primer pairs were subsequently designed using the Primer3 program (http://simgene.com/Primer3) (Table 1). Fluorescent-labeled primers were purchased from Applied Biosystems.
For each PCR assay, DNA amplification was carried out by mixing 0.375 µL of each primer (20 µM), 1 X LightCycler 480 Probes Master (Roche Diagnostics) and approximately 100 ng of template DNA in a total volume of 30 µL. PCR was performed using a LightCycler 480 System (Roche Diagnostics) with the thermal cycling conditions described in Table 1. PCR amplicons were subsequently verified via agarose gel (2%) electrophoresis.
VNTR PCR product size was determined via capillary electrophoresis. Aliquots of the PCR products were first diluted 1∶100 in sterile water, which was further diluted 1∶100 in a solution containing 25 µL Hi-Di Formamide 3500 Dx Series (Applied Biosystems) and 0.5 µL GeneScan 500 LIZ Size Standard (Applied Biosystems). The fluorescent end-labeled amplicons were analyzed using an ABI PRISM 310 Genetic Analyzer (Applied Biosystems) with POP-7 Polymer (Applied Biosystems). Finally, amplicon size was determined using GeneMapper v.3.0 software (Applied Biosystems).
To better characterize the V. cholerae strains responsible for the epidemic, whole-genome sequencing was performed on a strain isolated at the onset of the epidemic (strain G298_Guinea) using a GS FLX+ System (454 Life Science, a Roche company). The DNA sequence was assembled using Newbler, from GS De novo Assembler (http://454.com/products/analysis-software/index.asp).
To perform a phylogenetic assessment of the core V. cholerae genome based on genome-wide SNPs (single nucleotide polymorphisms), strain G298_Guinea DNA was re-sequenced using a HiSeq Illumina System (Illumina).
For the spatiotemporal description of the epidemic, rainfall data were aggregated weekly and graphically represented in parallel with cholera morbidity. Cholera attack rates were calculated and mapped, by prefecture and sub-prefecture, for various time periods using shapefiles of administrative divisions obtained from the HealthMapper application (WHO, Geneva, Switzerland) and Quantum GIS v1.8.0 (QGIS Geographic Information System, Open Source Geospatial Foundation Project, available at: http://qgis.osgeo.org).
MLVA-based genotypes were compared at each of the 6 VNTR loci. Genetic relatedness between the strains was first assessed using eBURSTv3 (http://eburst.mlst.net/), which aims to identify the founding genotype. A simple network of all possible links between genotypes was also assembled using Gephi.0.8.1 beta software (https://gephi.org/). Molecular epidemiology analyses were completed via the sequential mapping of each genotype by month at the prefecture level.
After the first whole-genome sequence was obtained with the GS FLX+ System, proteins were predicted using Prodigal software (http://prodigal.ornl.gov/). Data was then annotated employing the GenBank database (http://www.ncbi.nlm.nih.gov/genbank) and the Clusters of Orthologous Groups database using BLASTP with an E-value of 10−5. Allelic polymorphism of the cholera toxin B subunit and other virulence factors was characterized by comparing the obtained sequence with the genome description of V. cholerae strains available in GenBank and recent literature.
For phylogenetic analyses, the paired-end read data obtained with the HiSeq Illumina System and sequence data from 198 previously sequenced strains available in the NCBI SRA database were mapped to the reference N16961 El Tor strain (NCBI accession numbers AE003852 and AE003853) using SMALT software (http://www.sanger.ac.uk/resources/software/smalt). A whole-genome alignment was obtained for each strain in this analysis, and SNPs were called using the approach described by Harris et al. [19]. The reads that did not map to the N16961 genome were filtered out during SNP calling, and any SNP with a quality score less than 30 was excluded. A true SNP was only called if there were at least 75% of the reads at any heterogeneously mapped ambiguous sites. High-density SNP clusters indicating possible recombination sites were excluded using the methodology previously described by Croucher et al. [20]. Maximum Likelihood phylogenetic trees were estimated using the default settings of RAxML v0.7.4 [21] based on all the SNPs called in the manner explained above. M66 (accession numbers CP001233 and CP001234), a pre-seventh pandemic strain, was used to root the final phylogenetic tree of the seventh pandemic strains [13]. FigTree (http://tree.bio.ed.ac.uk/software/figtree/) was used to visualize and order the nodes of the phylogenetic tree.
Taking into account both the national database and patient line list, this epidemic was responsible for an estimated 7,743 suspected cases (global attack rate: 6.3 cases/10,000 inhabitants) and 138 deaths (case fatality ratio: 1.8%). The initial case was reported on February 2, 2012 (epidemiological week 5) in the midst of the dry season (Figure 1). The weekly number of new cases remained below 100 until July. The epidemic then peaked in August, 5 months after the onset of the rainy season, with nearly 1,188 new cases recorded during week 34. Cholera incidence began to markedly decline in September. The final case was recorded on December 11, 2012, and the Minister of Health officially declared the end of the epidemic on February 6, 2013.
Overall, the capital of Conakry reported 4,642 cases (25.9 cases/10,000 inhab.), which represented more than half of the national case total, but only 24 deaths (case fatality ratio: 0.5%). Moreover, 2,178 additional patients were located in the 5 other prefectures that border the Atlantic Ocean, with the highest attack rate observed in Coyah (55.1 cases/10,000 inhab.) (Figure 2). Twelve other prefectures were also affected, including distant inland prefectures such as Kerouane (Figure 2).
The initial cholera cases in Guinea emerged on February 2, 2012 on Kaback Island (Prefecture of Forecariah) (Figure 3), which is located in a remote mangrove zone close to the border with Sierra Leone, where an epidemic of acute diarrhea and vomiting had been reported in January. The Guinean index case was a fisherman who had just traveled by boat from Sierra Leone (a village on Yeliboyah Island, Kambia District) and arrived in the fishing village of Khounyi, on a land strip of the southern tip of Kaback Island. During the first month of the epidemic, this small village, which lacked safe water and improved sanitation facilities, recorded over 100 cases and represented the most affected community in the prefecture.
The cholera epidemic then progressively diffused northwestward along the Guinean coast, striking the prefectures of Boffa on February 23 and Boke on April 22 (Figure 3). As observed in Forecariah Prefecture, the initial cases in Boffa and Boke were also reported in fishing camps, namely Sakama and Yongosale, respectively. At each of these lowland fishing locales, the index case was a fisherman who had recently returned from an already affected area (i.e., travelling from Kaback to Sakama and from Koukoude (Boffa prefecture, Douprou sub-prefecture) to Yongosale). Concomitant with the expansion of the epidemic along the coast, cholera had also begun to spread inland. However, the inland prefectures were not significantly affected until the onset of the rainy season. Likewise, although Conakry is situated on a peninsula between the early affected regions of Kaback and Boffa, cholera did not strike the capital until a month after the inception of the rainy season. The first case in Conakry was officially recorded on May 29, who appeared to be a merchant returning from the Kaback market. Conakry subsequently acted as an amplifier of epidemic spread, especially towards the interior portions of the country, where several identified index cases were found to be drivers, merchants or students recently returning from the capital.
Fourteen samples out of 50 were not positive by culture and 2 additional samples were heavily contaminated. However, direct DNA extraction from transport tubes was successful for 4 culture-negative isolates. Genotype analysis with the 6-VNTR panel was thus performed on 38 V. cholerae isolates. All strains displayed constant results for the VC1, VC5 and LAV8 assays, while the VC4, VC9 and LAV6 assays revealed 4, 3 and 6 allelic variants, respectively. Based on the MLVA results, the strains were grouped into 12 different genotypic profiles, all of which were very closely related (Figure 4). All strains seemed to have arisen from genotype #1, which was identified as the founder genotype using the eBURST algorithm. Genotype #1 represented the earliest genotype isolated during the 2012 epidemic (on Kaback Island in February 2012) as well as the most frequent genotype identified (Figure 4). Subsequent diversification of this clone occurred via 1 or 2 mutational events during its propagation across the country (Figure 4).
The genome of a genotype #1 strain isolated on February 28, 2012 in Kaback was examined via whole-genome sequencing. The cluster composition of the virulence genes displayed one “hybrid” CTXϕ prophage on chromosome 1 but no RS1 fragment. Sequence results showed that this “hybrid” CTXϕ harbors a majority of El Tor allele genes (e.g., zot, ace and cep) with a classical ctxB gene (encoding the B subunit of the cholera toxin) and a classical rstR gene. Strain phylogeny based on genome-wide SNP analysis situated this Guinean “atypical” El Tor variant within a new clade of the third and most recent wave of the seventh pandemic (Figure 5). This strain was thus distinct from both strains isolated in Mozambique in 2004–2005 (second wave) and strains isolated between 2005 and 2010 in Eastern Africa (i.e., the Kenyan clade within the third wave, indicated in purple on Figure 5). The Guinean 2012 strain was also found to be clearly separated from two South Asian clades (indicated in sky blue on Figure 5), which includes the Haitian clone. The closest relative of the Guinean strain was a strain isolated in 1994 in Bangladesh.
While tracking the origin of the 2012 Guinean cholera epidemic, this multidisciplinary study demonstrates the monoclonal nature of the epidemic, as clinical V. cholerae strains exhibited a progressive genetic diversification that paralleled outbreak diffusion from Kaback Island. Molecular results confirmed the epidemiological findings, as the single ancestral and most abundant genotype was the sole V. cholerae strain isolated during the onset of the epidemic in February, at the initial focus of Kaback. According to field investigations, the index case was a fisherman arriving from a nearby cholera-affected district of Sierra Leone. Cholera then bounced along the Guinean coast, likely carried by other infected fishermen, before exploding during the rainy season in the capital Conakry and subsequently spreading inland. This clone was found to be an “atypical” El Tor variant of V. cholerae, as determined via whole-genome sequencing. Furthermore, this Guinean strain phylogenetically grouped into a new clade of the third wave of the current pandemic, and the closest known relative was a strain isolated in Bangladesh in 1994. This study represents the first such molecular analysis of a cholera epidemic conducted in West Africa.
Overall, these results strongly suggest that cholera spread along the coast of Guinea due to human-driven diffusion of the bacterium. According to the molecular analyses, this epidemic was caused by a single clone, which rapidly evolved in parallel with the spatiotemporal spread of the epidemic. A few weeks after identification of the founder clone in Kaback, the same genotype was identified in new outbreak foci further along the coast, where it was likely transported by infected traveling fishermen. Likewise, isolates characterized by descendant genotypes were found to have spread across the country throughout the year, with strains of the most distant genotypes primarily identified in distant prefectures, such as Kerouane, several months later (e.g., August and September). Such genotype analysis has rarely been conducted to assess cholera epidemic diffusion from the onset. However, similar genetic diversification from an initial V. cholerae clone has been recently observed throughout the current epidemic in Haiti [22], where the human-associated importation of cholera is largely undoubted [23], [24]. Furthermore, the diffusion of cholera by traveling fishermen has already been documented in West Africa. For example, the arrival of the seventh cholera pandemic in Ghana in 1971 was linked to the repatriation of a man who had succumbed to the disease while fishing in the waters of Togo, Liberia and Guinea [25].
Conversely, had the 2012 cholera epidemic originated from a local aquatic reservoir of proliferating vibrios, the diversity of V. cholerae strains found in the environment would have resulted in the early identification of several distinct clones [8], [26]. Therefore, the emergence of a unique V. cholerae genotype in clinical samples isolated on Kaback Island in February does not correlate with environment-to-human transmission of the disease. Furthermore, this period was not characterized by the wet and warm climatic conditions that are considered to be a favorable to V. cholerae proliferation in water bodies [8]–[12]. Finally, a recent review addressing cholera epidemics in African coastal areas has indicated that no perennial environmental reservoir of toxigenic V. cholerae O1 has yet been identified in West Africa, which may be attributed to the lack of appropriate studies [27].
The epidemiological data rather suggest that cholera was imported to Guinea from Sierra Leone. Indeed, Kaback is situated less than 30 km away from this neighboring country. Nearby districts of Sierra Leone, including Kambia and Port Loko, were already affected by the disease in early January 2012 [28]. Furthermore, the index case identified in Kaback was a travelling fisherman who had just arrived from a fishing village in Kambia. Unlike Guinea, where an efficient early alert system [4] enabled the detection, report, investigation, laboratory-confirmation and official declaration of the outbreak within 8 days after the appearance of the first cholera case observed in the past 3 years, health authorities in Sierra Leone did not perform similar investigations. Thus, the origin of this cholera epidemic in Sierra Leone remains unclear, although possible importation events by fishermen travelling from Liberia and Ghana have been reported [29].
Finally, according to whole-genome sequence analysis, this epidemic was caused by an “atypical” El Tor variant of V. cholerae O1, a type of strain that harbors both El Tor biotype genetic elements and the Classical biotype ctxB gene [30]. Such “atypical” El Tor strains initially emerged in Asia in 1991 and were first detected on the African continent in 2004 [31]. This may also present major public health implications as these strains have been suggested to be associated with more severe clinical symptoms compared with conventional El Tor strains [32], [33]. Furthermore, genome-wide SNP-based phylogeny analysis grouped the Guinean 2012 clone into a recent clade within the third wave of the seventh pandemic. Several studies have shown that this monophyletic radiation is largely distinct from the vast diversity of V. cholerae environmental strains [14], [34], which suggests that cholera epidemics are clonal and caused by a specific subset of related V. cholerae strains often spread via human-to-human transmission [14], [35].
Nevertheless, to confirm the origin of the V. cholerae clone responsible for this epidemic, it would have been ideal to analyze pre-epidemic environmental isolates as well as isolates from previous epidemics in Guinea, isolates from Sierra Leone and strains from other countries the region. However, earlier Guinean isolates were not stored and we did not have access to strains from Sierra Leone. Furthermore, no study of environmental V. cholerae strains had previously been performed in the region.
In conclusion, by tracking the origin of the 2012 cholera epidemic in the Republic of Guinea, this study identified fishermen as cholera victims and vectors during the early phase of epidemic propagation. Improving water and sanitation infrastructures, implementing enhanced hygiene education programs and targeting oral cholera vaccination campaigns in high-risk coastal areas could thus benefit these vulnerable populations and prevent the spread of future cholera outbreaks. The likely Sierra Leonean origin of this Guinean epidemic highlights the importance of encouraging transborder collaboration in the surveillance and control of highly mobile populations and main communication routes so as to rapidly identify emerging foci and organize coordinated targeted responses. These results also support the implementation of biobanks dedicated to prospective clinical and environmental V. cholerae isolates, to perform molecular epidemiological analyses, which have become essential to interpret field investigation data. Such an integrated approach would provide valuable insights concerning cholera in other African regions, where the key determinants of all too frequent epidemics still remain poorly understood and prevention or control strategies are not always accurately oriented.
|
10.1371/journal.pbio.1001869 | Network Analyses Reveal Pervasive Functional Regulation Between Proteases in the Human Protease Web | Proteolytic processing is an irreversible posttranslational modification affecting a large portion of the proteome. Protease-cleaved mediators frequently exhibit altered activity, and biological pathways are often regulated by proteolytic processing. Many of these mechanisms have not been appreciated as being protease-dependent, and the potential in unraveling a complex new dimension of biological control is increasingly recognized. Proteases are currently believed to act individually or in isolated cascades. However, conclusive but scattered biochemical evidence indicates broader regulation of proteases by protease and inhibitor interactions. Therefore, to systematically study such interactions, we assembled curated protease cleavage and inhibition data into a global, computational representation, termed the protease web. This revealed that proteases pervasively influence the activity of other proteases directly or by cleaving intermediate proteases or protease inhibitors. The protease web spans four classes of proteases and inhibitors and so links both recently and classically described protease groups and cascades, which can no longer be viewed as operating in isolation in vivo. We demonstrated that this observation, termed reachability, is robust to alterations in the data and will only increase in the future as additional data are added. We further show how subnetworks of the web are operational in 23 different tissues reflecting different phenotypes. We applied our network to develop novel insights into biologically relevant protease interactions using cell-specific proteases of the polymorphonuclear leukocyte as a system. Predictions from the protease web on the activity of matrix metalloproteinase 8 (MMP8) and neutrophil elastase being linked by an inactivating cleavage of serpinA1 by MMP8 were validated and explain perplexing Mmp8−/− versus wild-type polymorphonuclear chemokine cleavages in vivo. Our findings supply systematically derived and validated evidence for the existence of the protease web, a network that affects the activity of most proteases and thereby influences the functional state of the proteome and cell activity.
| Proteases modify the structure and activity of all proteins by peptide bond hydrolysis and are increasingly recognized as integral regulatory components of numerous biological mechanisms. Deregulated protease activity is a common characteristic of many diseases. However, protease drug development is complicated by an incomplete understanding of protease biology. One missing piece in this puzzle is the interplay between proteases: Some proteases activate other proteases, whereas some proteases inactivate inhibitors, leading to currently unpredictable cleavage of additional proteins. Using database annotations we mathematically modeled protease interactions. Our model includes 1,230 proteins and shows connections between 141,523 pairs of proteases, substrates, and inhibitors. Thus, proteases interact on a large scale to form the protease web, which links most studied groups of proteases and their inhibitors, indicating that the potential of regulation through this network is very large. We found that this interplay is robust to targeted or untargeted pruning of the protease web and that protease inhibitors are central to network connectivity. Our model was used to decipher proteolytic pathways that drive inflammatory processes in vivo. Consequently, protease regulatory interactions should be recognized and explored further to understand in vivo roles and to select better drug targets that avoid side effects arising from inhibition of unexpected activities.
| Proteolysis, the hydrolysis of peptide and isopeptide bonds in protein substrates by proteases (also termed peptidases or proteinases [1]), affects every protein at some point during its lifetime. The outcomes of proteolysis are of two kinds: Protein degradation ablates protein function by breakdown to amino acids, whereas proteolytic processing is an irreversible posttranslational modification to precisely produce modified, stable protein chains. The length of this cleavage product is defined by the substrate site specificity of the protease catalyzing the reaction, which can be exquisite. Processed proteins often have radically altered activity, protein interactions, structure, or cellular location and hence are implicated in many human diseases [2]–[4]. Recent research has focused on identifying the cleavage products of protease activity in cell culture and in vivo as a means of understanding their biological roles and hence guiding drug target identification and validation [5]. This need has led to the development of genomics and proteomics approaches that have come to be termed degradomics [6],[7] in which the specialized subfield known as terminomics that identifies N termini [8]–[10] and C termini [11],[12] has seen recent rapid development. In one such terminomics analysis of murine skin in vivo, ∼44% of identified N termini mapped to internal positions in proteins, revealing proteolytic cleavage after translation as part of protein maturation and function [13]. With ∼68% of identified N-termini being internal, human erythrocytes have been found to possess an even higher proportion of processed proteins [14]. These recent findings demonstrate that proteolytic processing is a widespread and functionally important posttranslational modification. Thereby, proteolytic processing modifies the activity of many more proteins than currently appreciated from conventional shotgun proteomics analyses and biological studies.
As exemplified by N-terminal cleavage of chemokines [6], the activity of a protein often depends on the exact position and nature of its N and C termini [15]. Therefore, identifying the termini of proteins is essential for functional insight into protein bioactivity, annotation of proteins in the Human Proteome Project, and drug development [14]. However, deeper biological insight requires identifying the protease responsible for generation of neo-termini that distinguish cleavage products from the original protein termini. Whereas low- and high-throughput methods to identify the in vitro substrate repertoire of proteases, also known as the substrate degradome [7], are well established, in vivo identification is problematic [16]. In vitro experiments can only indicate potential cleavage in vivo because of difficulties assigning precise parameters governing cleavage in the actual biological system, such as protease and substrate colocalization spatially and temporally, presence of inhibitors, zymogen activation, pH, ion concentrations, interaction with nonprotein compounds [17], as well as O-glycosylation or phosphorylation of the protease or substrate [18]. Hence, posttranslational modifications of proteases, inhibitors, and their substrates add complexity to the dynamic nature of the proteome and cell responses. Thus, an observed cleavage in vitro might not occur in vivo—that is, “just because it can (in vitro) does not mean it does (in vivo)” [5].
In vivo studies, which rely on comparing samples of protease knockout or inhibition to controls, are hampered in particular because the underlying biological system reacts to the removal of a protease or inhibitor in complex and unpredictable ways. For example, a protease knockout can lead to alterations in gene expression profiles of proteases, inhibitors, and substrates [13],[19], due to the biological consequences of altered substrate cleavages in vivo, including cleavage of transcription factors [20]. Another factor is the activation of other proteases in the system through increasingly recognized activation cascades of protease zymogens by other proteases and the proteolytic regulation of protease inhibitor activity by nontarget proteases that cleave and inactivate the inhibitor. For example, serpins and cystatins inhibit serine and cysteine proteases, respectively, but when cleaved by a matrix metalloproteinase (MMP), the inhibitor is inactivated and the protease remains active [13],[21]–[23]. Through activating and inactivating cleavages of other proteases and inhibitors, a protease thereby indirectly influences the activity of additional proteases. Such interactions can lead to knock-on effects that alter the cleavage of a range of additional protein substrates that are not direct substrates of the protease. Furthermore, titration of inhibitors upon covalent or tight interaction with one protease can reduce the availability of free inhibitors to regulate other proteases. Consequently, phenotyping protease and inhibitor genetic knockout mice is complicated, which also hampers biological understanding and drug target validation of proteases.
Protease biology is also complex due to the large protease numbers in humans (460) and mice (525), which form the second largest enzyme family after ubiquitin ligases in these organisms [24]. Moreover, an additional 93 and 103 are predicted to be inactive proteases in human and mouse, respectively, which often can function as dominant negative counterparts [24]. Protease numbers are almost equally distributed in the intracellular and extracellular environments, and other than some proteases that segue between these two compartments, this distribution partitions and limits their potential interactions with each other. In an effort to systematically comprehend this complex biology, proteases are grouped by the MEROPS database, which is assembled from biochemical experimental data curated from the literature, into seven classes, five of which are found in human and mouse, according to the active site residue catalyzing substrate cleavage, and into clans based on the structure of the active site [25]. Similarly, inhibitors are commonly grouped according to the class of proteases they inhibit, with several inhibitors exhibiting broad inhibitory activity against proteases from more than one class. Interactions between proteases of the same class are well established as part of classically described cascades of proteases such as the complement [26]–[28] and coagulation [29],[30] systems, and newer recognized cascades such as kallikreins [31] and caspases in apoptosis [2],[32]–[34]. However, wide-ranging additional protease interactions have also been proposed to extend more globally to link networks forming what was termed the protease web [35]. The protease web was defined as the universe of cleavage and inhibition interactions between proteases and their inhibitors. Stemming from examples in simple systems such as in vitro biochemical analyses and early in vitro and cell culture degradomics analyses of protease substrates [36]–[38], and mRNA transcript analyses in cancer upon administration of protease inhibitors or tissue inhibitor of metalloproteinase (TIMP) overexpression and knockout studies [19], the protease web concept has been well supported. Extending terminomics analyses to in vivo situations, for example skin inflammation in wild-type versus Mmp2 knockout mice in vivo, has revealed hitherto biologically relevant and unsuspected critical connections of MMPs in regulating the complement and coagulation cascades and the plasma kallikrein system, which regulates vessel permeability through bradykinin excision and release from kininogen [13].
Such interactions between protease families were shown to create small networks in specific cases [13],[19],[39],[40], but the full extent of the protease web, the fraction of proteases and inhibitors involved, and hence the regulatory potential of this network remain underexplored and underappreciated despite the potentially wide impact on the functional state of proteomes. Furthermore, the protease web is a black box with an unknown mechanism of regulation—it is unclear whether it follows a super structure of known cascades, where signals are amplified downstream, or forms more of a network, where signals can flow in multiple directions with multiple positive and negative feedback loops [35],[40]. Similarly, it is unclear which are the main regulatory protein switches controlling subparts of the network. Descriptions of the protease web are difficult to assemble, as many proteases remain poorly studied and characterized. Likewise, many proteases have no described inhibitors and many predicted inhibitors have unknown protease targets and deorphanization examples are uncommon [41].
Here, we assessed the global extent and structure of protease interactions computationally. Graph models are used to describe multiple interactions between many elements and have been applied extensively in research on various biological networks. We represented existing biochemically validated data on protease cleavages and inhibition as annotated in the manually curated database TopFIND [42] as organism-specific networks. TopFIND stores established biochemical information on substrate cleavage and protease inhibition from MEROPS [25], the most complete collection of such data, most of it published, and combines it with published high-throughput terminomics and degradomics datasets as well as protein annotations from UniProt [43] for five different organisms. Our analyses revealed a large and pervasive network spanning all known cascades and four of the five protease classes present in human and mouse tissues. The network is highly connected in that via a few connections a protease can potentially influence many other proteases, with inhibitors often taking a special role as key connectors in the protease web. We demonstrate the utility of our analysis by applying the network to gain mechanistic in vivo insights into protease web effects, which we then validated in vitro, in cell culture, and in vivo.
Functional protease interactions comprising cleavage and inhibition events influence the in vivo cleavage of substrates in many ways. Cleavage of a substrate by a protease is a direct event, and as shown in Figure 1, by cleaving other proteases and protease inhibitors, one protease can activate, inactivate, or alter the activity of a second protease, thereby indirectly influencing the cleavage of substrates of another protease. To assess the global extent of such effects, we represented protease interactions as a graph, connecting proteases and protease inhibitors to their established substrates and protease targets, respectively. The resulting graph contains nodes, which are proteins, and edges, which represent cleavages or inhibitions. Edges link proteases to their substrates and protease inhibitors to their target proteases. Therefore, edges are directed: an edge from protein X to protein Y signifies cleavage or inhibition of Y by X but does not contain information about cleavage or inhibition of X by Y. In graph theory, the latter would require another edge with the opposite directionality. Figure 1 outlines functional protease interactions and how they are represented in small graph models, which were then aggregated to represent the full complexity of the protease web based on curated biochemical data as described below.
As input to our analysis of the protease network, we used the TopFIND v 2.0 knowledgebase [44] to retrieve validated cleavage and inhibition data mostly annotated from published experiments. TopFIND contained 4,774 cleavages for Homo sapiens, 3,679 for Mus musculus, 426 for Escherichia coli, 190 for yeast, and 43 for Arabidopsis thaliana. Due to the low number of cleavages annotated for other organisms, we focused our analysis on human and mouse. Only proteins performing an annotated cleavage or inhibition were added, and then these were connected via edges representing the biochemical reactions as explained in Figure 1. These networks extend the protease web, which contains only proteases and inhibitors, by also including all other substrates of proteases, and hence represent the annotated functional proteolytic interactions between the substrates in the proteome and the protease web. The human and murine networks (with 1,230 and 1,393 nodes, respectively) are shown in high resolution upon click-to-zoom in Figure S1 and available for download as a Cytoscape file, gml file, and R objects at www.chibi.ubc.ca/ProteaseWeb and http://clipserve.clip.ubc.ca/supplements/protease-web.
The human and murine proteolytic networks show that the majority of proteins are connected and only very few are in unconnected components. Thus, in both networks, the Largest Connected Component (i.e., the biggest group of nodes directly or indirectly connected) encompasses the vast majority of these proteins—1,183 of 1,230 (96%) in human and 1,377 of 1,393 (99%) in mouse (Table 1). This remarkable connectivity is particularly surprising given the incompleteness of annotation currently available in the databases. Indeed, Table 1 shows that of 460 human proteases, only 244 (53%) have one or more known and annotated substrates. In mouse this number is even lower, with only 88 of 525 (17%) proteases having a substrate annotated. Furthermore, even the data on these proteases are incomplete and biased, with most substrates assigned to few, well-studied proteases. Figure S2 shows the out-degree (i.e., the sum of cleavages catalyzed by a protease or the sum of inhibitions caused by a protease inhibitor) for proteases and inhibitors having any annotated cleavage or inhibition, respectively. Although few proteases have a large known substrate repertoire (higher out-degree), most proteases have very few known substrates. Although this could be due to high substrate specificity, it is more likely that these proteases simply received less attention in studies dedicated to discover substrate repertoires. This effect is especially pronounced for the mouse data, where 80% of total cleavages (2,938 of 3,679) are assigned to three proteases—cathepsin D (UniProt: P18242), cathepsin E (UniProt: P70269), and MMP2 (UniProt: P33434)—and are mostly derived from high-throughput proteomics screens. Accordingly, the annotations differ strongly between human and mouse. Although the networks have similar size (1,230 and 1,393 nodes, respectively), they overlap minimally, with only 126 of 3,852 connections in mouse (3.3%) reflected in 122 of 4,905 human connections (2.5%). However, we suggest that the small overlap is mostly due to differences in the state of data annotation between the networks rather than to actual differences in the evolution of these networks.
The human data are further biased in that proteases and inhibitors are largely overrepresented as substrates themselves (Figure S3). Strong representation of protease–protease cleavages is expected because many proteases are synthesized as zymogens requiring proteolytic cleavage for activation by other proteases. Indeed, this strong enrichment is found in the human TopFIND/MEROPS data, but less so in mouse. We compared these values to a terminomics data set of cleavages in mouse skin [13], which more accurately reflects reality because terminomics analyzes N termini in an unbiased fashion. However, in this in vivo data set, inhibitors, and not proteases, were overrepresented as processed proteins, indicating that the overrepresentation of proteases as cleavage substrates in the human in vitro database is likely exaggerated.
The observed data biases likely resulted from the nature of biochemical studies, where many substrates were identified for some “interesting” proteases (target bias) and “interesting” proteins are more likely to be tested as substrates (substrate bias). Substrate bias is especially found for proteases themselves, which are preferably tested as substrates in zymogen activation studies. With the advent of degradomics utilizing proteomics methods dedicated to substrate discovery, we anticipate both an increase in target bias in the future with many substrates identified for a few proteases, and a decrease in substrate bias where any protein can be identified as a substrate without prior selection of interesting candidates. Therefore, the cleavages annotated represent a biased fraction of the biochemically possible cleavages in the organism compared with an unknown number of as yet uncharacterized cleavages. On these grounds, the high connectivity in both the mouse and human networks is even more noteworthy because future information can only further increase connectivity. The observed, extensive interactions between proteases and inhibitors are further characterized as described in the following.
In the interactions between proteases in proteolytic signaling pathways, there are major upstream regulators or initiation factors, whose proteolytic activity leads to the cleavage of downstream proteases, which in turn activate even further downstream factors that finally cleave and activate the effector molecules at the end of the pathway. A special case of proteolytic pathways are activation cascades, where signal amplification occurs to generate large quantities of the end protein products in seconds as classically described for coagulation [29],[30]. To investigate whether the connections in the overarching protease web follow such a pathway or cascade (hierarchical) structure, we used a graph measure termed reachability. Reachability of node X denotes the number of other nodes Y where there is a path from X to Y in the network. A path is a sequence of directed edges connecting X and Y, following the directionality of edges in the network. The path from X to Y can therefore be different from the path from Y to X (and the existence of one does not guarantee the existence of the other). In the protease web, reachability corresponds to the number of proteins that can be influenced by one protease or inhibitor. Figure 2A outlines reachability values of nodes in three theoretical examples: (i) an unconnected (single), (ii) a strongly connected (circle), and (iii) a cascade-like network (cascade). Figure 2B shows the respective distribution of reachability values of these three theoretical examples.
We next compared the theoretical reachability distributions with the distributions observed in our human and mouse protease networks. In order to specifically describe the selective connectivity between proteases and inhibitors, which form the protease web, we excluded from further analysis other simple substrates (nonprotease and noninhibitor proteins), whose reachability in the network is 1 by definition. Table 2 summarizes the resulting protease web networks for human (340) and mouse (220) proteins that have annotated cleavages or inhibitions. In analyzing the human and mouse protease webs, we further identified one dominant “largest connected component” comprised of 255 proteins for human and 187 proteins for mouse. Figure 2C compares the distribution of reachability scores in the largest connected component in mouse (blue curve) and human (red curve). In mouse, reachability indicates a cascade-like, hierarchical network, where most nodes have a very low reachability and fewer nodes have gradually higher reachability. In contrast, the reachability distribution of the human network follows a strongly bimodal distribution: 158 (62%) nodes reach 153 (60%) or more nodes. This is very high reachability that is most similar to the circle graph in Figure 2B, where any node can reach any other node. For a biological system, this implies that 158 proteases or inhibitors have the potential to regulate the activity of 153 or more other proteases and inhibitors in the network. In other words, there are one or more directed paths between 24,166 pairs of proteases in the human protease web, which are 37% of all 64,770 possible directed connections between pairs of 255 proteins. This number of connections between pairs rises to 141,523 paths when substrates are added (network with 1,230 nodes). This highlights the high degree of connectivity between proteases and inhibitors. Reachability between nodes does not take the path length between nodes into account and so might be the result of very long and hence biologically irrelevant paths in the network. However, this possibility can be excluded as most paths have a length of just four (Figure 2D). The lack of connectivity in the mouse network is not surprising given the small overlap between the two networks. We assume that this difference is due to data biases rather than a real biological difference, and accordingly we focused on characterizing the extensive and more complete human network.
High connectivity in the human protease web is due to a strongly connected component (87 nodes), a subgroup of nodes within the largest connected component, that can directly or indirectly reach each other and hence have the same reachability value of 153 (Table 3). We visualized this effect in Figure 3, where nodes of the human protease web are shown separated by their reachability. Upstream of the strongly connected component are 71 nodes with reachability higher than 153; these nodes can reach the strongly connected component, but cannot be reached from it. Downstream (with reachability smaller than 7) are 97 nodes, which cannot reach the strongly connected component. The nodes in Figure 3 are also colored according to their centrality in the network, as measured by node betweenness [45]. Betweenness is calculated by first finding the shortest paths (as explained above) between all 64,770 pairs of nodes in the network and then counting the number of times a node appears in these paths. Notably, all nodes with high betweenness are found in the strongly connected component; these nodes tether the network together. Nodes with high betweenness or reachability are listed in Table S1.
Figure 3 shows that our network data from MEROPS/TopFIND contain all the known proteolytic pathways (e.g., coagulation, complement system, apoptosis, and kallikreins) as they were discovered, published, and annotated previously in MEROPS (detailed in Figure S4). In addition, these proteolytic pathways are extended by connections linking known pathways with other pathways and additional proteases. Details of these connections can be found in Figure 4A, which shows separated protease groups in the strongly connected component after removing inhibitors. Figures 3, 4A, and S4 show that the observed connectivity in the protease web is caused by the concerted action of defined protease cascades and key protease inhibitors: alpha-2-macroglobulin (A2M, UniProt: P01023), amyloid precursor protein (APP, UniProt: P05067), kininogen 1 (KNG1, UniProt: P01042), and alpha-1-antitrypsin (also known as serpin A1) (SERPINA1, UniProt: P01009). Whereas intragroup connections are pervasive as expected, intergroup connections are also considerable, in particular between coagulation factors and kallikreins or MMPs, but also including cathepsins and caspases. These findings are confirmed in Figure 4B, which shows that connections among four of the five classes of proteases and protease inhibitors in human are extensive. Importantly, Figure 4B also shows proteases frequently cleaving inhibitors of other protease classes, an important regulatory aspect of protease activity. Only threonine proteases, which are found exclusively in large specialized cell organelles termed the proteasome and immunoproteasome, remain isolated from connections with other proteases and inhibitors according to current data. Note added in proof: However, a recent publication shows that the threonine proteasomal proteases are cleaved by intracellular MMP-12. Thus, all five classes of proteases in human and mouse are interconnected [81].
From a biological standpoint, the highly interconnected (reachable) nature of the protease web was surprising and underappreciated in the literature. To explore the degree to which this result is statistically surprising given the properties of the proteins making up the network, we investigated theoretical network models as well as randomized versions of the network. We first compared the protease web to two commonly used generative network models, the Erdős-Rényi model (ER) and the Barabasi-Albert model (BA), with parameters chosen to mimic the properties of the real network's member proteins (see Materials and Methods). We found that neither model (each 500 networks) adequately explains the data, yielding networks that have either much higher (ER) or lower (BA) reachability on average (Figure S5A–C). These experiments therefore leave open the statistical nature of the process that generates the network, which we stress currently involves both biological components and experimenter biases, the latter being due to the incomplete nature of the underlying biochemical analyses (many potential edges have not been tested). We next generated two types of edge-shuffled networks, one maintaining in- and out-degree of each node (“Shuffled”) and a second preserving overall in- and out-degree distributions of the network, but not for each node (“Shuffled2”). The mean reachability was lower in the real network (72.09) than in 353 Shuffled networks (70.6% of all 500; average reachability was 73.96 across all 500 networks; see Figure S5D) but higher than all 500 Shuffled2 networks (average 34.8; Figure S5C). Taken together, these results indicate that high reachability emerges quite readily in a network composed of proteins with the measured in- and out-degrees found in a real biological network, such as the protease web described here. In fact, a network without such high reachability—as it is often assumed in biochemistry and cell biology—would be surprising from these results. Importantly, this further suggests that the current biochemical description of cascades and individual proteases working in isolation is unlikely.
To assess reliability of high connectivity in the protease web, which we observed assuming that all cleavage and inhibition data are trustworthy, we addressed the possibility of erroneous data passing through database annotations into our network. A possibility of validating our findings is to compare the network to another second network derived from an orthologous data source. However, MEROPS being the only database of similar coverage, we instead tested whether the same connectivity can be observed by removing nodes in anticipation that some interactions are wrongly annotated. Protease specificity is mostly influenced by three factors: substrate sequence, substrate folding, and the encounter of protease and substrate [46]. In MEROPS/TopFIND, annotations are mostly derived from in vitro experiments where a protease is incubated with a substrate. Although some proteases are specific for given substrate sequences, others will cleave a wider range of sequences, but in both cases, possible cleavage sites can be masked in the protein structure of the substrate. Hence, experimental parameters of protease cleavage assays are designed to preserve protein folding and activity of both the protease and substrate in order to prevent unspecific cleavage of denatured substrates. Colocalization of proteases and substrates in vivo is an important factor but not unambiguously determinable, with unexpected localization recently revealed [20],[37],[47]–[49],[81]. In addition, most experiments are only performed if it can be assumed that the protease and substrate will colocalize in vivo. Assuming that most annotations are correct but individual assignments can be wrong, we randomly and selectively removed edges from the protease web (focusing on the regulatory core, the largest connected component with 255 nodes) to test how reachability is maintained or influenced by such modifications. We utilized the term “physiological relevance,” as annotated in MEROPS and TopFIND, to first create a high-confidence network (abbreviated as “hc” in Figure 5A) by removing all edges that were annotated with physiological relevance other than “yes.” As a consequence, the reachability of the resulting network was markedly decreased (Figure 5A), with the area under the curve (AUC) reduced to 22% of the original network. This was mostly due to the removal of all inhibitors (abbreviated as “i” below) as all 131 human inhibitions in TopFIND have a physiological relevance annotation of “unknown”; that is, their physiological relevance is not annotated in MEROPS from which TopFIND data are largely derived. Upon adding back the inhibitors to the high confidence network (“hc+i”), but still removing all “low confidence” nonphysiological cleavages, high reachability was largely recovered as indicated by an AUC of 88% of the original network. The observation that limiting the cleavages to high-confidence cleavages only barely reduces network connectivity strengthens the result that the protease web is not due to incorrect annotations. Moreover, removing inhibitions from the network severely impacted reachability and thus connectivity, highlighting the essential role of inhibitors in connecting the protease web.
Given the observed importance of inhibitors, we assessed the possibility of incorrect annotation of cleavages of inhibitors. The molecular mechanism of cysteine or serine protease inhibition by serpins involves cleavage of the serpin at its flexible reactive loop, which displays “bait” amino acids. Following cleavage, an induced conformational change leads to entrapment and inactivation of the protease [50],[51]. Because the trap occurs after formation of the acyl intermediate during catalysis, the inhibited serine proteases, but also some cysteine proteases, remain covalently bound to the inhibitor. In contrast, metalloproteinase and aspartic protease cleavage of serpins in the reactive loop does not result in their inhibition, as the nucleophile of these proteases classes is a water molecule. Thus, these proteases are not trapped and therefore escape inhibition, but the serpin is now inactivated. Mechanisms of trapping upon cleavage have also been observed for some metalloprotease inhibitors [52] and for A2M or pregnancy zone protein (PZP, UniProt: P20742), which use a physical trapping mechanism to inhibit all classes of proteases, except exopeptidases [53],[54]. Therefore, annotated cleavages of a protease inhibitor comprise cleavages that reflect either a regulatory inhibition of the protease or a regulatory inactivation cleavage of the inhibitor. To date, this distinction is not annotated in the databases, but is one that we suggest implementing. As a conservative estimate, we removed all cleavages of serpins by serine or cysteine proteases and from any protease to A2M or PZP (“inh rm” in Figure 5B). Therefore 144 edges were deleted from the original 1,238 edges of the largest connected component of the protease web (“orig” in Figure 5B). Notably, this removal only moderately reduced reachability (AUC 74% of original) and preserved a bimodal distribution. Thus, the high connectivity is not a result of unspecific inhibitors. Hence, the observed connectivity in the network is not an artifact attributable to ambiguous annotation of inhibitor cleavage and so further supports the importance of inhibitors in connecting the protease web.
We next assessed the dependence of reachability on individual nodes of the network. By removing each node individually, we found that reachability in the protease web is not dependent on any one single node (Figure S6). Indeed, by iteratively removing all nodes with the highest betweenness from the network, we identified the six most important nodes: plasminogen (PLG; UniProt: P00747), alpha-1-antitrypsin, A2M, cathepsin L1 (CTSL1; UniProt: P07711), alpha-1-antichymotrypsin (also known as serpin A3) (SERPINA3; UniProt: P01011), and kallikrein-4 (KLK4; UniProt: Q9Y5K2) (Figure 5C). Removing all six nodes simultaneously removes 227 edges whereupon this significantly breaks down the bimodal distribution of reachability values, an effect not observed when removing any combination of five out of the six connectors. Thus, high connectivity in the protease web is robust in that it depends not on a single protein, but rather on six important connectors. Furthermore, even after removal of those six nodes the reachability for many proteins remains high with many long paths in the network. Notably, none of these six important nodes are digestive tract proteases, such as trypsin or chymotrypsin, which are broad-acting proteases and ones that might have been expected to form many connections. However, we predict that the identity and number of these key connector proteins will change as more information on the protease web is uploaded to the databases with further experimentation.
Finally, we addressed the possibility of incorrect annotations by removing a fixed percentage of edges, thereby simulating a situation where these edges are incorrect cleavage or inhibition annotations and therefore would have to be removed from the network (Figure 5D). We randomly removed 10%, 20%, 30%, and 40% of all edges (cleavages and inhibitions) 200 times and then plotted the worst case for each experiment. The AUC was reduced to 78%, 65%, 47%, and 52%, respectively, but nonetheless even removal of 40% of edges still preserved the bimodality of the reachability values. Therefore, again the protease web shows a strong resistance to removal of elements, which further increases confidence in the description of a highly connected protease web with inherent robustness to change. This also leads to biological resilience and shows the importance of proteases that can nonetheless be resiliently maintained in genetic deficiencies or pathological perturbations of the system.
Our analyses suggested that the protease web represents a robust regulatory system of high complexity and flexibility enabling complex patterns of regulation of proteins at the posttranslational level. We next assessed how this system is implemented in vivo where only a fraction of proteases and inhibitors is expressed or active at the same time in the same cell, compartment, or tissue. We constructed tissue-specific networks based on protease and inhibitor gene expression levels in 23 different human tissues quantified by CLIP-CHIP microarray (Kappelhoff et al., unpublished data available at http://clipserve.clip.ubc.ca/supplements/protease-web). We used negative control spots on this microarray to define a threshold of expression at detectable levels and then limited networks to those proteins expressed above this threshold. We next plotted the reachability of the nodes in the largest connected component of the resulting networks for all 23 tissue-specific protease webs (Figure 6A). Figure 6B shows liver, spleen, and skin results in more detail. Although most tissue-specific networks (e.g., skin) show low reachability values, some preserve the strong connectivity of the original network totally (e.g., kidney and liver) or partially (e.g., spleen, small intestine, pancreas, lung, colon). Notably, the tissue-specific networks also show that reachability is highly dependent on expression of the same six network connectors shown in Figure 5C (Figure S7).
In agreement with our findings based on biochemical interactions, general biological literature also shows that proteases and their inhibitors can be involved in multiple biological processes (Figure 7A). It is easy to imagine that this multifunctionality is partly due to the interplay in the protease web. Indeed most of the proteins in Figure 7A are found in the strongly connected component of our protease web, indicating that they serve in connecting different biological processes. One example is TIMP1 (UniProt: P01033). Protein expression levels of TIMP1, an MMP inhibitor mainly involved in extracellular matrix remodeling and organization, were found associated with hemostasis [55]. This finding, which is derived from orthogonal data to the protease web, primed us to search for connections linking TIMP1 to coagulation factors, which we could indeed identify (Figure 7B). Together, these provide a plausible mechanism of action of TIMP1 and hence MMPs on coagulation and could explain the association observed. Hence, the protease web can be used to explain multifunctionality of proteases, which in turn strengthens our conclusion of a large interplay between proteases.
We were able to test the utility of our graph representation of the protease web by deciphering a previously inexplicable result in vivo. We analyzed the MMP8-dependent cleavage of the murine chemokine C-X-C motif chemokine 5 (CXCL5, UniProt: P50228), also known as lipopolysaccharide (LPS)-induced C-X-C chemokine LIX (LIX). LIX is a potent chemoattractant chemokine for polymorphonuclear (PMN) leukocytes, and MMP8 (UniProt: O70138) is PMN specific. It was previously demonstrated in an in vivo airpouch model that MMP8 knockout mice showed reduced PMN migration in response to LPS [56]. This was attributed to MMP8 processing and activation of LIX at position Ser4↓Val5, with a second cleavage at Lys79↓Arg80 of the 92-residue protein. Indeed the MMP8-truncated activated form of LIX (5–79) showed equal cell migration in wild-type and knockout mice, validating LIX as a physiological MMP8-dependent mechanism for promoting neutrophil infiltration in vivo. However, a neoepitope antibody specific to the MMP8-generated neo-N terminus failed to detect truncations at Ser4↓Val5 in the airpouch model. Thus, cleavage of LIX is a MMP8-dependent but MMP8-indirect event in vivo that could not be explained, prompting a further analysis of alternate MMP8-dependent proteolytic pathways predicted using our representation of the protease web.
To examine the importance of neutrophil-derived MMP8 in LIX processing and activation, we isolated bone marrow neutrophils from wild-type and MMP8 knockout mice. Neutrophils were stimulated with phorbol myristate acetate (PMA) followed by incubation of the activated neutrophils with chemokine for up to 3 h. Truncations of LIX generating the bioactive products LIX (9–92) and LIX (9–78), as determined by MALDI-TOF mass spectrometry from the still inactive form LIX (1–78), were readily apparent, even after only 1 h of incubation (Figure 8A). However, both the MMP8 knockout and wild-type neutrophils showed identical cleavage sites (Ala8↓Thr9 and Ala78↓Lys79) and cleavage kinetics. Because these sites differ from the MMP8 cleavage sites (Figures S8, S9, and 8B), MMP8 is not the dominant neutrophil protease cleaving LIX in the cellular context. Investigating protease web effects that may account for this, we found that LIX cleavage by neutrophils was inhibited by the serine protease inhibitor 2-aminoethyl benzenesulfonyl fluoride hydrochloride (Figure 8C). This showed that one or more of the four serine proteases in neutrophils—neutrophil elastase (UniProt: Q3UP87), cathepsin G (UniProt: P28293) [57], proteinase-3 (UniProt: Q61096), or the recently described neutrophil serine proteinase 4 (UniProt: Q14B24) [58]—were responsible for LIX cleavage. Using low concentrations of the endogenous serine proteinase inhibitors α1-proteinase inhibitor (α1-PI, UniProt: P07758) [21] and secreted leukocyte proteinase inhibitor (SLPI, UniProt: P97430) (Figure 8C), we excluded proteinase-3 and neutral serine proteinase 4 as candidates, as SLPI does not inhibit these proteinases [58],[59]. Moreover, neutral serine proteinase 4 has a stringent substrate specificity that does not fit our observed cleavage sites. Cathepsin G did not cut after Ala8 and required high enzyme concentrations (>100 nM) in generating the C-terminal cleavage (Figure S9) as it was inefficient with a kcat/KM 60 M−1 s−1. Thus, neutrophil elastase was the strongest candidate, and indeed 1 nM elastase efficiently cleaved LIX with a kcat/KM 1,200 M−1 s−1 at Ala8↓Thr9 and Ala78↓Lys79 (Figures 8D, S8, and S9). Because MMP8 cleaves N-terminal to the Ala8↓Thr9 elastase site and C-terminal to the Ala78↓Lys79 elastase site, truncations by elastase will remove evidence of any MMP8 cleavage. Furthermore, MMP8 is less efficient (kcat/KM 600 M−1 s−1) than elastase in cleaving LIX. Thus, elastase is the dominant protease for LIX cleavage by neutrophils in vivo.
To explain the paradoxical result that in the Mmp8−/− mouse LIX is not cleaved in vivo despite the presence of neutrophil elastase, we employed path finding in the protease web to identify potential regulatory effects from MMP8 on neutrophil elastase. Although no path was found in the murine network, the more extensive human network contains a path that had potential to explain this perplexing result (Figure 8E). Human MMP8 is known to cleave and inactivate human α1-PI [21], the potent inhibitor of neutrophil elastase, but SLPI is resistant to MMP8 cleavage [60]. We verified α1-PI cleavage by MMP8 using mouse proteins for the first time at various enzyme-to-substrate ratios and in time course experiments (Figure 8F) from which we found that murine MMP8 efficiently cleaves and inactivates murine α1-PI in vitro with a kcat/KM 7.7×103 M−1 s−1.
We next validated the in vitro results in vivo. In murine bronchioalveolar lavage collected following 24 h of treatment with LPS, both the full-length and high molecular weight forms of α1-PI, which were present as inhibitor-serine protease complexes, were greatly enhanced in Mmp8−/− mice compared to wild type (Figure 8G). Together, these in vitro and in vivo data show that efficient cleavage of α1-PI occurs by MMP8 in vivo and indicates the importance of MMP8 in modulating the balance of functional α1-PI protein and activity in vivo and hence elastase activity. This result further shows that MMP9, which also cleaves alpha1-PI in vitro, does not functionally compensate for MMP8 in vivo. This is despite MMP9 being in the same cytosolic granules as MMP8 and being present at elevated concentrations in the neutrophils from the MMP8 knock out mouse. Finally, we confirmed neutrophil elastase-dependent LIX cleavage in vivo using a specific neutrophil elastase chemical inhibitor (GW311616). Specific elastase inhibition reduced the relative numbers of neutrophils in wild-type mouse bronchioalveolar lavage similar to the decrease in cell migration in the MMP8 knockout versus the wild-type mouse bronchioalveolar lavage (Figure 8H). We conclude that MMP8 cleaves and inactivates α1-PI in vivo acting as the “metallo-serpin” switch leading to increased neutrophil elastase activity and LIX activation, which thereby promotes neutrophil infiltration in vivo. Evidence of LIX cleavage by MMP8 is lost following elastase cleavage in vivo, which is also catalytically more efficient than MMP8. Thus, the protease web enabled deconvolution of a complex biologically relevant proteolytic event and in turn formulation of a testable hypothesis that was confirmed in vitro and in vivo.
To our knowledge, this is the first systematic bioinformatics analysis of the extent and structure of the protease web. We assembled in silico networks comprising all biochemically annotated interactions between proteases and their inhibitors, which therefore represent the potential of regulation among proteases based on current biochemical data. By representing the human protease web as a graph, we show the depth of how proteases and inhibitors regulate each other across families and even catalytic classes. Thus, known cascades and proteases do not act in isolation, as often assumed, but crosstalk extensively. The structure of the human protease web is not cascade-like and hierarchical but multidirectional with connections between top and bottom proteins of known cascades with six proteases and inhibitors identified as key connectors in this network. Although other connectors might be identified in future versions of the network, this shows how regulatory switches, especially inhibitors, tether subnetworks of the overall network. Notably, the observed potential for regulatory crosstalk between proteases and inhibitors is not an artifact of data annotation as it persists robustly despite various perturbations we tested (Figure 5). On the contrary, the extent of such crosstalk is an underestimation because current data on protease cleavage and inhibition are largely incomplete.
As high-throughput terminomics analyses continue to massively add new information, more connections will undoubtedly be found, thereby further increasing the observed connectivity. In fact, a decrease in connectivity can only occur if current annotations are proven wrong and are corrected by removing edges from the network. However, we demonstrated that connectivity in the protease web is highly robust against such modifications, further validating the existence of a pervasive network of proteases and inhibitors embedded in different proteomes. Investigating tissue-specific implementations of the protease web, we found that gene expression shapes the protease web specifically in various tissues. Thus, subnetworks of the entire network are active at any place and time in different tissues. Some human tissues exhibit a protease web with connectivity close to the global network, further validating the existence of such a network in vivo. Mouse annotations are currently focused on few proteases and can therefore not yet display large-scale network features. Despite this and the current lower connectivity in the murine network (Figure 2C), we expect that with further annotations the murine network will morph to form more of a multidirectional, highly connected structure similar to the described human network.
The utility of the protease network as a concept and as a tool was demonstrated in successfully deciphering a paradoxical in vivo result involving cleavage of the murine chemokine LIX by neutrophils, an important inflammatory cell in innate immunity, which had been previously shown to be a substrate of the neutrophil-specific MMP8 [56]. Our analyses showed that even though MMP8 cleaves LIX in vitro and in the Mmp8−/− mouse LIX cleavage is also reduced, it was not cut by MMP8 in vivo. Rather, we identified neutrophil elastase as the relevant protease in vivo. Path finding in the protease web enabled us to then prove that MMP8 potently but indirectly facilitated LIX cleavage through direct MMP8 cleavage and inactivation of the elastase inhibitor α1-PI in cellular contexts and in vivo. Thus, combining individual interactions stored in TopFIND/MEROPS through interrogation of the protease web by random and directed walks generated a testable hypothesis that was experimentally validated. This revealed the mechanistic importance of MMP8 in mediating the cleavage of LIX—not directly as observed in vitro, but indirectly by enabling elastase activity through removal of the biologically relevant blocking inhibitor, thus forming a metallo-serpin switch to regulate the concentrations of active versus inactive α1-PI in vivo. The biological outcome of path walking in the network will depend on the relative concentrations of the individual nodes in different tissues or tissue conditions and pathologies. Thus, what is biological meaningful in one situation may not be in another and so requires experimental validation, as we performed here. Hence, the overall workflow of path prediction and validation can now be transferred to other investigations of complex in vivo protease biology.
Critical control of protease activity is exerted at the protein level. Proteases from one class (e.g., metalloproteases) frequently cleave proteases from other classes (e.g., serine proteases) or their cognate inhibitors (serpins), and subnetworks can thereby be activated or inactivated. In this process, we found that protease inhibitors take an important connecting role in the web—they are highly enriched as substrates of all classes of proteases and removal of inhibition strongly decreases reachability of all nodes in the network. Protease inhibitors often lack specificity and inhibit families of proteases rather than just individual enzymes. Thus, inhibitors function as key on/off switches of entire subnetworks within the protease web, enabling rapid and efficient activation of proteolytic processes upon their cleavage. We provided a new example of a metallo-serpin switch controlling chemokine activation. As an important biological consequence of this, removal of inhibition is therefore recognized to be as important as zymogen activation in cascades in controlling proteolysis. Indeed this was recently demonstrated in skin inflammation in vivo, where MMP2 was found to cleave and inactivate serpin G1, also known as complement C1 inhibitor [13]. Dynamically regulating the activity levels of serpin G1 inhibition allowed complement activation to cascade, which otherwise was greatly reduced in the Mmp2−/− mouse, where excess amounts of intact functional serpin G1 were proteomically quantified by TAILS terminomics. The central role of this metallo-serpin inhibitor switch in the protease web was further shown in the regulation of another subnetwork involving plasma kallikrein cleavage of kininogen to release the vasoactive peptide bradykinin. The network representation of the protease web emphasizes that proteases of one family and class can markedly regulate the activity of proteases from different families and classes.
Understanding a complex biological network, such as the protease web, can only be achieved via systematic storing and sharing of biochemical information in order to enable network-based predictions to generate testable hypotheses. Applying this strategy, we gained in silico insights into in vivo processes and validated these biochemically, in culture and in vivo. We forecast that through further identification and biochemical characterization of cleavage and inhibition events, the representation of protease interactions can be improved to strengthen its predictive power. The resulting network could then be used to simulate the effects of protease and inhibitor knockouts and protease drug targeting in disease, which will enhance confidence of targeting the correct protease and thereby increase the success rate of clinical trials by reducing unexpected side effects.
In conclusion, our analysis of the protease web reveals a multidirectional rather than a hierarchical structure, as has been proposed [40], with deep connections in regulation of the proteome by specific proteolytic processing in addition to degradation. As the structure of the human protease web is multidirectional rather than cascade-like and hierarchical, it has high connectivity that is robust to change. Biologically this implies that regulation by proteolysis is a consistent and pervasive force in all tissues. In comparison to phosphorylation, which is limited to intracellular proteins and pathways, proteolysis affects all proteins and pathways inside and outside the cell, and it is irreversible and pervasive and needs to be considered in functional analyses of the proteome.
Tables containing proteases and their substrates (cleavages) and protease inhibitors and their target proteases (inhibitions) as well as tables mapping UniProt IDs to MEROPS IDs and gene names were collected from the TopFIND MySQL database (http://clipserve.clip.ubc.ca/topfind/; downloaded January 15, 2012).
Proteases were classified based on their MEROPS IDs in TopFIND. Determining the inhibitor class specificity of human protease inhibitors was performed by downloading lists of UniProt ACs for Gene Ontology [61] annotations cysteine-type (GO:0004869, n = 49 proteins), metallo- (GO:0008191, n = 11 proteins), or serine-type (GO:0004867, n = 95 proteins) endopeptidase inhibitor from neXtProt [62] on May 24, 2012. A term “aspartic-type endopeptidase inhibitor” (GO:0019828) exists, but no proteins are annotated with this term. Inhibitors were labeled “broad” if they are annotated to inhibit more than one class of protease based on (i) their GO terms from neXtProt or (ii) their annotated inhibitions from TopFIND.
The network representation of cleavages and inhibitions was obtained via R [63] scripts, heavily relying on the use of the igraph library [64]. Proteins are represented as nodes. Cleavages are represented as directed edges from the proteases node to the substrate node. Accordingly, inhibitions were represented as directed edges from the inhibitor to the inhibited protease. Reachability of a node was calculated by counting all proteins where a shortest path can be found using the shortest.path function of igraph. Betweenness of nodes was calculated using the betweenness function of the igraph package. By recalculating betweenness after removing each node, the iterative identification of nodes with the highest betweenness was performed. Paths from MMP8 to neutrophil elastase were identified in the network using the get.all.shortest.paths function of the igraph package. Erdős-Rényi networks with the same number of nodes and edges as the original graph were generated using the erdos.renyi.game function of the igraph package, and Barabasi-Albert networks were generated with the barabasi.game function, forcing the same out-degree distribution as the protease web. Edge-shuffled random graphs were generated using the degree.sequence.game function once keeping out- and in-degree distributions the same so that each node has the same in- and out-degree as in the original network (Shuffled) and once shuffling those distributions before passing them to the method (Shuffled2). Inverse empirical cumulative distribution functions were calculated and plotted using an inverted version of the empirical cumulative function “ecdf” in R. The AUC was calculated by calling the integrate function in R on the cumulative function.
Mouse and human networks were compared by identifying connections, which occur between homologous proteins. The homology mapping between UniProt ACs of the two species was performed by mapping UniProt ACs to Ensembl protein IDs via the Ensembl database of the biomaRt package [65] in R obtained from Bioconductor [66]. The homology mapping between Ensembl protein IDs was performed using the InParanoid [67] database via the hom.Hs.inp.db [68] package in R/Bioconductor.
Network figures were plotted using Cytoscape 2.8.3 [69].
Proteins involved in selected, protease-specific biological processes were identified by obtaining Gene Ontology [61] annotation of proteins using the org.Hs.eg.db package [70] in R/Bioconductor on August 8, 2013.
N-terminal cleavage sites in normal and inflamed murine skin were obtained from Supplementary table S8 from [13].
The data for the analysis of the protease and inhibitor expression profile was achieved by analysis of commercially available RNAs from 23 different healthy human tissues on the protease- and inhibitor-specific oligonucleotide-based CLIP-CHIP microarray [71]. Data from 84 CLIP-CHIP microarrays representing biological and technical replicates of antisense RNA of these tissues were used, and average signal intensity values (A-Value) of each gene were combined. An expression cutoff was determined at an A-Value of 7.5, where 95% of the intensities of the negative oligonucleotide probes on the microarray were below this cutoff (data are available at http://clipserve.clip.ubc.ca/supplements/protease-web).
|
10.1371/journal.pntd.0000916 | The Potential Economic Value of a Trypanosoma cruzi (Chagas Disease) Vaccine in Latin America | Chagas disease, caused by the parasite Trypanosoma cruzi (T. cruzi), is the leading etiology of non-ischemic heart disease worldwide, with Latin America bearing the majority of the burden. This substantial burden and the limitations of current interventions have motivated efforts to develop a vaccine against T. cruzi.
We constructed a decision analytic Markov computer simulation model to assess the potential economic value of a T. cruzi vaccine in Latin America from the societal perspective. Each simulation run calculated the incremental cost-effectiveness ratio (ICER), or the cost per disability-adjusted life year (DALY) avoided, of vaccination. Sensitivity analyses evaluated the impact of varying key model parameters such as vaccine cost (range: $0.50–$200), vaccine efficacy (range: 25%–75%), the cost of acute-phase drug treatment (range: $10–$150 to account for variations in acute-phase treatment regimens), and risk of infection (range: 1%–20%). Additional analyses determined the incremental cost of vaccinating an individual and the cost per averted congestive heart failure case. Vaccination was considered highly cost-effective when the ICER was ≤1 times the GDP/capita, still cost-effective when the ICER was between 1 and 3 times the GDP/capita, and not cost-effective when the ICER was >3 times the GDP/capita. Our results showed vaccination to be very cost-effective and often economically dominant (i.e., saving costs as well providing health benefits) for a wide range of scenarios, e.g., even when risk of infection was as low as 1% and vaccine efficacy was as low as 25%. Vaccinating an individual could likely provide net cost savings that rise substantially as risk of infection or vaccine efficacy increase.
Results indicate that a T. cruzi vaccine could provide substantial economic benefit, depending on the cost of the vaccine, and support continued efforts to develop a human vaccine.
| The substantial burden of Chagas disease, especially in Latin America, and the limitations of currently available treatment and control strategies have motivated the development of a Trypanosoma cruzi (T. cruzi) vaccine. Evaluating a vaccine's potential economic value early in its development can answer important questions while the vaccine's key characteristics (e.g., vaccine efficacy targets, price points, and target population) can still be altered. This can assist vaccine scientists, manufacturers, policy makers, and other decision makers in the development and implementation of the vaccine. We developed a computational economic model to determine the cost-effectiveness of introducing a T. cruzi vaccine in Latin America. Our results showed vaccination to be very cost-effective, in many cases providing both cost savings and health benefits, even at low infection risk and vaccine efficacy. Moreover, our study suggests that a vaccine may actually “pay for itself”, as even a relatively higher priced vaccine will generate net cost savings for a purchaser (e.g., a country's ministry of health). These findings support continued investments in and efforts toward the development of a human T. cruzi vaccine.
| Chagas disease (American trypanosomiasis), caused by the parasite Trypanosoma cruzi (T. cruzi), is a leading etiology of non-ischemic heart disease worldwide [1] and has a substantial impact on Latin America, resulting in an estimated 750,000 productive life years lost and 1.2 billion dollars lost annually [2], [3], [4]. Chagas disease has three established phases: acute, indeterminate, and chronic. While acute disease is primarily asymptomatic, cases often transition to the chronic phase and clinically manifest as cardiomyopathy and subsequent congestive heart failure (CHF) decades after infection [1], [5], [6]. Furthermore, those who develop Chagas-related CHF have poorer prognoses and higher mortality rates than those with other CHF etiologies [2].
The substantial burden of Chagas disease and the limitations of current interventions have motivated efforts to develop a vaccine against T. cruzi. Although currently available drugs (benznidazole and nifurtimox) are moderately efficacious when administered during the acute phase, they have been minimally successful in treating chronic infection [7], [8], [9]. Low rates of symptomatic acute illness limit the utilization (and thus, the benefit) of these drugs [5], [8]. The lack of an available vaccine has left insecticide spraying for T. cruzi vectors (reduvidae insects) as the primary control strategy. However, implementing successful mass spraying can be challenging [10], [11]. Mass spraying requires both repeated and consistent reapplication, which in turn necessitates funding, personnel, and equipment. Due to a lack of national-level funding, local communities may not have the resources to maintain spraying. Furthermore, increased use of insecticides has elicited resistance among vectors in Argentina and Bolivia and may lead to untoward health effects for humans [11].
Several T. cruzi vaccine candidates have demonstrated protective effects against challenge in mouse models and offer promise for the future development of a human vaccine [12], . Much attention has focused on DNA vaccines, consisting of one or more antigen-coding plasmids, which may provide sufficient protection without the possibility of reverting back to the infectious form [12].
Understanding the potential economic and health benefits of a T. cruzi vaccine could help guide vaccine investment, development, targeting, and implementation, thereby assisting vaccine scientists, manufacturers, policy makers, and other decision makers. It can be helpful to construct economic models early in a vaccine's development, before key decisions about the vaccine are made and while important aspects of the vaccine can still be altered [14]. We developed a computer model to evaluate the economic value of a T. cruzi vaccine for the control of Chagas disease in Latin America. Different scenarios helped determine the effects of varying various key vaccine characteristics such as vaccine efficacy (to guide development), vaccine cost (to help set future price points), and risk of infection (to identify appropriate target populations).
We constructed a Markov decision analytic computer simulation model to assess the potential cost-effectiveness of a T. cruzi vaccine in Latin America using TreeAge Pro 2009 (TreeAge Software, Williamstown, Massachusetts). The model assumed the societal perspective and evaluated the economic value of vaccination versus no vaccination of a cohort of children <1 year of age to prevent T. cruzi infection and Chagas disease. Each cycle length was one year. Figure 1 illustrates the general model structure, including the following six Markov states of the disease model and an individual's possible transitions between states:
All Markov states were mutually exclusive. Individuals entered the model at age 0 and began in the ‘Susceptible/Well’ state (i.e., well with no prior history of infection). Each passing year, the individual either remained in the same state or transitioned to another state. The individual continued in the model until he or she entered the Death state from (1) acute/symptomatic infection, (2) cardiomyopathy with or without CHF, or (3) mortality unrelated to T. cruzi infection, where they remained for the remainder of the trial [5], [6].
Figure 2a–e shows the possible paths an individual could have traveled through after entering each Markov state. After entering the ‘Susceptible/Well’ state, an individual had probabilities of becoming infected (symptomatic or asymptomatic infection), dying of unrelated causes, or remaining uninfected. Only symptomatic cases had a probability of seeking T. cruzi treatment during the acute phase. Consistent with standard medical operating procedure, developing severe side effects resulted in discontinuation of drug treatment and therefore eliminating any chance the treatment could be successful. Asymptomatic cases did not receive treatment and proceeded directly to the indeterminate (latent) disease phase. The reported time from acute illness to development of chronic cardiac manifestations in the literature ranged from 10 to 30 years; infected individuals in the model therefore had to stay in the indeterminate phase for at least 10 cycles (years) before having an annual probability of developing cardiomyopathy (with or without CHF) [5], [6], [15], [16], [17]. Those who sought a form of treatment for chronic infection once continued to do so throughout the remainder of their life span.
Each simulation run sent 1,000 individuals through the model 1,000 times each for a total of 1,000,000 individual realizations. For each simulation run, the following equation computed the incremental cost-effectiveness ratio (ICER), or cost per disability-adjusted life year (DALY) avoided, through vaccination:Quantification of DALY decrements for vaccinating and not vaccinating included the years lost due to disability from Chagas (YLD) as well as the years of life lost as a result of Chagas-related mortality (YLL). Costs included both direct and indirect costs (such as productivity loss) that resulted from becoming a Chagas case. The cost-effectiveness of vaccination for each scenario was evaluated using the GDP per capita of Colombia ($5,048.41), as it represented the approximate average GDP per capita for all of Latin America [18], [19]. Vaccination was considered highly cost-effective if the ICER value was $5,048.41 per DALY avoided or less. Scenarios that yielded an ICER of $15,145.23 per DALY (3 times the GDP per capita) avoided or less still indicated that vaccination was cost-effective, while ICERs greater than 3 times the GDP per capita indicated that vaccination was not a cost-effective strategy. The cost of vaccinating an individual was calculated by comparing the average cost accrued on the vaccine arm of the model to the average cost accrued on the no vaccine arm over the 1 million trials for each simulation. The cost per avoided congestive heart failure (CHF) case was calculated by dividing the cost difference between the vaccine and no vaccine arm for the entire cohort by the total number of CHF cases avoided on the vaccine arm/branch (compared to the no vaccine branch) for the simulation.
Table 1 shows the cost, probability, and DALY model input values and their corresponding sources. The probability of an acute case being symptomatic in the model was 1%. Based on a report of treatment-seeking behavior for febrile illness, the probability that a symptomatic individual sought treatment was 34% [5], [8], [20]. For treatment costs, life expectancy, and crude mortality rates, data from Colombia, where Chagas prevalence is high, served as a proxy for Latin America [19]. World Health Organization (WHO) sources provided disability weights for cardiomyopathy with and without CHF as well as crude mortality rates (0–11 mos, 1.65%; 1–4 yrs, 0.32%; 5–9 yrs, 0.15%; 10–14 yrs, 0.18%; 15–19 yrs, 0.55%; 20–24 yrs, 0.97%; 25–29 yrs, 1.08%; 30–34 yrs, 1.02%; 35–39 yrs, 1.02%; 40–44 yrs, 1.12%; 45–49 yrs, 1.58%; 50–54 yrs, 2.11%; 55–59 yrs, 3.12%; 60–64 yrs, 4.99%; 65–69 yrs, 9.07%; 70–74 yrs, 14.66%; 75–79 yrs, 23.85%; 80–84 yrs, 30.62%; 85–89 yrs, 40.46%; 90–94 yrs, 51.70%; 95–99 yrs, 64.50%; 100 yrs and older, 100%) and life expectancies (0–11 mos, 75.5 yrs; 1–4 yrs, 75.8 yrs; 5–9 yrs, 72 yrs; 10–14 yrs, 67.1 yrs; 15–19 yrs, 62.2 yrs; 20–24 yrs, 57.5 yrs; 25–29 yrs, 53.1 yrs; 30–34 yrs, 48.6 yrs; 35–39 yrs, 44.1 yrs; 40–44 yrs, 39.5 yrs; 45–49 yrs, 35 yrs; 50–54 yrs, 30.4 yrs; 55–59 yrs, 26 yrs; 60–64 yrs, 21.8 yrs; 65–69 yrs, 17.8 yrs; 70–74 yrs, 14.3 yrs; 75–79 yrs, 11.4 yrs; 80–84 yrs, 9.2 yrs; 85–89 yrs, 7.1 yrs; 90–94 yrs, 5.2 yrs; 95–99 yrs, 3.6 yrs; 100 yrs and older, 2.5yrs) [21], [22].
Sensitivity analyses evaluated the impact of varying key model parameters such as vaccine cost (ranging from a low of $0.50 up to the cost at which the vaccine was no longer cost-effective), vaccine efficacy in preventing infection (range: 25% to 75%), the cost of acute-phase drug treatment (range: $10–$150 to account for variations in acute-phase treatment regimens), and risk of infection (range: 1%–20%) [23]. As many studies report a cost associated with routine surveillance (i.e., clinic visits, radiographs, electrocardiograms, and laboratory tests) during the indeterminate phase, additional sensitivity analyses varied the probability that an individual accrued a cost while in this phase from 25%–75% [6], [24]. Probabilistic sensitivity analyses determined the effects of simultaneously varying parameter values across their respective ranges.
Results suggest that a T. cruzi vaccine would be cost-effective across a wide range of vaccine prices and efficacies and T. cruzi infection risks. In fact, in many cases, a T. cruzi vaccine could actually save costs (i.e., that would be associated with treating the disease) in addition to providing health benefits. Vaccination remained cost-effective even up to a vaccine price of $75 when infection risk was 5% and vaccine efficacy was greater than 50% and up to a vaccine price of $200 when the vaccine efficacy was 75% or greater.
The majority of our modeled scenarios demonstrated vaccination to be very cost-effective, and in many cases highly cost-effective, especially with lower vaccine price points, compared to no vaccination. Table 2 indicates the incremental cost-effectiveness ratio (ICER) of Chagas vaccination at various vaccine costs, infection risks and vaccine efficacy rates. When the total vaccine cost was less than $5, vaccination was highly cost-effective across all scenarios of infection risk (1%, 5%, 10% and 20%) and vaccine efficacy (25%, 50%, and 75%). Increasing the vaccine price to $10 alters the vaccine strategy to cost-effective across all scenarios, and finally when the vaccine cost is $20 and vaccine efficacy is 25%, some scenarios become cost-ineffective. However, when the vaccine efficacy is greater than 50%, vaccination remains cost-effective until a $50 vaccine price point at 1% infection rate, a $75 vaccine price point at 5% infection rate, and a $100 vaccine price point at 10% infection rate. Vaccination remains cost-effective at a $200 price point when the efficacy is greater than 50% and infection risk is 20%. The model was fairly robust and displayed minimal sensitivity to variation in treatment costs for acute infection and the probability of an indeterminate phase-associated cost.
Table 3 shows that vaccinating an individual can actually be cost savings under many explored circumstances. Negative values in the table indicate that vaccinating an individual resulted in cost savings (i.e., there was actually net monetary gain from vaccinating an individual). For example, when a $1 vaccine was only 25% efficacious and the risk of infection was 1%, vaccinating an individual would on average save $1.52. At infection rates as low as 1%, vaccination was cost-saving up to a $5 vaccine price point as long as the vaccine remained at least 50% efficacious, with savings overall ranging from −$0.07 to −$81.26 per vaccinated individual. Vaccination remained cost-saving for all scenarios under a $10 vaccine price point, when the risk of infection was 5% or greater, averting the most cost (−$81) per vaccinee when vaccine price was $1, efficacy was 75%, and infection risk was 20%. Administering vaccine only resulted in net cost ($2.44–$7.26) at low infection risk (1%) and vaccine price point of $10 for the entire range of evaluated vaccine efficacies. The cost benefit persisted when the vaccine efficacy was 50% at vaccine costs of $20–$30 when infection risk was greater of equal to 10%. When the vaccine was 75% efficacious, a cost savings exists for vaccine prices of ≤$50. For a vaccine cost of $200, the incremental cost of vaccinating an individual ranged from $111.03 (20% infection rate, 75% vaccine efficacy) to $197.36 (1% infection rate, 25% vaccine efficacy).
Table 4 displays the cost per CHF case averted by vaccination. As can be seen, since the vaccine in most cases is cost savings, there is actually net savings per CHF case averted. The greatest cost savings per CHF case avoided were seen at the lowest vaccine cost ($1) and the lowest infection rates (1%). Under every scenario with a vaccine cost of ≤$10 (regardless of indeterminate phase-cost association or vaccine efficacy) and infection rates of 5% or above, the cost per CHF case averted was negative, indicating a cost savings per case avoided through vaccination. When infection rates were 1% and vaccine cost increased to $10, the cost per avoided CHF case was $50, 883 at 75% vaccine efficacy. At lower efficacies (50% and 25%), cost per CHF case averted increased to $126,009 and $660,350, respectively. At a vaccine price point of $5, vaccinating to prevent a case of CHF ranged from costing over $100,000 to saving over $100,000, depending on vaccine efficacy.
Our results support further development and future implementation of a T. cruzi vaccine and have several implications for vaccine policy makers, developers, and manufacturers. A T. cruzi vaccine could not only decrease infection burden, but also severe cardiac disease in Latin America. Our results suggest that vaccination would be highly cost-effective (and in some cases, economically dominant) over a wide range of infection rates, treatment-seeking behaviors, and vaccine costs. Such findings imply that vaccination may be beneficial even in areas with low risk of infection or in endemic areas that are improving (i.e., decreasing infection risk) over time. Demonstrating that the vaccine could provide net cost savings even at relatively higher vaccine price points (such as $50) could encourage manufacturers to pursue development and eventual commercialization of the vaccine. Additionally, as our analysis suggests that the target efficacy window for a vaccine can be fairly wide; scientists do not necessarily have to design the near “perfect” vaccine that confers protection close to 100% in order to establish economic value. Even vaccines that offer lower levels of protection can be valuable, especially at lower vaccine prices. Moreover, our study suggests that a T. cruzi vaccine may actually “pay for itself”: even a relatively higher priced vaccine of $50 can generate net cost savings for a purchaser (e.g., a country's ministry of health).
As our results demonstrate, vaccine price helps drive the net costs and cost-effectiveness of a T. cruzi vaccine. A vaccine that costs $30 or less is cost-effective under most explored conditions. However, a vaccine that costs $100 or more is cost-effective under a much more limited set of circumstances. Our results may help decision-makers map out the appropriate prices for a given situation.
A T. cruzi vaccine could affect the epidemiology of cardiac disease in endemic populations and countries. As management of cardiomyopathy and CHF are associated with intensive and costly medical care, preventing such outcomes would likely alleviate a substantial burden on the healthcare system. A study conducted by Mendez et al. reported that 20% of all cardiovascular disease patients seen in a particular Brazilian health institution had Chagas-related cardiomyopathy, and suggests that unresolved Chagas infection accounts for nearly 30% of all CHF cases in Brazil [25]. Additional literature suggests that a similar proportion of CHF cases (20%) in Colombia result from Chagas disease [2]. Chagas disease may play an even larger role in causing cardiac disease in other parts of Latin America.
Limitations of current alternatives have spurred efforts to develop a T. cruzi vaccine. Although drugs exist for the treatment of Chagas disease, they have many complications and are associated with long treatment regimens and low cure rates when administered after the acute disease phase [26]. DNA vaccines have recently shown promise against several protozoan parasitic diseases, such as malaria, leishmaniasis, and Chagas disease, as they have proven to be safe and elicit a complete immune response. Furthermore, their thermo-stability and affordability may make them highly practical for use in resource poor settings where these diseases are endemic. While this technology has yet to yield an efficacious Chagas vaccine for the human population, the development and licensure of DNA vaccines against West Nile and infectious hematopoietic necrosis viruses in animals give promise for a future human vaccine [12]. Certainly bringing a T. cruzi vaccine to market would involve surmounting a variety of scientific hurdles, but our study suggests that surmounting such obstacles could be very worthwhile [12].
In developing our model, we attempted to remain conservative about the benefits of a vaccine. The actual costs of diagnosing and treating CHF may be much higher than the numbers used. In addition, CHF could have substantial repercussions on an individual's life which were not captured by the model. The model assumed that individuals were otherwise healthy, but the addition of co-morbidities or co-infection with other pathogens such as the human immunodeficiency virus (HIV) could worsen outcomes for individuals infected with T. cruzi. Moreover, our model did not consider how the vaccine may reduce the transmission of T. cruzi by reducing available human hosts. The risk of infection may be higher than reported, as many cases go undiagnosed. Our range of infection probabilities came from Global Health Statistics reports but may not include regions and populations with higher risk [2], [25], [27]. Some studies have reported the direct progression from acute to chronic disease as well as sudden death due to cardiac complications during the indeterminate phase; however, these possibilities were difficult to measure accurately and therefore not incorporated into the model parameters [5].
All computer models are simplifications of real world scenarios and therefore cannot capture all possible outcomes of Chagas disease or the efficacy of concurrent existing regional control methods. Our model focuses on the individual rather than a population and therefore, does not explicitly represent herd immunity. However, increasing herd immunity could decrease Chagas risk, and our study demonstrates how changes in Chagas disease risk would affect the vaccine's economic value. Additionally, this model cannot account for the variation in high risk exposure resulting from factors such as environmental conditions or individual behaviors across Latin America. Although model assumptions and data inputs were drawn from an extensive review of the literature, sources may vary in quality and input values may not hold under all conditions.
Our model suggests that introducing a T. cruzi vaccine to Latin America would provide economic value. Such a vaccine could be highly cost-effective and many cases could be economically dominant, providing both cost savings and health benefits. Even a vaccine with fairly low efficacy (25%) can provide cost savings. Moreover, a vaccine could be cost-effective even at relatively low infection risks (1%). Such findings support continued efforts to develop a human vaccine against T. cruzi to help reduce the significant burden of Chagas disease.
|
10.1371/journal.pbio.1000484 | A Key Commitment Step in Erythropoiesis Is Synchronized with the Cell Cycle Clock through Mutual Inhibition between PU.1 and S-Phase Progression | Hematopoietic progenitors undergo differentiation while navigating several cell division cycles, but it is unknown whether these two processes are coupled. We addressed this question by studying erythropoiesis in mouse fetal liver in vivo. We found that the initial upregulation of cell surface CD71 identifies developmentally matched erythroblasts that are tightly synchronized in S-phase. We show that DNA replication within this but not subsequent cycles is required for a differentiation switch comprising rapid and simultaneous committal transitions whose precise timing was previously unknown. These include the onset of erythropoietin dependence, activation of the erythroid master transcriptional regulator GATA-1, and a switch to an active chromatin conformation at the β-globin locus. Specifically, S-phase progression is required for the formation of DNase I hypersensitive sites and for DNA demethylation at this locus. Mechanistically, we show that S-phase progression during this key committal step is dependent on downregulation of the cyclin-dependent kinase p57KIP2 and in turn causes the downregulation of PU.1, an antagonist of GATA-1 function. These findings therefore highlight a novel role for a cyclin-dependent kinase inhibitor in differentiation, distinct to their known function in cell cycle exit. Furthermore, we show that a novel, mutual inhibition between PU.1 expression and S-phase progression provides a “synchromesh” mechanism that “locks” the erythroid differentiation program to the cell cycle clock, ensuring precise coordination of critical differentiation events.
| Hematopoietic progenitors that give rise to mature blood cell types execute simultaneous programs of differentiation and proliferation. One well-established link between the cell cycle and differentiation programs takes place at the end of terminal differentiation, when cell cycle exit is brought about by the induction of cyclin -dependent kinase inhibitors. It is unknown, however, whether the cell cycle and differentiation programs are coordinated prior to cell cycle exit. Here, we identify a novel and unique link between the cell cycle clock and the erythroid (red blood cell) differentiation program that takes place several cell division cycles prior to cell cycle exit. It differs from the established link in several respects. First, it takes place at the onset, rather than at the end, of erythroid terminal differentiation, preceding the chromatin changes that enable induction of red cell genes. Second, it is initiated by the suppression, rather than the induction, of a cyclin -dependent kinase inhibitor. It therefore causes the cell to enter S-phase, rather than exit the cell cycle. Specifically, we found that there is an absolute interdependence between S-phase progression at this time in differentiation, and a key commitment step in which, within a short few hours, cells become dependent on the hormone erythropoietin, undergo activating changes in chromatin of red cell genes, and activate GATA-1, the erythroid master transcriptional regulator. Arresting S-phase progression at this time prevents execution of this commitment step and subsequent induction of red cell genes; conversely, arresting differentiation prevents S-phase progression. However, once cells have undergone this key commitment step, there is no longer an interdependence between S-phase progression and the induction of erythroid genes. We identified two regulators that control a “synchromesh” mechanism ensuring the precise locking of the cell cycle clock to the erythroid differentiation program during this key commitment step.
| Hematopoietic progenitors execute a cell division program in parallel with a differentiation program in which lineage choice is followed by lineage-specific gene expression. In many differentiation models, cell cycle exit, driven by cyclin-dependent kinase inhibitors (CDKI), is a prerequisite for terminal differentiation, establishing a key interaction between the cell cycle and differentiation programs [1]–[3]. However, it is unclear how the cell cycle and differentiation programs might be linked prior to cell cycle exit. Such links are presumably required to ensure the correct number of differentiated progeny. In addition, it has been speculated that the reconfiguration of chromatin at sites of lineage-specific genes, a necessary step preceding lineage-specific gene expression, may be innately dependent on DNA replication [4],[5]. An intriguing possibility is that the clockwork-like mechanisms regulating orderly cell cycle transitions may also be used, in the context of differentiating cells, to coordinate key steps in differentiation.
Here we studied differentiation of the enucleated red blood cell lineage, which first arises from hematopoietic stem cells in the fetal liver on embryonic day 11 (E11). It replaces a transient, nucleated yolk-sac erythrocyte lineage and persists throughout life. Although many of the committal events that lead to the erythroid phenotype are known, their precise timing in erythroid differentiation, and the manner in which they are coordinated with each other and/or with the cell cycle machinery, is poorly understood. Thus, survival of erythroid progenitors requires both the hormone erythropoietin (Epo), and its receptor, EpoR, a class I cytokine receptor expressed by erythroid progenitors [6]. However, the precise time in erythroid differentiation when progenitors become dependent on Epo had not been defined. The master transcriptional regulator GATA-1 is responsible for the erythroid gene expression profile, in combination with a number of additional transcriptional regulators, including FOG-1, EKLF, SCL/Tal-1, LMO2, Ldb1, E2A, and Zbtb7a [7]–[10]. Though GATA-1 functional activation must precede erythroid gene induction, its precise timing in primary differentiating progenitors is not known. GATA-1 functions are antagonized by PU.1, an Ets transcription factor that acts as a master regulator in the myeloid and B-cell lineages. The mutual inhibition between PU.1 and GATA-1 is thought to underlie cell fate choice in multipotential progenitors [11]–[14]. PU.1 has been implicated in erythroleukemia [13],[15], but its physiological function in erythropoiesis is not known.
Erythroid gene induction by GATA-1 requires an “open chromatin” conformation in the vicinity of erythroid-specific genes. The erythroid-specific β-globin locus is one of the best studied models of lineage-specific gene expression [16],[17]. The active locus is characterized by early replication during S-phase, higher sensitivity to DNase I digestion, low levels of DNA methylation, and post-translational histone tail modifications associated with actively transcribed genes. Conversely, the same locus in non-erythroid cells is DNase I resistant, replicates late in S-phase, and contains histone tail modifications characteristic of silent chromatin. In spite of the detailed knowledge contrasting chromatin states in erythroid cells with non-erythroid cells, the precise time during erythroid differentiation when chromatin reconfiguration occurs is not known. Furthermore, it is not known whether this reconfiguration involves a number of sequential stepwise alterations occurring over a number of cell cycles/differentiation stages or whether the many changes entailed in chromatin activation occur simultaneously.
Here we studied erythroid differentiation using a flow-cytometric assay that identifies sequential stages in erythroid differentiation directly within primary hematopoietic tissue. We found that in mouse fetal liver in vivo, upregulation of CD71 marks cells that are synchronized in S-phase of a single cell cycle, corresponding to the last generation of erythroid colony-forming cells, approximately three cell cycles prior to terminal cell cycle exit. A number of differentiation milestones, whose precise timing in erythroid development was previously unknown, occur during early S-phase of this cycle. These include the onset of Epo dependence, activation of GATA-1 function, and the opening up of chromatin at the β-globin locus. We show that S-phase progression during this specific cell cycle is dependent on downregulation of p57KIP2 and is required for execution of these differentiation milestones, including the reconfiguration of chromatin at the β-globin locus. Further, this S-phase dependent rapid differentiation transition is regulated by PU.1 through a newly identified, mutual antagonism between S-phase progression and PU.1 expression that coordinates the precise locking of the differentiation program to the cell cycle clock as cells enter a terminal differentiation phase.
Mouse fetal liver between E11 and E15 is primarily an erythropoietic tissue. Cell surface markers CD71 and Ter119 may be used to identify differentiation-stage specific subsets, directly in primary tissue [18]–[20]. Here we divided freshly harvested fetal liver cells into six CD71/Ter119 subsets that we termed S0 to S5 and that form a developmental sequence (Figure 1A). Cells isolated from subsets S1 to S5 show morphological features characteristic of erythroid maturation, including decreasing cell and nuclear size, nuclear condensation, and hemoglobin expression (Figure 1A, right panel). The precise proportion of fetal liver cells within each of the CD71/Ter119 subsets is a function of embryonic age, with the majority of cells being in the early, S0 and S1 subsets in E12. The more mature, S3 to S5 subsets are gradually populated with cells during subsequent embryonic days (E13 to E15) [20].
The EpoR−/− fetal liver is small and lacks morphologically identifiable hemoglobinized erythroblasts of the enucleated (definitive) lineage [6]. Here we found that EpoR−/− fetal liver does not contain subsets S1 to S5 (Figure 1B). This suggested that in the definitive erythropoietic lineage that gives rise to adult-type enucleated red cells, EpoR becomes essential on or prior to the transition from S0 to S1; subsets S1 to S5 are composed almost entirely of Epo-dependent erythroblasts. Of note, the small number (≈5%) of Ter119+ cells in the EpoR−/− fetal liver are all nucleated erythrocytes of the transient yolk-sac (primitive) lineage (Figure S1A).
Erythroid progenitors have traditionally been identified by their in vitro colony-forming potential. “Colony forming unit-erythroid” (CFU-e) are defined as cells that give rise to colonies containing 8 to 32 hemoglobinized cells after 2–3 days of in vitro culture in Epo [21]. We investigated the colony-forming potential of cells sorted from each of the S0 to S3 subsets (Figure 1C). CFU-e potential was exclusive to S0 and S1 and was lost with the transition to S2. Cells in S2 and S3 gave rise to small, 2 to 4 cell clusters (Figure 1C, right panel).
The frequency of CFU-e obtained from sorted S0 cells was 65%–70% of the frequency from sorted S1 (Figure 1C). S1 consists entirely of Epo-dependent cells of similar maturation, with CFU-e potential (Figure 1A–C). Assuming similar plating efficiency for sorted S0 and S1 (of ≈30%, Figure 1C), this suggested that CFU-e make up 65%–70% of the S0 subset. This is in agreement with our finding that fetal liver cells expressing non-erythroid lineage markers, which were limited to S0, formed up to 30% of this subset (Figure 1D, Figure S1B). Non-erythroid colony-forming progenitors were also restricted to S0, where they formed less than 5% of all colony-forming cells (Figure 1C). Our conclusion that 65%–70% of S0 cells are CFU-e was further supported by single cell RT-PCR, which showed that 68% of S0 cells expressed EpoR mRNA (Figure S1C).
In all the experiments that follow, “S0” refers to S0 cells from which cells expressing non-erythroid markers were excluded by flow-cytometric gating or sorting.
To examine the cell cycle status of erythroid subsets S0 to S5 in vivo, we injected pregnant female mice with the nucleotide analogue bromodeoxyuridine (BrdU) and harvested fetal livers 30 min post-injection. We sorted cells from each of S0 to S5 and stained them with antibodies directed at BrdU (Figure 1E,F). Cells that incorporated BrdU were in S-phase of the cell cycle at the time of harvesting. Subsets S4 to S5 showed a rapid decline in the number of S-phase cells, consistent with cell cycle exit of terminally differentiating cells. Unexpectedly, we noted that ≈90% of S1 cells were in S phase, as compared with ≈50% of cells in S0 (Figure 1E,F). In addition, the intensity of the BrdU fluorescence within S1 cells was approximately 50% higher than in S0, suggesting a higher rate of DNA synthesis (Figure S1D). Similar experiments with EpoR−/− fetal liver showed that EpoR appears to have no effect on progenitor cell cycle status (Figure S1E).
Consistent with the higher number of S-phase cells in S1, we found a corresponding increase in the E cyclins in S1 compared with S0 (Figure 1G). Strikingly, we noted >30-fold decrease in the CDKI p57KIP2 mRNA, but no significant change in the mRNA of other members of the CIP/KIP CDKI family; there was induction in p27KIP1 later in differentiation, in subsets S2 and S3 (Figure 1G and Figure S1F) [22],[23]. The p57KIP2 protein also decreased at the S0 to S1 transition (Figure 1G lower panel).
The finding that nearly all S1 cells were in S-phase could be due to an unusual cell division cycle with short or no gap phases. Alternatively, S1 cells may be synchronized in S-phase of the cycle. The latter explanation would require that cells spend only a brief period of a few hours in S1, lasting through part or all of a single S phase. The preceding G1 phase of this same cell cycle would have occurred prior to the transition from S0 to S1. The G2 and M phases of this same cycle would occur as cells upregulate Ter119 and transition into S2.
To investigate these possibilities, we isolated S0 cells by flow-cytometry, labeled them with the cell-tracking dye carboxyfluorescein diacetate succinimidyl ester (CFSE), and followed their Epo-dependent differentiation into S1 in vitro (Figure 1H). By 10 h, 53% of S0 cells transitioned into S1 in the absence of cell division, as indicated by a single CFSE peak for S1 (solid red histogram, t = 10 h) that was identical in intensity to that of the CFSE peak for S0 (blue histogram, t = 10 h; median CFSE fluorescence for both S1 and S0 peaks = 4,400). This suggested that the transition from S0 to S1 occurred in the absence of cell division, within a single cell cycle. Four hours later, at t = 14 h, essentially all S1 cells had divided once, as indicated by the halving of the CFSE signal (red histogram at t = 14 h, CFSE fluorescence = 2,100). The simultaneous division of S1 cells suggested they were synchronized in their cell cycle phase. By contrast, only a portion of S0 cells, which were presumably asynchronous in their cell cycle phase, had divided at this time, resulting in a biphasic CFSE peak (blue histogram, t = 14 h).
Taken together, these results suggest that the most mature CFU-e progenitor (“CFU-e.2”, Figure 1I), capable of giving rise to an eight-cell colony, traverses S0, S1, and enters S2 within a single cell cycle. This progenitor arises in S0, becomes Epo dependent, and upregulates CD71, transitioning into S1 during S-phase of its cell cycle. Upregulation of Ter119 occurs at approximately the same time that it completes its cycle and divides, giving rise to progeny that lack CFU-e activity in S2 (Figure 1C). These conclusions are consistent with essentially all S1 cells being in S-phase (Figure 1E,F), and with our finding that nearly all S1 cells are sensitive to hydroxyurea, a drug that specifically targets S-phase cells (Figure S2A). These conclusions are consistent with a number of other observations: the loss of CFU-e activity with Ter119 expression (Figure 1C, [24]), the short time span (<15 h) that freshly sorted S0 cells require to transition through S1 and into S2 (compare with an estimated cell cycle length of 16 h for a CFU-e cell that will undergo three cell divisions in 48 h, giving rise to an eight cell colony), and with early work suggesting that Epo dependence first occurs in early S-phase of a specific CFU-e cell generation [25]. These conclusions are also consistent with the finding that EpoR−/− embryos have normal numbers of CFU-e [6]: though EpoR−/− embryos lack S1 cells, all the CFU-e in S1 first arise as Epo-independent cells in S0, where they are presumably retained in the EpoR−/− fetal liver.
There are two ways to explain how upregulation of CD71, a differentiation event, might coincide with S-phase, a cell cycle event. These events may have each been initiated in parallel by a common upstream regulator, such as the EpoR, since both occur at the time that cells become EpoR dependent. Alternatively, there may be a direct mechanistic link between the differentiation and cell cycle programs. To distinguish these possibilities, we examined whether a block to S-phase progression would interfere with CD71 upregulation (Figure 2). We incubated sorted S0 cells in vitro for 10 h in the presence of Epo, and either in the presence or absence of aphidicolin, an inhibitor of DNA polymerase that arrests S-phase progression [26]. At t = 10 h, cells were washed free of aphidicolin and incubated in Epo alone for an additional 10 h (Figure 2A). In the initial 10 h of incubation, there was an Epo-dependent transition of cells from S0 to S1 (Figure 2C, rows 1 and 5). However, the presence of aphidicolin blocked this transition (Figure 2C, rows 2 & 3, t = 10 h). Both S-phase and the transition into S1 resumed once the cells were washed free of aphidicolin (Figure 2C, rows 2 & 3, t = 20 h). These observations suggested that the transition from S0 to S1 occurred during S-phase and required both Epo and S-phase progression.
We also examined the effect of mimosine, a plant amino acid that blocks cell cycle progression in late G1 [27]. We incubated sorted S0 cells in Epo and in the presence or absence of mimosine. By 4 h of incubation, the majority of cells were arrested in G1. However, a small fraction of cells (12%) could be seen in S-phase at t = 4 h (Figure 2C, row 4, BrdU/7AAD at t = 4 h). Presumably, at the time mimosine was added, these cells were advanced in their cell cycle beyond the point at which mimosine exerts its block. BrdU/7AAD analysis showed that these cells were in the early half of S-phase and expressed the highest CD71 levels within the S0 subset (Figure 2D, cells marked in red). By t = 10 h, no S-phase cells were seen in S0, presumably because they have now transitioned into S1, where a similar number of cells (15%) had newly appeared (Figure 2C, row 4, BrdU/7AAD for S0 at t = 10 h, and CD71/Ter119 for S1 at t = 10 h). These observations were consistent with the onset of CD71 upregulation occurring in early S-phase in S0, culminating in the transition to S1 later within that same S-phase.
CD71, the transferrin receptor, is required during erythroid differentiation in order to facilitate cellular uptake of iron for hemoglobin synthesis. CD71 is also expressed, albeit at lower levels, on all cycling cells. We therefore examined whether, in the context of S1 cells, CD71 might be required specifically for S-phase progression. We used RNAi to prevent CD71 upregulation in S0 cells during their incubation in Epo (Figure S2B,C). The failure of these cells to upregulate CD71 did not interfere with the number of cells in S-phase (Figure S2B). Therefore, the link between S-phase progression and CD71 upregulation in S1 cells is not due to a cell cycle function for this gene.
To investigate the link between S-phase and the erythroid differentiation program, we examined expression of erythroid transcriptional regulators and erythroid-specific genes in freshly sorted fetal liver subsets and in fetal brain (Figure 3A). We found that the GATA-1 mRNA was present in S0 cells, at 200-fold higher levels than in fetal brain (Figure 3A) and 40-fold higher level than in Mac-1+ cells (Figure S3A). It increased a further ≈2-fold with the transition from S0 into S1 and continued to increase in S2 and S3. Of note, total RNA per cell decreased 4-fold over the course of differentiation from S2 to S4 (Figure S3B), suggesting an overall modest increase in GATA-1 mRNA per cell over this period. Other erythroid transcriptional activators and GATA-1 associated factors, including EKLF, NF-E2 [28], SCL/Tal-1, and Lmo2, showed a similar expression pattern to that of GATA-1 (Figure 3A). Therefore, expression of GATA-1 and of other activators of the erythroid transcriptional program precedes the transition from S0 to S1. By contrast, we found that PU.1, a repressor of GATA-1 function, and GATA-2, a target of GATA-1-mediated repression [29], were both downregulated ≈30-fold and ≈20-fold, respectively, at the S0 to S1 transition, becoming undetectable with further differentiation (Figure 3A). Prior to its downregulation, the level of PU.1 in S0 cells was comparable to that of myeloid Mac-1+ cells (Figure S3A). PU.1 protein levels also declined with the transition from S0 to S1 (Figure S3C).
EpoR−/− fetal liver cells, though apparently arrested at the S0 stage (Figure 1B), have a similar expression pattern of transcriptional regulators to wild-type S1 (Figure 3A). Therefore, downregulation of PU.1 and GATA-2 at the S0 to S1 transition, as well as the preceding induction of GATA-1, are independent of EpoR signaling.
We examined expression of several erythroid-specific GATA-1 target genes: β-globin (Hbb-b1); the first enzyme of heme synthesis, aminolevulinic acid synthase 2 (ALAS2); and the anion exchanger Band 3 (Slc4a1), a major erythrocyte membrane protein [30]. There was a modest increase in their expression at the S0 to S1 transition, followed by a 30–100-fold induction during subsequent differentiation in S2 and S3 (Figure 3A). Expression of the EpoR gene, itself a GATA-1 target, increased 10-fold above its S0 level with the transition to S1 (Figure S3D). Taken together, induction of erythroid GATA-1 target genes and repression of GATA-2 suggest that GATA-1 function is activated at the S0 to S1 transition. The modest increase in GATA-1 mRNA at this time suggests that its activation may be principally a result of PU.1 downregulation.
We had found that S-phase progression at the transition from S0 to S1 was required for CD71 upregulation (Figure 2). We therefore examined whether S-phase progression at this time was also required for induction of erythroid-specific genes. We cultured sorted S0 cells in Epo for 10 h, a period sufficient for 25%–50% of cells to transition into S1 (Figures 1H, 2C), and examined the effect of adding aphidicolin to the culture. Cells were then washed free of aphidicolin, continuing incubation in Epo alone. Cells incubated in Epo alone for the entire period showed ≈50- to 100-fold induction in the mRNAs for β-globin, Band 3, and ALAS2 (Figure 3B, red curves). By contrast, cells that were subject to aphidicolin treatment during the initial 10 h showed reduced mRNA induction by the end of the culture period (Figure 3B, blue curves). The reduced mRNA levels corresponded closely to the levels predicted had there been a 10 h delay in the time course of induction for each of the genes (Figure 3B, black curves). Therefore, induction of erythroid-specific genes was likely blocked during the incubation period in aphidicolin.
We also examined whether S-phase arrest interferes with erythroid gene induction if applied at the S1 stage of differentiation. We sorted S1 cells and incubated them in Epo, either in the presence or absence of aphidicolin. Unlike S0 cells, aphidicolin-mediated S-phase arrest of S1 did not interfere substantially with their induction of erythroid specific genes, as shown by the unperturbed induction of β-globin, Alas2, and Band 3 (Figure 3C, Figure S3E) or with the upregulation of Ter119 (Figure S3F). Therefore, S-phase progression is required for activation of erythroid-specific genes, specifically at the S0 to S1 transition, but not a few hours later when the cells have traversed into S1. The lack of effect of aphidicolin on mRNA induction in S1 suggests its effects in S0 are not due to non-specific suppression of transcription.
Transcripts for PU.1 and GATA-2 are markedly downregulated at the transition from S0 to S1 (Figure 3A). We examined whether S-phase arrest interferes with their downregulation. Sorted S0 cells were incubated in Epo for 4 h, at which time, just prior to their transition into S1 (Figure 2C), aphidicolin was added to the cultures for a period of 10 h. Cells were then washed free of aphidicolin and incubated in Epo for a further 10 h. Aphidicolin halted the downregulation of both PU.1 and GATA-2, which resumed once the cells were washed free of the drug (Figure 3D,E). Similar results were obtained in cells treated with mimosine (Figure S3G). Therefore, S-phase progression is required for downregulation of PU.1 and GATA-2 at the S0 to S1 transition. Of note, GATA-1, Nfe2, and Lmo2 mRNAs, which did not change significantly during the transition from S0 to S1 (Figure 3A), were not altered significantly by the aphidicolin treatment (Figure 3E, Figure S3H).
We also examined the effects of aphidicolin or mimosine treatment on morphological maturation of S0 cells cultured in Epo. Following 10 h in Epo in the presence of aphidicolin or mimosine, cells appeared larger than cells incubated in Epo alone. This suggested that, while S-phase progression and the erythroid differentiation program had both arrested, cell growth was not perturbed (Figure 3F). Cells were then washed free of aphidicolin or mimosine and cultured in Epo alone. By 20 h, erythroid maturation had resumed in cells that were initially incubated in cell cycle blocking drugs, as judged by decreasing cell size, nuclear condensation, and decreased nuclear to cytoplasmic ratio, but was nevertheless delayed when compared with control cells. These results are consistent with the effect of S-phase arrest on gene expression (Figure 3B,D,E) and suggest that S-phase progression at the S0 to S1 transition is a key requirement for activation of the erythroid differentiation program.
Expression of p57KIP2 mRNA decreases over 30-fold at the S0 to S1 transition, and this is associated with downregulation of the p57KIP2 protein (Figure 1G). To examine the effect of preventing p57KIP2 downregulation, we generated a point mutant of p57KIP2, p57T329A, analogous to a proteolysis-resistant human p57KIP2 mutant [31]. Sorted S0 cells were infected with bicistronic retroviral vectors expressing either wild-type p57KIP2 or p57T329A, linked through an internal ribosomal entry site (IRES) to a human CD4 (hCD4) reporter; control cells were infected with retroviral vector expressing the IRES-hCD4 construct only (MICD4). To allow expression of the transduced p57KIP2, infected cells were cultured for 15 h in stem-cell factor (SCF) and interleukin 3 (IL-3), cytokines that sustain viability of progenitors but, unlike Epo, do not support differentiation from S0 to S1. Infected S0 cells were then transferred to Epo for 14 h (Figure 3G). Expression of either wild-type (unpublished data) or mutant p57KIP2, but not expression of MICD4, resulted in a block to S-phase progression and inhibited the transition from S0 to S1 (Figure 3G). Further, PU.1 mRNA was >3-fold higher in cells expressing p57KIP2 compared with control cells expressing vector only (Figure 3H), suggesting that, as in the case of aphidicolin-mediated S-phase arrest, p57KIP2-mediated S-phase arrest prevents downregulation of PU.1 at the transition from S0 to S1. Erythroid morphological maturation, but not cell growth, of p57T329A-transduced cells was also arrested (Figure S3I).
Taken together, upregulation of CD71, which defines the transition from S0 to S1, identifies a key differentiation transition within the last generation of CFU-e (“CFU-e.2”, Figure 3I). It marks the onset of EpoR dependence and occurs exclusively during S-phase of the cell cycle. Induction of GATA-1 and other activators of the erythroid transcriptional program precedes this transition, whereas induction of erythroid-specific genes such as β-globin and Ter119 follows it. The S0 to S1 transition coincides with rapid downregulation of p57KIP2, PU.1, and GATA-2. Both Epo and S-phase progression are required for upregulation of CD71. S-phase progression at the S0 to S1 transition requires the downregulation of p57KIP2 and is in turn required for the downregulation of PU.1 and GATA-2 and the subsequent activation of erythroid-specific genes. By contrast, S-phase arrest in S1 cells does not affect erythroid gene activation (Figures 3C, S3E–F).
Both PU.1 and GATA-2 were rapidly and dramatically downregulated at the transition from S0 to S1 (Figure 3A, Figure S3A). We examined the effect of preventing this downregulation by expressing either PU.1 (Figure 4A–D) or GATA-2 (Figures 4E, S4C,D) in S0 cells using retroviral constructs and a similar strategy to that described above for p57KIP2. Following infection, S0 cells were cultured for 15 h in IL-3 and SCF and then transferred to Epo for 24 h, when CD71/Ter119 and cell cycle profiles were examined (Figure 4A–C). We divided the PU.1 expression profile at t = 24 h into 7 sequential hCD4 gates labeled (i) to (vii) (Figure 4A), each containing cells with increasing levels of the hCD4 reporter and, therefore, increasing levels of PU.1. By measuring PU.1 protein directly in fixed and permeabilized cells using a PU.1-specific antibody and flow-cytometry, we found that hCD4 protein expression was a reliable reporter of exogenous PU.1 protein expression in our system (Figures 4D, S4A–B); expression of transduced PU.1 was also measured by qPCR (Figure S4D). Sequential hCD4 gates were also obtained for control cells expressing the empty MICD4 vector. PU.1 expression blocked transition from S0 to S1, with the number of cells transitioning into S1 declining as PU.1 expression increased (Figure 4B, upper panels). PU.1 expression also resulted in a decrease in the number of S-phase cells, with cells arresting principally at the transition from G1 to S-phase, though there was also an increase in the number of cells within G2 or M (Figure 4B, lower panels). The decrease in the number of cells in S1 was paralleled by decreased S-phase cell number, suggesting a direct correlation between the PU.1-mediated block of the transition from S0 to S1, and its inhibitory effect on S-phase (Figure 4C). Therefore, PU.1 inhibits both S-phase and erythroid differentiation at the S0 to S1 transition.
Since the downregulation of both PU.1 and p57KIP2 are required for S-phase progression and for the transition from S0 to S1 (Figure 3G, Figure 4B,C), we examined whether PU.1 may be a regulator of p57KIP2. However, we found that exogenous expression of PU.1 did not prevent downregulation of p57KIP2 (Figure S4E). Therefore, PU.1's inhibitory effect on S-phase is not mediated via p57KIP2.
In contrast to PU.1, expression of GATA-2 in S0 cells did not prevent transition into S1, though it somewhat reduced the subsequent transition from S1 to S2 (Figure 4E). GATA-1 overexpression in S0 cells had the opposite effect, of promoting the transition from S1 to S2. There was no significant effect of either GATA-1 or GATA-2 on the cell cycle profile (Figure 4E).
A long-standing hypothesis suggests that DNA replication may provide an opportunity for the restructuring of chromatin at tissue-specific gene loci [4],[5]. Given the requirement for DNA replication for the transition from S0 to S1, we asked whether chromatin change may be taking place at this time. The β-globin gene locus (Figure 5A) is a well-studied model of tissue-specific gene expression. The features that characterize the open chromatin conformation at the actively transcribed locus in erythroid cells have been established, but the time during development when the active chromatin conformation is acquired is not known. We therefore set out to examine whether the S0 to S1 transition might coincide with an alteration in the structure or function of chromatin at this locus.
The timing of replication of the β-globin locus is correlated with its chromatin state. In higher eukaryotes the timing of replication of genes correlates with their transcriptional activity [32]. Housekeeping genes replicate early in S-phase, whereas silent chromatin and heterochromatin replicate late. The β-globin locus replicates in mid to late S-phase in non-erythroid cells, and early in S-phase in erythroid cells [33]. We examined the timing of replication of the β-globin locus in S0 and S1 cells sorted from fresh fetal liver. Individual alleles were identified using fluorescence in situ hybridization (FISH) with a probe directed at the β-major gene. Cells in S-phase were identified by positive staining for BrdU incorporation. Nuclei from at least 100 S-phase cells from either S0 or S1 were examined in each of two experiments (Figure 5B). Using this approach, two single dots (“SS”) suggest that neither of the β-globin alleles had yet replicated. Nuclei in which both alleles have replicated contain a pattern of two double dots (“DD”). Replication of only one allele results in one single and one double dot (SD) [33]. We found that the number of cells with a DD pattern increased from only 15% in S0 to over 50% in S1 (Figure 5B), suggesting a switch in the timing of replication from late to early S-phase. In addition, an average of 36% of S0 cells, but only 21% of S1, had an SD pattern, consistent with a switch from late, asynchronous replication in S0 to early, synchronous replication in S1 [33].
A key indicator of open chromatin at the β-globin LCR is the presence of hypersensitivity (HS) sites (Figure 5A). We prepared nuclei from freshly sorted S0 or S1 cells and tested their sensitivity to DNase I digestion. Following digestion, we measured remaining DNA using quantitative PCR, with amplicons within HS2, HS3, and HS4 [34]. Results were expressed as a ratio to the DNase I resistant, non-expressing neural gene, Nfm. We found that S0 cells were relatively resistant to DNase I, while S1 cells were hypersensitive at all tested HS sites (Figure 5C). Therefore, the S0 to S1 transition coincides with the onset of DNase I hypersensitivity at the β-globin LCR.
We also examined E12.5 EpoR−/− whole fetal livers, which do not contain S1 cells (Figure 1B). We found that EpoR−/− fetal livers were resistant to DNase I, whereas whole fetal livers from wild-type or heterozygous littermates showed the expected hypersensitive sites (Figure 5D). We therefore concluded that DNase I hypersensitivity develops at the S0 to S1 transition, synchronously with the onset of EpoR dependence.
Since the transition from S0 to S1 coincides with, and requires, S-phase progression, we examined whether development of DNase I hypersensitivity at the β-globin LCR also requires S-phase progression. We incubated sorted S0 cells in Epo, in the presence or absence of aphidicolin, for 10 h. Over this period 25%–50% of S0 cells transition into S1, a process arrested by aphidicolin (Figures 1H, 2C). At the end of a 10-h incubation period, nuclei were prepared and digested with varying concentrations of DNase I. There was a clear increase in DNase I sensitivity in cells incubated in Epo alone, relative to cells incubated in Epo and aphidicolin (Figure 5E). Therefore, the development of DNase I hypersensitivity at the S0 to S1 transition is dependent on S-phase progression.
The switch in timing of replication and in DNase I hypersensitivity at the S0 to S1 boundary suggested the β-globin LCR was undergoing structural changes. To investigate these, we used chromatin immunoprecipitation (ChIP) to determine specific histone tail modifications at the β-globin LCR in freshly sorted S0, S1, and in fetal brain. We used ChIP-qPCR for amplicons at the β-globin LCR HS sites, or at a control, neural gene, Nfm. Changes in histone modifications were expressed as a ratio, between S0 and either S1 or fetal brain (Figure 5F,G). Figure 5F summarizes data pooled from seven experiments with various immunoprecipitating antibodies as indicated. A comparison of S1 with S0 shows a 7-fold decrease in trimethylation of histone 3 lysine 27 (H3K27me3, p = 0.019, paired t test), a mark associated with silent chromatin, and a 2.5-fold increase in histone 3 lysine 4 dimethylation, a mark associated with active chromatin (H3K4me2, p = 0.032), at the HS2 site of the β-globin LCR. A similar trend for these two modifications was also found at other HS sites (p = 0.0006 and p = 0.011 for H3K4me2 and H3K27me3, respectively, pooling all HS sites). An increase in acetyl marks in histones H3 and H4 associated with active chromatin was also seen consistently across the HS sites tested, though it did not reach statistical significance. Of note, no significant changes in histone marks were found between S0 and S1 at the Nfm gene. Further, there was no significant change in total histone occupancy of the HS sites between S0 and S1, as determined by ChIP with antibodies directed against total H3 and H4 (Figure 5F).
We noted that H3K27me3, associated with silent chromatin, and H3K4me2, associated with active chromatin, were both enriched in S0 compared with fetal brain (Figure 5G, lower panel). These results were suggestive of bivalent chromatin at the β-globin LCR in S0, and loss of the repressive H3K27me3 mark with transition into S1 (Figure 5G, upper panel, 5F).
We examined DNA methylation of six CpG dinucleotides, three each at the HS1 and HS2 sites of the β-globin LCR (Figures 5A, 6A). Genomic DNA was prepared from sorted hematopoietic cell subsets from fresh fetal liver, including S0, S1, megakaryocytic CD41+, myeloid Mac-1+, and Lin−Sca1+Kit+ (LSK) cells, enriched for hematopoietic stem-cells. We also examined EpoR−/− fetal livers depleted of cells expressing lineage markers, and fetal brain. DNA methylation at each of the six CpGs was obtained following bisufite conversion of genomic DNA, PCR amplification at HS1 and HS2, and pyrosequencing. In fetal brain methylation levels were high, at ≈60%–80%, for all six CpG dinucleotides. Methylation levels were lower in all hematopoietic cell subsets (Figure 6A). Methylation levels were largely similar in all hematopoietic, Epo-independent cell subsets examined: LSK, Mac-1+, CD41+, S0, and EpoR−/− cells. The onset of Epo dependence in S1 was associated with a marked reduction in DNA methylation in all six CpG dinucleotides, with the level of methylation dropping to virtually undetectable levels in S1 for four of the six CpGs.
We found that DNA demethylation also took place in freshly sorted S0 cells allowed to differentiate in vitro (Figure 6B,C). Demethylation in vitro occurred earlier at the HS1A, B, C, and HS2C than at HS2A, B (Figure 6C, red lines), in agreement with results in vivo (Figure 6A). Demethylation in vitro was arrested at all CpGs if either aphidicolin or mimosine were added to the incubation medium, and resumed when these drugs were removed (Figure 6B,C). Therefore, DNA demethylation, initiated at the transition from S0 to S1, is dependent on S-phase progression. These results are suggestive of a passive demethylation process, due to loss of maintenance methylation at nascent DNA.
We examined HS1 and HS2 DNA methylation levels in S0 cells transduced with PU.1 (as in Figure 4A) and incubated in Epo for 24 h. DNA methylation was significantly higher at 3 of the 6 CpGs in S0 cells transduced with PU.1-ICD4, compared with control cells transduced with MICD4 (Figure 6D). Therefore, PU.1 expression, along with its inhibitory effect on erythroid differentiation, also impaired DNA demethylation, possibly due to its inhibitory effect on S-phase in these cells (Figure 4C).
We have identified a committal step in erythropoiesis in which the cell cycle clock is precisely synchronized with and coordinates an erythroid differentiation switch. It takes place during S-phase of the last CFU-e generation, at the transition from S0 to S1, when S-phase progression is required for several distinct committal differentiation events, including the onset of Epo dependence, a switch in chromatin at the β-globin locus into an open conformation, and activation of GATA-1 function with consequent transcription of GATA-1 target genes (Figure 7A). The transition from S0 to S1 can be replicated in vitro, where sorted S0 cells develop into a differentiation state characteristic of S1 within 10 to 12 h.
S-phase progression at the S0 to S1 transition and the ensuing differentiation switch are dependent on the downregulation of p57KIP2 (Figures 3G, 7B), a novel finding since, to date, the principal known role of CDKIs in differentiating cells had been to mediate terminal differentiation secondary to cell cycle exit [1]–[3]. Unlike other CDKIs, p57KIP2 is required for the development of multiple tissues [35], suggesting that its novel role in erythropoiesis, triggering S-phase progression during a committal differentiation event, may be replicated in other systems. Of note, p57KIP2 is the only CDKI of the CIP/KIP family to be downregulated at the S0/S1 transition.
The synchronization of S-phase progression with several rapid and committal differentiation transitions suggests they are co-regulated. A key mediator of this co-regulation is PU.1, whose expression declines at the transition from S0 to S1. We have identified a novel cross-antagonism between S-phase progression and PU.1 expression. We show that S-phase arrest, caused by high levels of p57KIP2 or by cell cycle blocking drugs, prevents downregulation of PU.1 (Figure 3E,H); conversely, failure to downregulate PU.1 arrests S-phase progression (Figure 4B,C). Either maneuver blocks erythroid differentiation, including a block of chromatin reconfiguration at the β-globin locus and blocked expression of erythroid-specific genes (Figures 3B, 3D–H, S3G, S3I, 4B, 5E, 6C–D).
We propose that the mutual inhibition between PU.1 and S-phase progression at the S0 to S1 transition (Figure 7B) simultaneously controls the transition in both differentiation and cell cycle states. Its function is analogous to a synchromesh mechanism in automotive transmission, matching the speeds of two rotating gears before allowing them to lock together during a gear-shift. The mutual antagonism between S-phase progression and PU.1 expression ensures that PU.1 downregulation does not occur prior to the cell's entry into S-phase; conversely, S-phase entry cannot occur before conditions for PU.1 downregulation are in place. In this manner the cell cycle and differentiation programs can only proceed when precisely synchronized.
Once cells have transitioned from S0 into S1, S-phase progression is no longer required for expression of erythroid genes (Figure 3C, Figure S3E–F). Therefore, the synchromesh mechanism is specific to the transition from S0 to S1, when committal decisions bring about an irreversible terminal differentiation phase. Our findings reveal a key organizational feature in erythroid differentiation and have implications for differentiation of other lineages, where similar synchronization events may occur. Further, the synchromesh mechanism we describe may be a target in leukemogenesis, consistent with reports that high levels of PU.1 promote erythroleukemia [15]. Similarly, although no reports at present implicate p57KIP2 specifically in erythropoiesis, mutations in p57KIP2 are implicated in the familial Beckwith-Weidemann syndrome, which predisposes to pediatric tumors [22].
PU.1, whose physiological function in erythropoiesis had not been clear, plays a pivotal role at the S0 to S1 transition, through its cross-antagonism with S-phase progression. This cross-antagonism is lineage and differentiation stage-specific, since it presumably does not operate in myeloid and B-cell lineages where PU.1 is an essential transcriptional activator. Similarly, within the erythroid lineage, this mutual antagonism must be activated specifically in the last generation of CFU-e. Its premature activation at an earlier CFU-e cycle may be predicted to result in premature transition into S1 and consequently, in a reduced number of differentiated progeny. This prediction helps explain previous observations, where erythroid cells from PU.1-null embryos were found to differentiate prematurely and to have reduced self-renewal capacity [36]. These observations are consistent with the PU.1-null phenotype mimicking premature downregulation of PU.1. The mutual inhibition between PU.1 and S-phase may also explain findings in the T-cell lineage, where exogenous expression of PU.1 at the pro-T cell stage was found to block both thymocyte expansion and differentiation [37].
Previous work documented cross-antagonism between PU.1 and GATA-1, showing them to interfere with each other's transcriptional functions through a variety of mechanisms including direct physical binding [11]–[14]. We propose that the activation of erythroid terminal differentiation at the S0/S1 boundary is due to functional activation of GATA-1 (Figure 7B). Though present in S0 cells prior to the transition into S1 (Figure 3A), GATA-1 function is inhibited by PU.1. Downregulation of PU.1 at the S0 to S1 transition alleviates this inhibition, allowing GATA-1-mediated activation of erythroid gene induction. Among its known targets, GATA-1-mediated transcriptional repression of GATA-2 [29] would account for our observation that GATA-2 is downregulated at the S0 to S1 transition (Figure 3A). This scheme places the decrease in GATA-2 downstream of the cross-antagonism between PU.1 and GATA-1 (Figure 7B) and explains why exogenous high levels of GATA-2, unlike PU.1, do not block the transition from S0 to S1 (Figure 4).
Based largely on immortalized progenitor-like cells, the antagonism between GATA-1 and PU.1 was proposed to underlie a binary cell fate choice in cells expressing both GATA factors and PU.1. An increase in GATA-1 would result in PU.1 suppression and the erythro-megakaryocytic cell fates, whereas an increase in PU.1 would suppress GATA-1 and give rise to the myelo-lymphocytic lineages [11]–[14]. However, our data show that CFU-e cells, considered committed erythroid progenitors, express PU.1 at levels equivalent with those found in cells of the myeloid lineage (Figure S3A). Our results are consistent with previous reports of PU.1 expression in early erythroid progenitors, including the expression of a GFP reporter “knocked in” to the PU.1 gene locus in S0 (CD71lowTer119negative) fetal liver cells [36],[38].
The biochemical nature of commitment to the erythroid lineage is unknown at present. It is possible that CFU-e cells prior to PU.1 downregulation, which are also expressing GATA-1 and GATA-2, are in fact multipotential cells that may give rise to either myeloid or erythro-megakaryocytic lineages. In this case, the cross-antagonism between PU.1 and GATA-1 would simultaneously be responsible for a lineage choice, as well as facilitate activation of the erythroid gene expression program at the S0 to S1 transition should this choice be in favor of the erythroid lineage. Our ability to isolate CFU-e cells expressing high levels of PU.1 prior to their transition into S1 should facilitate further study of this issue.
Why is S-phase progression coupled to the erythroid differentiation program at the S0 to S1 boundary? Linking developmental transitions to cell cycle phases may serve as a strategy for their correct developmental timing [39] and may ensure the correct number of differentiated progeny. Another possibility is that S-phase progression plays a direct role in the re-configuration of chromatin at erythroid-specific gene loci. DNA replication was proposed to provide an opportunity for structural changes in chromatin, since the passage of the replication fork transiently disrupts nucleosomes [4],[5]. Indeed, S-phase is essential for activation or silencing of some genes in yeast [40],[41] and metazoa [39],[42]–[45], though it is not known that this is due to a requirement in the reconfiguration of chromatin. However, S-phase is not required for activation of other developmental genes [46]–[50]. Further, in recent years the structure of chromatin was found to be much more dynamic outside S-phase than originally suspected [51]. It is therefore unclear whether there is an innate requirement for DNA replication in the reconfiguration of chromatin during activation of lineage-specific genes, or what specific aspects of chromatin restructuring might require S-phase.
Here we found that S-phase is required for DNA demethylation and for formation of DNase I hypersensitive sites. The requirement for DNA replication suggests that DNA demethylation is passive, due to a decrease in maintenance methylation of the nascent DNA strand [52]. This raises the possibility that formation of DNase I hypersensitive sites may require DNA replication because it might be contingent on DNA demethylation. Alternatively, DNase I hypersensitivity may require S-phase progression in order to lift a direct repressive effect of PU.1 on chromatin [13].
Our examination of EpoR−/− fetal liver shows that the EpoR becomes essential for erythroid differentiation at the S0/S1 boundary. The principal function of EpoR at this time is its pro-survival signaling: EpoR−/− erythroid progenitors undergo apoptosis but their cell cycle status is unaltered, suggesting that EpoR signaling is not required for S-phase progression (Figure S1E). These findings are consistent with the established role of EpoR as a survival factor that does not affect the erythroid cell cycle [53]. EpoR signaling is probably also dispensable for downregulation of PU.1 at the S0 to S1 transition, since both PU.1 and GATA-2 are low in EpoR−/− cells (Figure 3A).
In spite of both S-phase progression and PU.1 downregulation being apparently unimpaired, EpoR−/− cells fail to develop DNase I HS sites and fail to undergo DNA demethylation at the β-globin LCR (Figures 5D, 6A). It has been reported that exogenous expression of bcl-xL facilitates Epo-independent differentiation of erythroblasts [54] arguing against a direct requirement for EpoR signaling in chromatin reconfiguration. Therefore, EpoR−/− cells may be undergoing rapid apoptosis prior to the time when the chromatin change would have otherwise taken place.
Other than its survival function, EpoR is probably directly required for CD71 upregulation, via Stat5 [55],[56]. However, EpoR signaling results in CD71 upregulation only if S-phase is allowed to proceed (Figure 2).
Thus, while the onset of Epo dependence occurs synchronously with committal chromatin and transcriptional events in erythroid differentiation, there is apparently no direct requirement for EpoR signaling in these events, other than ensuring cell survival. The principal function of Epo in erythropoiesis is to determine the number of differentiated erythrocytes, via Epo concentration [57]. The S0 to S1 transition may have evolved as the time of onset of Epo dependence as it represents a biochemical commitment to erythroid differentiation, setting in motion chromatin and transcriptional transformations that lead to expression of erythroid-specific genes. This therefore represents the earliest time in erythroid differentiation when Epo may regulate cell number specifically within the erythroid lineage, with minimal lateral effects on other hematopoietic cells.
The β-globin LCR had long been studied as a model of chromatin at sites of lineage-specific genes. However, the time in erythroid differentiation when the locus switches from a “closed” to an “open” conformation had not been clearly defined. Further, it was not known whether activation of the locus develops in a step-wise fashion over several cell cycles and differentiation stages or whether it occurs rapidly in a single step.
Our findings show that, strikingly, the locus transitions to an active conformation rapidly, within S-phase of a single cell cycle. Further, several distinct functional and biochemical changes that characterize the active chromatin conformation appear to develop simultaneously. We found marked differences between S0 and S1 cells in DNA methylation and in DNase I hypersensitivity at the LCR. These transformations could be reproduced when purified S0 cells transitioned into S1 in vitro (Figures 5, 6). Both DNA demethylation and DNase I hypersensitivity required S-phase progression for their development. Further, we also found that development of histone-tail modifications characteristic of active chromatin, as well as the switch in the timing of replication of the locus from late to early S-phase, both coincide with the transition from S0 to S1 (Figure 5). Therefore, our findings support an “all or none” model for the state of chromatin, previously hypothesized based on the probabilistic nature of developing DNase I hypersensitivity in a range of mutated chicken β-globin enhancer constructs [58].
Previous work showed that while the highest levels of DNase I accessibility at the β-globin LCR are attained in mature erythroid progenitors, the β-globin LCR is already poised for expression in earlier multipotential progenitors, contributing to low-level β-globin transcription (“priming”) [59],[60]. The β-globin LCR was found to already contain DNase I hypersensitive sites in cell lines resembling early hematopoietic progenitors [61]. Here we find that the β-globin LCR appears poised for change prior to the transition from S0 to S1. Thus, LSK and S0 cells have similar DNA methylation levels that are substantially lower than in fetal brain, suggesting chromatin already primed for expression at the LSK stage (Figure 6A). Histone tail modifications in the LCR similarly suggest that chromatin in S0 is poised for change, as it is enriched with both H3K4me2, a mark associated with active chromatin, and with H3K27me3, a mark found in silent chromatin (Figure 5G). The LCR is therefore marked as a bivalent domain, which may denote chromatin that is silent but primed for activation [62],[63]. Regardless of the precise state of chromatin readiness in earlier hematopoietic progenitors, however, our results show a clear switch in chromatin conformation at the S0 to S1 transition.
The clear switch we identified at the β-globin LCR occurs in synchrony with other switch-like transformations at the transition from S0 to S1, including the onset of Epo dependence and activation of GATA-1 function. Our ability to identify this transition with precision in vivo and manipulate it genetically in vitro should facilitate further study of the pivotal link between the cell cycle clock and the committal chromatin decisions that bring about the erythroid phenotype.
Fetal livers (E12.5–E14) were mechanically dissociated and immunostained as described [64]. Immunofluorescence was measured on an LSRII (BD Biosciences, CA) and data analyzed using FloJo (Tree Star, CA). Cells were sorted on a FACSAria, FACSVantage (BD Biosciences), or MoFlo (Beckman Coulter) cell sorters. In a small number of experiments StemSep columns (StemCell Technologies) were used.
Freshly harvested fetal liver cells were sorted and cultured in medium containing 20% fetal calf serum and 2 U/ml Epo (Amgen) for up to 48 h.
BrdU (100 µl of 10 mg/ml) was injected intra-peritonealy to pregnant mice and embryos were harvested 30–50 min later. In vitro, cells were pulsed with BrdU for 30 min. BrdU incorporation was detected using BrdU flow kit (BD Biosciences). Cell tracking with CFSE (carboxyfluorescein diacetate succinimidyl ester) was performed on sorted S0, incubated with 2.5 µM CFSE (Invitrogen) for 10 min at 37°C.
Retroviral transduction was by spin infection of sorted S0 cells at 2,000 rpm, 37°C on fibronectin coated dishes in 5 µg/ml polybrene (Sigma). Transduced cells were incubated overnight in the presence of 100 ng/ml SCF and 10 ng/ml IL3 (Peprotech, Rocky Hill, NJ) and were then transferred to Epo-containing medium for the indicated times.
Quantitative RT-PCR was performed as described [64].
DNase I hypersensitivity assays were performed as described [34] with modifications to amplicons (see Supplemental Methods).
ChIP-qPCR was performed on 106 cells/sample of sorted S0, sorted S1, or fetal brain. Cells were cross-linked in 1% formaldehyde, sonicated, and incubated overnight with a range of antibodies (see Supplemental Data), followed by 3–4 h of incubation with Protein G-magnetic beads (Invitrogen). Cross-links were reversed and purified DNA measured by qPCR using the same amplicons as in the DNase I hypersensitivity assay.
Genomic DNA or cells were treated with sodium bisulfite (Zymo Research, Orange, CA). Bisulfite-converted DNA was amplified by PCR and methylation levels measured using pyrosequencing at EpigenDx (Worcester, MA).
See Text S1 for additional methods, primer sequences, and antibodies.
|
10.1371/journal.pbio.1001830 | Discovery of a “White-Gray-Opaque” Tristable Phenotypic Switching System in Candida albicans: Roles of Non-genetic Diversity in Host Adaptation | Non-genetic phenotypic variations play a critical role in the adaption to environmental changes in microbial organisms. Candida albicans, a major human fungal pathogen, can switch between several morphological phenotypes. This ability is critical for its commensal lifestyle and for its ability to cause infections. Here, we report the discovery of a novel morphological form in C. albicans, referred to as the “gray” phenotype, which forms a tristable phenotypic switching system with the previously reported white and opaque phenotypes. White, gray, and opaque cell types differ in a number of aspects including cellular and colony appearances, mating competency, secreted aspartyl proteinase (Sap) activities, and virulence. Of the three cell types, gray cells exhibit the highest Sap activity and the highest ability to cause cutaneous infections. The three phenotypes form a tristable phenotypic switching system, which is independent of the regulation of the mating type locus (MTL). Gray cells mate over 1,000 times more efficiently than do white cells, but less efficiently than do opaque cells. We further demonstrate that the master regulator of white-opaque switching, Wor1, is essential for opaque cell formation, but is not required for white-gray transitions. The Efg1 regulator is required for maintenance of the white phenotype, but is not required for gray-opaque transitions. Interestingly, the wor1/wor1 efg1/efg1 double mutant is locked in the gray phenotype, suggesting that Wor1 and Efg1 could function coordinately and play a central role in the regulation of gray cell formation. Global transcriptional analysis indicates that white, gray, and opaque cells exhibit distinct gene expression profiles, which partly explain their differences in causing infections, adaptation ability to diverse host niches, metabolic profiles, and stress responses. Therefore, the white-gray-opaque tristable phenotypic switching system in C. albicans may play a significant role in a wide range of biological aspects in this common commensal and pathogenic fungus.
| The capacity of the yeast Candida albicans to grow in several cellular forms—a phenomenon known as phenotypic plasticity—is critical for its survival and for its ability to thrive and cause infection in the human host. In this study, we report a novel form of C. albicans, the “gray” phenotype, which may enhance fitness and confer an adaptive advantage for this important pathogenic yeast in certain host environments. The gray cell type, together with the previously discovered “white” and “opaque” cell types, forms a tristable phenotypic switching system. The three phenotypes differ in their cellular and colony appearance, their global transcriptional profiles, their production of secreted aspartyl proteinases (enzymes that degrade host tissues and release nutrients), and their virulence in different infection models. Moreover, gray cells exhibit a level of mating competency that is intermediate between that of white and opaque cells. We further demonstrate that two key transcriptional regulators, Wor1 and Efg1, play central roles in the regulation of the “white-gray-opaque” tristable transitions. Our study reveals a multi-stable and heritable switching system, indicating that the adoption of distinct morphological forms in response to environmental change could be much more elaborate than previously thought.
| The ability of a single genotype to generate a number of different phenotypes in response to environmental stimuli, known as phenotypic plasticity, enables microorganisms to rapidly adapt to their changing environment and to survive and thrive in certain ecological niches. The human pathogenic yeast C. albicans can switch among several morphological phenotypes in response to a variety of environmental cues [1],[2]. The ability to grow in different morphological forms is critical for both its commensal lifestyle and its existence as a pathogen [3],[4]. The “white-opaque” transition is a well-known bistable phenotypic switching system in C. albicans [5]. White and opaque cells are two morphologically distinct cell types [6]. White cells are small and round and form white and shiny colonies on solid media, while opaque cells are larger and elongated and form flatter and rougher colonies. White and opaque cells also differ in their gene expression profiles, mating competency, and virulence characteristics [7],[8]. For instance, opaque cells can mate more efficiently and are better at cutaneous infections than white cells, while white cells are more virulent in systemic candidiasis [8].
The white and opaque phenotypes are heritable and stable for many generations of cell divisions [5],[6]. There does not appear to be a stable intermediate phase between white and opaque in C. albicans [5],[9], although transient intermediate phenotypes have been observed at the cellular level during the process of high temperature-induced transitions [10]. C. tropicalis and C. dubliniensis, two human fungal pathogens closely related to C. albicans, can also undergo white-opaque switching [11]–[13]. We recently observed an intermediate phase between the white and opaque phenotypes in C. tropicalis and proposed that the phenotypic switching system in this species may be tristable [13].
The white-opaque transition is regulated by the bistable expression of the master regulator gene WOR1 [14]–[16], and therefore, there is not an intermediate phase between white and opaque phenotypes. In the laboratory strain SC5314 and its derivatives, the mating type locus (MTL) controls white-opaque switching via repressing WOR1 expression by the a1-α2 complex [7],[14]. We have recently reported that a subset of clinical isolates of C. albicans MTLa/α heterozygous strains can undergo white-opaque switching when cultured in N-acetylglucosamine (GlcNAc)-containing media [17], which is thought to mimic the host environment. The key regulators, including Wor1, Wor2, Efg1, and Czf1, constitute an interlocking transcriptional circuit controlling white-opaque switching via positive and negative feedback loops [18].
In this study, we report a novel morphological phenotype of C. albicans, referred to as the “gray” phenotype. This phenotype is heritable but distinct from the previously identified white and opaque phenotypes in cellular and colony appearance, global gene expression profiles, secreted aspartyl proteinase (Sap) activities, and virulence characteristics. The gray phenotype, together with the white and opaque phenotypes, forms a novel tristable and heritable switching system in C. albicans. We further demonstrate that neither Wor1 nor Efg1 are required for the maintenance of the gray phenotype. Deletion of WOR1 blocks white-to-opaque and gray-to-opaque transitions, but not white-gray transitions. Deletion of EFG1 blocks opaque-to-white and gray-to-white transitions, but not gray-opaque transitions. Deletion of both WOR1 and EFG1 locks cells in the gray phenotype. Therefore, Wor1 and Efg1 may coordinately regulate the “white-gray-opaque” tristable phenotypic switching system in C. albicans.
We isolated a C. albicans strain (BJ1097) from the genital tract of a female patient at a women's health hospital in Beijing, China. We sequenced the internal transcribed spacers (ITS) and 5.8S rDNA region and verified that BJ1097 is a C. albicans strain. When this strain was grown on yeast extract-peptone-dextrose (YPD) agar plates for several days, we observed a novel colony phenotype, hereafter referred to as the “gray” phenotype, in addition to the typical white and opaque colony phenotypes (Figure 1A). Gray colonies appeared smooth and gray, while typical opaque colonies were gray and rough or “opaque,” and typical white colonies were relatively white and smooth. On YPD agar containing the dye phloxine B, the white colonies remained white and the opaque colonies were stained pink, while the gray colonies exhibited a distinctly lighter pink color (Figure 1B and 1C). The color of the gray colonies was noticeably different than that of the opaque colonies on phloxine B containing media. The cellular morphologies of the white, gray, and opaque phenotypes were also distinguishable on YPD medium (Figure 1C). Consistent with previous reports, white cells were round and small, while opaque cells were elongated and large. Gray cells were also elongated, but appeared much smaller than opaque cells (Figure 1C). The cellular and colony morphologies of the three phenotypes on Lee's glucose and Lee's GlcNAc medium are shown in Figures S1 and S2. Similar to the phenotypes on YPD medium, the order of coloration from darkest to lightest on Lee's media was opaque>gray>white. Cellular morphologies of white, gray, and opaque cells on Lee's media were also similar to those on YPD medium. The cellular morphology of gray cells was very similar to that of opaque cells of the haploid C. albicans strains recently reported by Hickman and colleagues [19]. We, therefore, performed fluorescence activated cell sorting (FACS) to assess ploidy, and found that all three cell types of BJ1097 are in fact diploid (Figure S1C).
The switching frequencies of white-gray-opaque transitions in air at 25°C are shown in Figure 2A. On YPD medium plates, white-to-gray and opaque-to-gray switching frequencies were 5.7%±0.7% and 89.7%±3.0%, respectively, indicating that the white and especially opaque phenotypes are not stable under this culture condition. On Lee's glucose and Lee's GlcNAc media, the white and the opaque phenotypes were relatively stable when cultured in air at 25°C, while the gray-to-opaque switching frequencies were 21.3%±0.4% and 17.6%±3.1%, respectively. The white, gray, and opaque phenotypes were also stable in liquid Lee's media (Figure S3). Scanning electron microscopy (SEM) examinations demonstrated that the cell surfaces of white and gray cells were smooth, while at least a part of the opaque cell surface exhibited a pimpled appearance (Figure S4).
As shown in Figure S5, the strain BJ1097 could switch among the white, gray, and opaque phenotypes in Lee's and YPD media at 37°C. Switching frequencies are shown in Figure 2B. On glucose containing media (YPD and Lee's glucose), opaque cells underwent a mass conversion to the gray phenotype (colony switching frequency = 100%). On Lee's GlcNAc medium, opaque cells were much more stable than on the other two media.
GlcNAc and CO2 are white-to-opaque switching inducers, which are abundant in the host gut, a major niche of C. albicans [20],[21]. Consistent with our previous studies [17],[20],[21], the combination of GlcNAc and CO2 promoted white-to-opaque switching on Lee's medium (Figure 2C). On YPD medium, the induction effect of CO2 on the opaque phenotype was not obvious (Figures 2C and S6). The switching frequencies from opaque-to-gray were 89.7%±3.0% in air and 51.1%±4.4% in 5% CO2, suggesting that CO2 has an effect on stabilizing the opaque phenotype on YPD medium. In 5% CO2, gray cells are more stable than white cells, which showed a high frequency to switch to the opaque phenotype on Lee's glucose and Lee's GlcNAc media (Figure 2A and 2C).
At 37°C, opaque-to-gray switching on Lee's glucose and Lee's GlcNAc medium were 100% and 0.5%±0.4%, respectively, suggesting that GlcNAc can also stabilize the opaque phenotype (Figure 2B). However, neither GlcNAc nor CO2 had a notable effect on white-gray transitions on three different media both at 25°C and at 37°C (Figure 2A–2C).
To test whether other clinical strains of C. albicans could form the gray phenotype, we plated 30 clinical isolates of C. albicans on Lee's media and cultured them at 25°C. These strains are all competent at white-opaque switching ([17] and our unpublished data). We found that a subset of strains could switch to the gray phenotype. Six examples are shown in Figure S7. The cellular morphology of gray cells of all these strains was similar to that of BJ1097. These results suggest that the white-gray-opaque transition is a general feature of clinical strains of C. albicans.
To better understand the differences among the three cell types, we performed RNA-Seq analysis to investigate their global gene expression profiles. As shown in Figure 3A and 3B, the three cell types exhibit distinct and overlapping gene expression patterns. A more detailed analysis of the differentially expressed genes in white, gray, and opaque cells are shown in Tables S1 and S2. Key findings are summarized as follows: (1) Gene expression profiles of white and opaque cells are consistent with our previous study performed in a different MTLa/α strain (CY110) and other reports in MTL homozygous strains [17],[22],[23]. For example, WOR1, WOR3, and OP4 were enriched only in opaque cells, while EFG1 was expressed in white and gray cells but not in opaque cells. WH11 was enriched only in white cells. (2) Gray-enriched genes included cell wall-related (e.g., PGA26 and SUN41), drug resistance-related genes (e.g., CSA2, orf19.3348 and orf19.3475), stress-response-related (e.g., HSPs), metal ion metabolism-related genes (e.g., FRE7 and FRE30), and some secreted enzymes (SAPs and LIP9). Notably, two oral infection-upregulated genes (orf19.6200 [24] and orf19.6070 [25]) were exclusively enriched in gray cells. (3) The expression profiles of metabolism-related genes, especially those involved in carbohydrate metabolism, exhibited distinct features from that of white and opaque cells. A small proportion of genes (<10%), such as the NADH oxidase gene AOX2 and the glycerol permease gene HGT10, were highly expressed in gray cells. About 30% of metabolism-related genes showed a similar expression level in gray cells to that of white cells, ∼20% to that of opaque cells, and ∼30% exhibited an intermediate level between white and opaque. Consistent with the previous study [22], white cells express a fermentative metabolism-gene profile, while opaque cells adopt an oxidative one. These results indicate that gray cells may have a unique metabolic mode, which is different from that of white and opaque cells. (4) A large set of genes encoding signaling peptide- or glycosylphosphatidylinositol (GPI)-containing proteins were enriched in gray cells, including orf19.3378, orf19.3376, orf19.3117, orf19.6200, PGA26, and IFF8. (5) The absolute expression levels in opaque cells (indicated by the reads per kilobase per million mapped reads, or RPKM values) of SAP genes were much higher than those in white and gray cells. For example, the RPKM value of SAP1 is over 20,000 in opaque cells, while it is less than 30 in both white and gray cells. Although the expression level of SAP7 in gray cells is about 10-fold higher than that in opaque cells, the absolute expression levels in both cell types are very low (the RPKM values are 22 and 2 in gray and opaque cells, respectively).
The transcriptional profiles of the SAP genes, which are known major virulence factors [26], were different in the three cell types. We therefore tested Sap activities using the yeast carbon base (YCB)-bovine serum albumin (BSA) medium assay. As expected, white cells showed lowest Sap activity, which is consistent with low expression levels of the SAP genes. However, we surprisingly found that gray cells exhibited higher Sap activity than opaque cells, indicated by the white halos of precipitated BSA (Figure 4A). Quantitative Sap activity assays verified these results (Figure 4B). This result is inconsistent with the expression profiles of SAP genes in gray and opaque cells. Since the RNA-Seq analysis was performed in Lee's glucose medium (without BSA), we predicted that the YCB-BSA medium induced the expression of SAP genes in gray cells and thus increased Sap activity. To test this hypothesis, we performed quantitative Sap activity assays. As predicted, the Sap activity of opaque cells was higher than that of gray cells in Lee's glucose medium, but was lower than that of gray cells in the YCB-BSA medium. Using a green fluorescent protein (GFP) reporter system (Figure 4C and 4D) and quantitative real-time PCR (Figure 4E) assays, we further found that SAP1 was constitutively expressed in opaque cells in both media. SAP2 exhibited extremely low expression levels in all three cell types in Lee's medium, while its expression level was increased over thousands of times in gray cells cultured in the YCB-BSA medium. Taken together, these results imply that the three cell types are likely to have differential virulence characteristics at different host niches.
White cells were much more virulent than gray and opaque cells in the survival assay of systemic infections (Figure 5A and 5B). At the higher inoculation concentration (3.75×106 cells per mouse), the order of virulence observed from highest to lowest was white cells>opaque cells>gray cells (Figure 5A). At the lower inoculation concentration (1×106 cells per mouse), both gray and opaque cells exhibited similarly low virulence, while white cells killed all the infected mice in seven days (Figure 5B). We further performed competitive infections to evaluate the fungal burdens in different organs. As shown in Figure 5C, the three types of cells differed in fungal burdens in different organs, suggesting that the three cell types may have distinct abilities to invade and colonize different host tissues. For example, the fungal burden of white cells was higher than that of gray and opaque cells in the kidney, while the fungal burdens of gray and opaque cells were relatively higher than those of white cells in the liver and spleen. The fungal burdens of opaque cells were significantly higher than those of gray cells in four (liver, kidney, lung, and brain) of the five organs examined. These results were consistent with the data of survival assays (Figure 5A).
Secreted extracellular proteinases (such as Sap enzymes and lipases) play critical roles in degrading host tissues, which is thought to help fungal growth by releasing nutrients as well as facilitate fungal penetration during infections [26]. As shown in the ex vivo tongue infection assay (Figure 5D), the order of growth rates for the three cell types from fastest to slowest was gray cells>opaque cells>white cells, indicating that gray cells are better suited for nutrient acquisition from the animal tissue. Consistently, SEM examinations at 48 hours post-inoculation revealed that gray and opaque cells caused significantly greater damage of the skin than white cells. The skin infected with white cells remained largely intact and smooth, while the skin infected with gray or opaque cells exhibited obvious degradation and damage (Figure S8).
The basic morphology of gray cells is similar to that of opaque cells, although the former is much smaller in cell size (Figures 1, S1, and S2). Moreover, the key white-opaque switching regulator genes, WOR1 and EFG1, were differentially expressed in the three cell types. We therefore set out to test whether the gray phenotype is governed by Wor1 and Efg1. Deletion of WOR1 blocked opaque cell formation but allowed white-gray transitions (Figures 6A, S9, and S10). Deletion of EFG1 blocked white cell formation but allowed gray-opaque transitions (Figures 6B, S9, and S10). The frequencies of white-gray switching in the wor1/wor1 mutant and gray-opaque switching in the efg1/efg1 mutant are shown in Figure S10. Given that Wor1 and Efg1 are essential for the formation of opaque and white cell types, respectively, we predicted that inactivating Wor1 and Efg1 simultaneously would block the formation of both the white and the opaque cell type and thus could only allow cells to exist in the gray phenotype. We, therefore, constructed a wor1/wor1 efg1/efg1 double mutant. As predicted, the double mutant was locked in the gray phenotype and could not switch to the white or opaque phenotype under all culture conditions tested (Figures 6C, S10E, and S10F). The wild type (WT) control is shown in Figure 6D. These results indicate that (1) neither Wor1 nor Efg1 is essential for gray cell formation; (2) both Wor1 and Efg1 could repress the formation of the gray phenotype; (3) Wor1 and Efg1 may work coordinately in the regulation of the white-gray-opaque phenotype. A regulatory model of the tristable switching system by Wor1 and Efg1 is shown in Figure 6E.
Over 90% of clinical isolates of C. albicans are MTL heterozygotes (a/α). The strain BJ1097 and five strains used in Figure S7 are all MTL heterozygotes (a/α). We therefore examined whether the MTL locus regulates the gray phenotype. As shown in Figure 7A, both the MTLa/Δ and the Δ/α strains exhibited white, gray, and opaque phenotypes and could switch among the three phenotypes frequently (unpublished data). The cellular morphologies of gray cells of the MTLa/Δ or Δ/α were similar to those of the parent a/α gray cells. The cell size of opaque cells of the MTLa/Δ or Δ/α strain were larger than that of a/α opaque cells. More importantly, we found that the natural MTLα/α strain 19F could also undergo the white-gray-opaque transitions (Figure S7) [27]. Therefore, the white-gray-opaque tristable switching system is independent of the MTL locus.
We further demonstrated that gray cells mate over 1,000 times more efficiently than white cells, but hundreds of times less efficiently than opaque cells in the wild type strains (Figure 7B). The relatively high mating efficiency of gray cells could be partly due to their high switching frequency to the opaque phase (Figure 2A). To rule out this possibility and characterize the mating ability of gray cells, we next performed mating assays in the wor1/wor1, efg1/efg1, and wor1/wor1 efg1/efg1 mutants. Gray cells of the wor1/wor1 mutant mated about 2,000 times more efficiently than their white cell counterparts and 248 times more efficiently than white cells of the wild type. In the efg1/efg1 mutant, gray cells exhibited a mating ability comparable to that of gray cells of the wild type. Consistently, the wor1/wor1 efg1/efg1 double mutant, locked in the gray cell type, showed an intermediate mating competence between the white and the opaque phenotype (Figure 7B). These results suggest that gray cells indeed mate more efficiently than white cells since cells of the wor1/wor1 and wor1/wor1 efg1/efg1 mutants cannot switch to the highly mating-competent opaque phenotype.
High-frequency switching of colony morphology was observed in several clinical isolates of C. albicans by the Soll lab [28],[29]. The strains could switch heritably and reversibly between at least seven different phenotypes, not including the white-opaque transition [28],[29]. Here we report a novel morphological phenotype, the gray cell type, and a white-gray-opaque tristable switching system in C. albicans. Our new findings, together with previous reports [28],[29], suggest that C. albicans is capable of undergoing multiple-stable phenotypic transitions under certain environmental conditions. Compared with white and opaque cells, gray cells exhibit several unique characteristics: (1) distinct cellular and colony appearance; (2) high Sap activity in BSA-containing media; (3) tissue-specific infection ability; and (4) differential global gene expression profiles. A more comprehensive comparison of features of the three phenotypes is presented in Table 1.
The tristable and heritable phenotypic switching system reported in this study may confer the pathogen with higher plasticity and capacity of environmental adaptation. White-gray-opaque transitions can both occur spontaneously and be induced by environmental cues. Although CO2 and GlcNAc promote white-to-opaque switching and stabilize the opaque phenotype [21], they have no obvious effect on the induction of the gray phenotype (Figure 2). However, the rich YPD medium facilitates the formation of gray cells (Figure 2). These results suggest that white and gray cells differ in response to environmental cues and that these three cell types may have different abilities to adapt to specific host niches.
White, gray, and opaque cells show distinct global gene expression profiles. The gray cell type-enriched genes include these encoding secreted enzymes, cell wall and surface proteins, metabolism- and antifungal-related proteins, and filamentation-regulators. This unique gene expression profile could contribute to a number of important biological traits of C. albicans, such as filamentation and biofilm formation, virulence, and resistance to antifungals. For example, a large set of genes encoding cell wall, membrane, and extracellular proteins, was upregulated in gray cells. These proteins directly interact with the extracellular environment and may play critical roles in sensing and responding to environmental changes.
Secreted proteinases, including Saps in C. albicans, play important roles in nutrient acquisition, tissue adherence, invasion, and infection [26]. Interestingly, although the expression of a number of SAP genes was only enriched in opaque cells in Lee's glucose medium, gray cells exhibited higher Sap activity than opaque cells in the BSA-containing medium (Figure 4A and 4B). The inducible expression mode of SAP2 could confer advantages to gray cells over white and opaque cells (Figure 4D and 4E). Previous studies have demonstrated that SAP2 mRNA is highly upregulated in a reconstituted human epithelial (RHE) infection model and critical for epithelial tissue damage [30]–[32]. Constitutive expression of SAP genes in opaque cells could cost and waste a lot of energy to the cell, and the damaging effects of Saps could evoke strong host immune response [4],[26]. The expression of SAP genes would not favor the commensal life style of C. albicans. The low Sap activity of white cells limits certain infection abilities, especially the ability to cause cutaneous infections. Inducible expression of Sap activity could also play a balancing role in the transition between the commensal and pathogenic life styles in C. albicans. The major components of the skin surface are proteins, such as collagen, elastin, and keratin. Like BSA, these proteins could also induce Sap activity in gray cells. Consistent with this idea, the skin damage and nutrient acquisition abilities of white, gray, and opaque cells correspond to their respective Sap activities exhibited in BSA-containing media (Figures 4, 5D, and S8). Gray cells and opaque cells are less virulent in the systemic infection model possibly due to their weaker abilities to filament compared to white cells (unpublished data). Gray cells and opaque cells propagated faster than white cells in an ex vivo tongue infection model and led to more serious damages in an in vivo skin infection model (Figure S8). These results are consistent with previous studies in terms of the correlation between Sap activity and cutaneous infections [26].
Moreover, the three cell types differ in fungal burdens in different organs (Figure 5C), suggesting that they may play distinct and specific roles in systemic infections. The distinct global transcriptional profiles and differential Sap activities of the three different cell types may contribute to the adaptability of C. albicans to inhabit a diverse number of host niches and to propagate in different tissues.
The MTL locus is involved in the regulation of white-opaque switching and sexual mating in C. albicans [7]. Here we found that the white-gray-opaque tristable switching system is independent of the regulation of the MTL locus. Although only a subset of C. albicans clinical strains undergo the tristable switch under the culture conditions tested, we suggest that this phenotypic switching system may be a general feature of natural C. albicans strains, and most, if not all, strains could do this under certain conditions (e.g., in certain niches of the host).
Wor1 is the master regulator of white-opaque switching and is essential for opaque cell formation (Figures 6, S9, and S10) [14]–[16]. Efg1 plays a negative role in white-to-opaque switching and is essential for white cell formation (Figures 6 and S10). Wor1 and Efg1, together with transcription factors Czf1, Wor2, and Wor3, form interlocking transcriptional feedback loops controlling white-opaque switching [18]. Although neither Wor1 nor Efg1 is required for the formation of the gray phenotype, deletion of both WOR1 and EFG1 locks cells in the gray phenotype (Figures 6, S9, and S10). These results suggest that different transcriptional circuitries may be involved in the regulation of white-gray and gray-opaque transitions. Together with other unidentified regulators, Wor1 and Efg1 may coordinately govern the formation of gray cells. On phloxine B-containing media, the cellular and colony morphologies of the wor1/wor1 efg1/efg1 double mutant we generated are similar to those of the wor1/wor1 efg1/efg1 double mutant in the SC5314 background constructed by Hnisz and colleagues [33], suggesting that derivatives of SC5314 also have the potential to switch to the gray phenotype. Both CO2 and GlcNAc induce white-to-opaque switching predominantly via activating the Wor1 regulator [21]; however, neither of these stimuli had an obvious effect on the induction of white-to-gray switching (Figure 2). This is reasonable because the expression level of WOR1 is very low in gray cells and Wor1 is not required for the formation of gray cells (Table S1).
Gray cells also differ from the recently reported GUT cell type (gastrointestinally induced transition) in several aspects [4]. First, GUT cells resemble opaque cells in shape and cell size but lack cell wall surface pimples [4]. Gray cells are much smaller than both opaque and GUT cells. Second, GUT cells have been reported to exist only in the animal gut and have not been found to be stable in vitro culture conditions, while gray cells are very stable in a variety of laboratory culture conditions. Third, several virulence genes including SAPs, are downregulated in GUT cells, while gray cells have a higher Sap activity than white and opaque cells in the presence of BSA. The switching characteristic, together with the unique aspects of gray cells discussed above, indicates that the gray phenotype is a novel morphological phenotype.
A number of reasons suggest that the gray cell type is not an intermediate of the white and opaque phenotypes. First, white cells can directly switch to the opaque phenotype under some culture conditions at high frequencies without an intermediate phase (Figure 2), and vice versa. Consistently, the frequencies of white-to-opaque switching is even higher than that of gray-to-opaque switching under certain conditions, suggesting that conversion to the gray phenotype is not necessary to facilitate the formation of opaque cells. Second, the key regulators of white-opaque switching, Wor1 and Efg1, are not required for the maintenance of the gray phenotype. If the gray phenotype were an intermediate of the white and opaque phenotypes, one would expect that the expression levels of white- or opaque-phase specific genes, such as WOR1, EFG1, WH11, and OP4 [2], would be at intermediate levels in gray cells. However, this was not the case (Tables S1 and S2). Third, the order of Sap activities from highest to lowest is gray>opaque>white cells. In this sense, the opaque phenotype is an intermediate of the white and gray phenotypes. Fourth, the gray phenotype is not transient, but heritable, and can be maintained for many generations. Fifth, similar to white and opaque cells, gray cells exhibit a unique global gene expression profile.
Understanding the regulatory mechanisms of phenotypic changes will provide insights into several fundamental questions such as how pathogens adapt to the host and survive and propagate under diverse niches. Here we add the gray morphology and the white-gray-opaque tristable transitions to the list of known fungal morphological switching systems. This study provides an example of multiple stable and heritable switching systems, indicating that the regulation of morphological forms to adapt to environmental changes could be much more elaborate than previously thought. Our study also sheds new insight on the regulatory mechanism of the transition between commensal and pathogenic life styles in C. albicans. Exploring the molecular basis of this tristable phenotypic switching system and its role in host adaptation will be important for seeking new strategies to treat infections caused by the major fungal pathogen of humans.
The strains used in this study are listed in Table S3. Lee's glucose medium and YPD medium (20 g/l glucose, 20 g/l peptone, 10 g/l yeast extract) were used for routine growth of C. albicans. Lee's glucose, Lee's GlcNAc [21],[34], and YPD media were used for phenotypic switching assays. 2% agar was added to the media to make solid nutrient plates. The dye phloxine B (5 µg/ml), which stains opaque colonies dark pink and gray colonies light pink, was added to the solid media. The two plasmids for WOR1 deletion, pSFS2A-WOR1KOa and pSFS2A-WOR1KOb, were used to delete WOR1 in BJ1097 as described previously [17]. The two plasmids for EFG1 deletion, pSFS2A-EFG1KOa and pSFS2A-EFG1KOb, were generated by inserting two fragments containing sequences homologous to the 5′- and 3′-terminus of the EFG1 gene. The primers used for PCR to generate these fragments are listed in Table S4. To generate the wor1/wor1 efg1/efg1 double mutant, the two alleles of EFG1 were deleted in the wor1/wor1 mutant using the same EFG1 knockout plasmids.
The same strategy we described previously [17] was used to construct the SAP1p-GFP and SAP2p-GFP reporter strains, BJ1097 was transformed with PCR products of the GFP-caSAT1 fragment (amplified from the template plasmid pNIM1 [35] with GFP reporter primers) (Table S4).
The tristable switching assays were performed similarly to previously described white-opaque bistable switching assays [21]. Briefly, white, gray, or opaque cells from cultures of five days were replated onto agar media and incubated in air or 5% CO2 at temperatures indicated in the main text for 4 (at 37°C) or 5 (at 25°C) days of growth. Switching frequency = (number of colonies containing the second or third alternative phenotype/total colony number)×100%. For example, white-to-gray switching frequency = (number of gray colonies plus colonies with gray sections/total colony number)×100%. To verify colony phenotypes, several representative colonies of each type were examined for cellular morphology.
SEM assays were performed as described previously [36]. White, gray, and opaque cells were grown on Lee's GlcNAc medium for 3 days at 25°C and used for SEM assays.
We first deleted one allele of the MTL locus in the parent strain BJ1097 (MTLa/α) with the plasmid L23.14 [37], generating BJ1097Na (MTLa/Δ Clon+) and BJ1097Nα (MTLΔ/α Clon+). The two strains were then grown in YPmal medium (1% yeast extract, 2% peptone, 2% maltose) for FLP-mediated excision of the SAT1/flipper cassette, generating BJ1097a (MTLa/Δ Clon−) and BJ1097α (MTLΔ/α Clon−). Experimental strains of the wor1/wor1, efg1/efg1, wor1/wor1 efg1/efg1 double mutants (MTLΔ/α, URA3+ Clon−) were generated by using a similar strategy. To generate the ura3− Clon+ tester strain (WTa, opaque in Figure 7B), the plasmid pNIM1, which contains a caSAT1 gene, was linearized and integrated into GH1012 (MTLa/a ura3−) [20], generating GH1012N. Quantitative mating assays were performed according to our previous publication [17]. Briefly, the mating experiments were performed on Lee's glucose medium at 25°C. The experimental white, gray, and opaque cell samples were collected from Lee's glucose medium plates. To test the mating efficiencies, 1×106 of GH1012N opaque cells and 1×106 of experimental cells (in white, gray, or opaque phase) were mixed and cultured on Lee's glucose medium plates for 48 hours at 25°C. The mating mixtures were resuspended, diluted, and plated onto three types of selectable plates for growth. Mating efficiencies were calculated as previously described [17].
YCB-BSA assay. Sap activity was monitored on YCB medium agar containing 0.2% BSA as the sole nitrogen source as described previously [38]. 5×106 cells of each cell type in 5 µl ddH2O were spotted onto the plates. The white halos indicate secreted enzyme activity. The size of the halo ring indicates the activity level. This experiment was repeated six times.
Quantitative Sap activity assays were performed according to Ray and colleagues [39]. Briefly, white, gray, and opaque cells were grown in Lee's glucose or YCB-BSA medium overnight at 25°C. An aliquot of 20 µl of the cell suspension (1×106 cells) was inoculated into 2 ml of fresh Lee's glucose or YCB-BSA medium. After incubation for 48 h at 25°C, the cell number of each sample was determined. The cultures were centrifuged at 13,000 rpm for 1 min. For activity assays, 250 µl of culture supernatant was mixed with 500 µl of 1% BSA in 0.1 M sodium citrate-HCl buffer (pH 3.0). The reaction mixtures were incubated at 37°C for 1 h with gentle agitation. The reaction was stopped by adding 1.25 ml of ice-cold 5% trichloroacetic acid (TCA). Precipitated material was removed by centrifugation and the protein concentration of the supernatant was determined in a modified Bradford assay according to the manufacturer's protocol (Sangon Biotech). The activities were calculated for 108 cells of each cell type. One arbitrary unit was defined as an extinction increase at 595 nm of 0.1/h.
All animal experiments were performed according to the guidelines approved by the Animal Care and Use Committee of the Institute of Microbiology, Chinese Academy of Sciences. The present study was approved by the Committee.
Systemic infection of mice was performed according to the previous studies [17],[36], with slight modifications. Female BALB/c mice aged 4–5 weeks were used for survival and competition experiments. 10 mice were used for injection of each cell type (white, gray, or opaque). 3.75×106 cells of each type were injected into a mouse via tail vein. For competition experiments, 5×105 cells of each type (white, gray, or opaque) of BJ1097 were mixed with 5×105 cells of SC5314N (NouR, locked in white phase) in 250 µl PBS and then injected into a mouse via tail vein. Mice were humanely killed at 24 hours after injection. Different organs were used for fungal burden assays. BJ1097 is sensitive to 100 µg/ml nourseothricin (clonNAT). To generate the nourseothricin resistant strain SC5314N, pNIM1 was linearized and integrated into the laboratory strain SC5314. Organ tissues (liver, kidney, spleen, lung, and brain) were homogenated, diluted in PBS, and plated on Lee's medium for 3 days of growth at 37°C in 5% CO2 for colony-forming unit (CFU) calculation. Colony morphologies of SC5314N and BJ1097 were distinguishable. SC5314N formed wrinkled and filamentous colonies, while BJ1097 formed smooth colonies under this culture condition. SC5314N and BJ1097 were also subject to the nourseothricin susceptibility test. Both colony morphology and nourseothricin susceptibility assays were used to distinguish SC5314N and BJ1097 formed colonies. Competitive index = the ratio of CFU number of BJ1097 to CFU number of SC5314N in each organ.
Skin infection assays were performed as described previously [17], with modifications. Newborn BALB/c mice (aged 2–4 days) were used. 4×106 cells of each cell type in 2 µl ddH2O were spotted on the skin on the back of a new born mouse. After water evaporated, a small sterile filter paper was covered and fixed on the fungal spot with First Aid tape. After 24 h, the infected areas were excised for SEM assays.
Ex vivo tongue infection assays were performed as described by Kamai and colleagues [40], with modifications. Tongues (of similar size and weight) were excised from humanely killed female BALB/c mice aged 4–5 weeks and added to each well of a 24-well polystyrene plate containing 1×107 cells of white, gray, or opaque cells in 1 ml PBS. 50 µg/ml ampicillin and 50 µg/ml kanamycin were added to each well to inhibit bacterial growth. After 24 hours of incubation at 37°C, cells in the liquid and on the tongue (after homogenization) were separately plated onto YPD agar for CFU assays. The total cell number of each well (including cells in the liquid and attached to the tongue) was calculated. The total cell number of each well indicates fungal growth rate. The experiment was repeated three times. For each time, three tongues were used for each cell type.
White, gray, and opaque cells were grown at 25°C in liquid Lee's glucose medium for 24 h and total RNA was extracted using GeneJET RNA Purification kits according to the manufacturer's instructions. RNA-Seq analysis was performed by the company BGI-Shenzhen according to the company's protocol (http://www.genomics.cn/) [41]. Approximately 10 million (M) reads were obtained by sequencing each library. The library products were sequenced using the Illumina HiSeq 2000. Software Illumina OLB_1.9.4 was used for basecalling. The raw reads were filtered by removing the adapter and low quality reads (the percentage of low quality bases with a quality value ≤5 was >50% in a read). Clean reads were mapped to the genome of C. albicans SC5314 using SOAP aligner/soap2 software (version 2.21) [42]. The gene expression level is calculated using the RPKM method [43].
The RNA-seq dataset has been deposited into the NCBI Gene Expression Omnibus (GEO) portal (accession number GSE53671).
|
10.1371/journal.pgen.1002765 | Geographic Differences in Genetic Susceptibility to IgA Nephropathy: GWAS Replication Study and Geospatial Risk Analysis | IgA nephropathy (IgAN), major cause of kidney failure worldwide, is common in Asians, moderately prevalent in Europeans, and rare in Africans. It is not known if these differences represent variation in genes, environment, or ascertainment. In a recent GWAS, we localized five IgAN susceptibility loci on Chr.6p21 (HLA-DQB1/DRB1, PSMB9/TAP1, and DPA1/DPB2 loci), Chr.1q32 (CFHR3/R1 locus), and Chr.22q12 (HORMAD2 locus). These IgAN loci are associated with risk of other immune-mediated disorders such as type I diabetes, multiple sclerosis, or inflammatory bowel disease. We tested association of these loci in eight new independent cohorts of Asian, European, and African-American ancestry (N = 4,789), followed by meta-analysis with risk-score modeling in 12 cohorts (N = 10,755) and geospatial analysis in 85 world populations. Four susceptibility loci robustly replicated and all five loci were genome-wide significant in the combined cohort (P = 5×10−32–3×10−10), with heterogeneity detected only at the PSMB9/TAP1 locus (I2 = 0.60). Conditional analyses identified two new independent risk alleles within the HLA-DQB1/DRB1 locus, defining multiple risk and protective haplotypes within this interval. We also detected a significant genetic interaction, whereby the odds ratio for the HORMAD2 protective allele was reversed in homozygotes for a CFHR3/R1 deletion (P = 2.5×10−4). A seven–SNP genetic risk score, which explained 4.7% of overall IgAN risk, increased sharply with Eastward and Northward distance from Africa (r = 0.30, P = 3×10−128). This model paralleled the known East–West gradient in disease risk. Moreover, the prediction of a South–North axis was confirmed by registry data showing that the prevalence of IgAN–attributable kidney failure is increased in Northern Europe, similar to multiple sclerosis and type I diabetes. Variation at IgAN susceptibility loci correlates with differences in disease prevalence among world populations. These findings inform genetic, biological, and epidemiological investigations of IgAN and permit cross-comparison with other complex traits that share genetic risk loci and geographic patterns with IgAN.
| IgA nephropathy (IgAN) is the most common cause of kidney failure in Asia, has lower prevalence in Europe, and is very infrequent among populations of African ancestry. A long-standing question in the field is whether these differences represent variation in genes, environment, or ascertainment. In a recent genome-wide association study of 5,966 individuals, we identified five susceptibility loci for this trait. In this paper, we study the largest IgAN case-control cohort reported to date, composed of 10,775 individuals of European, Asian, and African-American ancestry. We confirm that all five loci are significant contributors to disease risk across this multi-ethnic cohort. In addition, we identify two novel independent susceptibility alleles within the HLA-DQB1/DRB1 locus and a new genetic interaction between loci on Chr.1p36 and Chr.22q22. We develop a seven–SNP genetic risk score that explains nearly 5% of variation in disease risk. In geospatial analysis of 85 world populations, the genetic risk score closely parallels worldwide patterns of disease prevalence. The genetic risk score also predicts an unsuspected Northward risk gradient in Europe. This genetic prediction is verified by examination of registry data demonstrating, similarly to other immune-mediated diseases such as multiple sclerosis and type I diabetes, a previously unrecognized increase in IgAN–attributable kidney failure in Northern European countries.
| IgA nephropathy (IgAN) is a common kidney disease with a complex genetic determination. This disorder is diagnosed based on detection of mesangial proliferation and glomerular deposits of IgA1. Most frequently, IgAN has a progressing course and 20–50% of cases develop end-stage renal disease (ESRD) within 20 years of follow-up [1]. The disease has been detected among all ethnicities worldwide, but displays a striking geographic variation. It is the most common cause of kidney failure in East Asian countries, has intermediate prevalence in European and US populations but is rarely reported in populations of African ancestry. The diagnosis of IgAN requires a kidney biopsy, complicating accurate determination of heritability and population prevalence of disease. Autopsy and donor biopsy series suggest a prevalence of up to 1.3% in Finland [2] and 3.7% in Japan [3]. Familial aggregation of IgAN has also been recognized throughout the world [4], [5], [6], [7], [8], [9], [10], [11] and up to 14% of cases may be familial [8]. Moreover, family members frequently have aberrant glycosylation of the hinge region of circulating IgA1, a defect with an estimated heritability of 40–50% [12], [13]. These data suggest a strong genetic contribution to disease.
Recently, we have completed a large-scale genome-wide association study (GWAS) involving a cohort of 3,144 sporadic IgAN cases [14]. The discovery phase samples (1,194 cases and 902 controls) were recruited in Beijing, China and were comprised of individuals of Han Chinese ancestry. The most associated SNPs were then followed up in additional cohorts of Han Chinese and Europeans (1,950 cases and 1,920 controls). In the combined analysis, we discovered 5 novel susceptibility loci with consistent effects across individual cohorts. These include 3 distinct intervals in the MHC-II region on chromosome 6p21, with the strongest signal encompassing the HLA DQB1/DQA1/DRB1 locus (abbreviated as DQB1/DRB1 hereafter). Imputation of classical alleles suggested that this signal was partially conveyed by a strong protective effect of the DRB1*1501-DQB1*0602 haplotype. The second signal on Chr. 6p21 encompassed a ∼100 Kb region containing TAP2, TAP1, PSMB8, and PSMB9 genes (TAP2/PSMB9 locus) and the third signal on Chr. 6p21 contained the HLA DPA1/DPB1/DPB2 genes (DPA1/DPB2 locus). Independence of these three regions on Chr. 6p21 was demonstrated by their localization within distinct LD blocks as well as genome-wide significant associations after rigorous conditional analyses. We also detected significant association within the Complement factor H (CFH) gene cluster on Chr. 1q32, where alleles tagging a common deletion in the CFHR3 and CFHR1 genes imparted a significant protective effect (CFHR3/R1 locus). Finally, a fifth signal centered on the HORMAD2 gene on Chr. 22q12 and containing multiple genes demonstrated significant association with risk of IgAN (HORMAD2 locus). These five loci individually conferred a moderate risk of disease (OR 1.25–1.59), but together explained 4–5% of the variation in risk across the populations examined.
To follow-up these studies and better assess the risk imparted by susceptibility alleles in diverse populations, we performed a replication study in eight independent case-control cohorts and performed a meta-analysis of all available genetic data including the original GWAS, totaling in 10,755 individuals. The expanded sample size allowed us to formally assess locus heterogeneity, identify new independent risk variants by conditional analyses and search for first-order genetic interactions. Finally, we refined a genetic risk score for IgAN and analyzed differences in the distributions of the IgAN susceptibility alleles among the major world populations.
For replication we examined eight cohorts (five European, two East Asian, and one African-American cohort, totaling 2,228 cases and 2,561 controls, described in Table S1). While each individual cohort at best had 40–50% power to replicate original GWAS findings, the combined replication cohort (2,228 cases and 2,561 controls) provided essentially 100% power for replication across the range of allele frequencies and odds ratios initially observed (Table S2).
We genotyped the two top-scoring SNPs for the CFHR3/R1, TAP2/PSMB9, DPA1/DPB2, and HORMAD2 loci, but four SNPs were included for the DQB1/DRB1 locus to test for independent alleles at this interval by conditional analysis. After a standard assessment of genotype quality control, we performed association testing within each cohort using the standard Cochrane-Armitage trend test (Table S3). We also tested for heterogeneity of associations and performed a meta-analysis under both fixed and random effects models (Table 1).
Four of the five original GWAS loci displayed significant replication with direction-consistent ORs and no heterogeneity comparable to the original findings (Table 1). The strongest replication was at the DQB1/DRB1 locus and achieved genome-wide significance in the replication cohort (fixed effects OR 0.75, P-value 4×10−11). The CFHFR3/R1 locus on Chr.1q32, the HORMAD2 locus on Chr.22q12, and the DPA1/DPB2 locus on Chr.6p21 were also robustly replicated (fixed effects p-values 3×10−3–7×10−7), with minimal between-cohort heterogeneity (I2<25%). Accordingly, when combined with the four cohorts studied in the original GWAS, these four loci provided highly significant evidence of association (fixed effects p-values 3×10−10–5×10−32).
In contrast, the TAP2/PSMB9 locus on Chr. 6p21 displayed direction-consistent replication only in the Italian, German, Czech, and Japanese cohort but the full replication cohort did not support this association (Table 1, Table S3). However, when combined with the four cohorts from the original GWAS, this locus remained genome-wide significant (fixed effects p-values 1×10−8 and 6×10−10 for rs9357155 and rs2071543, respectively, Table 1). As expected, I2 and Q-tests provided evidence of heterogeneity and random effects meta-analysis, which explicitly models heterogeneity, was 1–3 orders of magnitude more significant than fixed effect meta-analysis at this interval (e.g. random effects p-value 3×10−11, I2 = 61% for rs9357155; Table 1). The heterogeneity was not attributable to differences in ethnicity or cohort size as the association results varied within Asian and European cohorts of differing size (Table S3).
The top signals in the original GWAS, represented by rs9275596 and located within the DQB1/DRB1 locus, were mediated by a very strong protective effect of the DRB1*1501-DQB1*602 haplotype [14]. However, the SNPs in this interval are in incomplete LD and conditional analyses in our GWAS [14] and in an independent study of Europeans [15] had indicated that additional independent haplotypes also contributed to the signal. Therefore, taking advantage of our expanded cohort size, we examined additional SNPs that were in partial LD with rs9275596 to detect potentially independent effects (rs9275224, rs2856717 and rs9275424, which had an r2 of 0.09 to 0.7 with rs9275596, Table S4).
After mutually conditioning each SNP on the remaining SNPs, three of the four SNPs in the DQB1/DRB1 region exhibited a genome-wide significant independent effect (rs9275596, rs9275224 and rs2856717, conditioned p-vales<5×10−8, Table 2). Interestingly, the conditioned effect of the minor allele of rs2856717 was reversed compared to the crude effect estimate, suggesting that the adjustment for LD structure has uncovered a risk haplotype in this region (conditioned OR 1.61, p = 2×10−10).
The above data indicated that there are multiple risk haplotypes within the DQB1/DRB1 locus. To better define these findings, we next phased four-SNP haplotypes at this locus and tested associations with disease (Table 3). We confirmed a very strong protective effect of the ATAC haplotype (freq. 0.21) which, based on our previous imputation analysis, carries the DRB1*1501/DQB1*602 classical alleles. In addition, we defined a new protective haplotype (ACAT, freq. 0.13) and a new risk haplotype (ATAT, freq. 0.05). The ATAC protective haplotype and the ATAT risk haplotype differ only by the rs9275596-C/T allele, explaining the reversal of OR for the rs2856717-T allele after conditioning for rs9275596 (Table 3). Additionally, the GCGT risk haplotype, tagged by the rs9275424-G allele, exhibited a weaker protective effect. These results were supported by both Asian and European cohorts (Table S5). Further support is provided by the global haplotype association test, which achieved a p-value of 3×10−43. Based on these analyses, we concluded that there are at least three independent haplotypes conferring risk of IgAN within this region.
Nonetheless, these 3 independent haplotypes in DQB1/DRB1 locus still did not explain associations in other Chr. 6p21 regions (TAP2/PSMB9 and DPA1/DPB2 loci, respectively represented by rs9357155 and rs1883414), and a fully adjusted model that included all independently associated SNPs continued to support the original GWAS findings of three discrete genome-wide significant intervals on Chr. 6p21 (Table 4).
We tested the possibility of interaction between the 7 risk-contributing SNPs and therefore tested for all possible pairwise interactions (Table S6). We detected strong evidence for a multiplicative interaction (defined as departure from additivity on the log-odds scale) between the CFHR3/R1 (rs6677604) and the HORMAD2 loci (rs2412971). In this interaction, the rs2412971-A allele has a strong and consistent protective effect among all genotypic subgroups, but its effects are reversed among homozygotes for the rs6677604-A allele, which closely tags a CFHR3/R1 deletion (Figure 1, Table S6). The significance of this interaction (p = 2.5×10−4) exceeds a Bonferroni-corrected threshold for 21 tests, and is most discernable among the European cohorts (p = 1.4×10−3), where both SNPs have higher minor allele frequencies. The 4-df genotypic interaction test was also significant for these two loci (p = 6.4×10−3), but the 1-df multiplicative interaction model provided a better fit.
The original IgAN risk score model was based on the genotypes of the top scoring SNPs at the 5 independent loci discovered in the GWAS [14]. We refined this risk score by incorporating the newly discovered independent effects of rs9275224 and rs2856717 and the interaction between the CFHR3/R1 and the HORMAD2 loci. A stepwise regression algorithm in the entire cohort defined a new risk score that retained the 7 SNPs exhibiting an independent effect as well as the rs6677604* rs2412971 interaction term (Table 4). When compared with the original GWAS model, the newly refined score was more strongly associated with disease risk and explained a greater proportion of the disease variance in both the replication and the original GWAS dataset (Table 5). Moreover, the refined risk score was a highly significant predictor of disease in each individual replication cohort (Table S7). In all datasets combined, the new risk score explained 4.7% in disease variance and was 13 orders of magnitude more significant than the original score. In this model, one standard deviation increase in the score was associated with nearly 50% increase in the odds of disease (OR = 1.47, 95% CI: 1.42–1.54, P = 1.2×10−72). This translates into nearly a 5-fold increase in risk between individuals from the opposing extremes of the risk score distribution (with tails defined by ≥2 standard deviations from the mean).
Similar to the GWAS study, we detected pronounced differences in the distributions of risk alleles among the three different ethnicities studied: for each of these seven risk loci, the frequency of the risk alleles was highest in East Asians and lowest in African-Americans (Figure S1). These differences were also reflected in highly significant disparities in the risk score distributions by ethnicity (Figure 2). Motivated by these observations, we examined global geographic variation in the genetic risk for IgAN by applying the newly refined IgAN risk score in 6,319 healthy individuals across 85 worldwide populations. We observed marked differences in the genetic risk across the world. Overall, the mean standardized risk score was lowest for Africans, intermediate for Middle Easterners and Europeans, and highest for East Asians and Native Americans (Figure 3 and Figure S2). Accordingly, the risk increased sharply with eastward distance from the prime meridian (Pearson's r = 0.27, p = 3.5×10−108). The same geospatial pattern were detected if we included only native populations of HGDP and HapMap-III (Figure S3), demonstrating that the findings are not biased by inclusion of control populations from the genetic association study. These data are consistent with the known East-West gradient in prevalence of IgAN, suggesting that genetic risk predicts prevalence.
Unexpectedly, higher resolution analysis of the European continent revealed an additional increase in the risk from South to North (Pearson's r = 0.11, p = 1.3×10−9). For example, northwestern Russians and northern inhabitants of Orkney Islands (Scotland) have the highest risk scores when compared with the rest of the European continent (Tables S8 and S9). To confirm these finding and test whether North-South variation in genetic risk is also reflected in differences in IgAN occurrence, we obtained genetic data from additional European populations (Belgian, British, Finnish, Swedish and Icelandic) and compared genetic risk scores with the incidence and point prevalence of IgAN among end-stage renal disease (IgAN-ESRD) populations across Europe (Table S10). As predicted by the genetic risk score, our analysis confirmed a strong North-South cline of both incidence and prevalence across the European continent (Figure 4). Notably, this analysis includes only patients with end-stage IgAN, on dialysis or after kidney transplantation, thus it underestimates the true incidence and population prevalence of IgAN. Because the point prevalence of IgAN-ESRD (Figure 4b) can be confounded by differential survival on renal replacement therapy and differences in kidney biopsy practice by country, we also examined IgAN-ESRD prevalence expressed as a percentage of all ESRD (Figure 4c), and ESRD due to biopsy-diagnosed primary glomerulonephritis (Figure 4d). Regardless of the metric used to quantify differences in IgAN occurrence, regression of the genetic risk score and the prevalence data on the average latitude resulted in positive correlations and parallel trends.
The co-variation in genetic risk score and IgAN-ESRD occurrence among world populations may also be in part influenced by differences in environment, or by other factors such as local medical guidelines for screening and treatment. To better distinguish these possibilities, we examined native populations that live under a uniform environment yet show variation in IgAN risk. In the densely sampled North Italian populations, the Alpine villagers of the Valtrompia region have a 3.5-fold higher prevalence of ESRD attributable to IgAN and primary glomerulonephritis when compared to the national average [16]. Consistent with this prevalence data, the median standardized risk score in this population was comparable to some of the Northern European countries and ranked as number one among the 17 Italian populations sampled in our study (Figure 5, Table S8).
Conversely, we compared the genetic risk score and IgAN-ESRD prevalence in populations in the United States, where diverse ethnicities live under different environments and health care systems compared to the ancestral populations. The analysis of the USRDS dataset confirmed the striking ethnic differences in IgAN-ESRD prevalence (Table S11): the percentage of ESRD attributable to IgAN was 5-fold greater for Caucasian and 15-fold greater for Asian Americans compared to African-Americans. This increased IgAN-ESRD occurrence in Asian- compared to African-Americans far exceeds the 50% increase in risk predicted by genetic risk-score (one standard deviation difference), suggesting the presence of additional unaccounted genetic and environmental factors (Figure 6).
In this study, we examined the largest IgAN case-control cohorts reported to date. We first verified the five top signals identified in a recent GWAS for IgAN in independent cohorts and demonstrated robust replication of four loci, and heterogeneity at one locus. Using combined dataset of 10,755 individuals, we also identified novel risk alleles for IgAN in the DQB1/DRB1 locus and detected a significant interaction between the CFHR3/R1 and the HORMAD2 loci. We also defined a more powerful genetic risk score that explained 4.7% in disease variance across all cohorts. Finally, in examination of 85 world populations, the genetic risk score paralleled the prevalence of IgAN, confirming the known East-West cline but also led to the detection of an association of IgAN-ESRD prevalence with latitude in Europe.
While ten of twelve tested SNPs (four susceptibility loci) were robustly replicated with direction-consistent ORs across all cohorts, the TAP2/PSMB9 locus demonstrated moderately high level of heterogeneity. This locus remained genome-wide significant in the combined analyses under both fixed and random effects model. Family-based studies [17], [18], sperm typing experiments [19] and HapMap data have identified a recombination hotspot directly centered over the TAP2 gene (22 cM/Mb, 5.5-kb centromeric from the 2 SNPs selected for replication). We can therefore hypothesize that high heterogeneity at this locus is due to the unusually high rates of recombination in this region, which perturbs LD patterns between tag-SNPs and causal variants; this situation has been shown to cause a “flip-flop” phenomenon in association results [20]. Therefore, higher density of SNP coverage on either side of the recombination hotspot will be needed to guide future replication and fine mapping efforts.
In addition to the independent replication of GWAS data, we identified two new signals in the DQB1/DRB1 region that exhibit independent genome-wide significant effect in conditional analyses, providing support for multiple causal variants at this locus. These findings are consistent with previous studies of IgAN [15], [21] and other autoimmune diseases [22], [23], [24], [25], highlighting the complexity of associations in the MHC region. In our study, the strongest association signal originates in a protective haplotype tagged by rs9275596-C that carries HLA-DRB1*1501 and DQB1*602, also associated with protection against type I diabetes [24]. The causal variants underlying the other haplotypes remain obscure and their discovery will likely require comprehensive re-sequencing to define classical alleles.
Genetic interactions have been seldom described in association studies [26]. We detected a multiplicative interaction between the CFHR3/R1 and the HORMAD2 loci, which was most evident in the European cohorts, likely because the frequencies of both protective variants are considerably higher in this population. While this interaction was robust to multiple-testing correction for 7 SNPs, it will require confirmation in additional independent cohorts or via functional studies that examine whether these two loci are involved in a common biological pathway. Because the rs6677604-A allele tags a deletion in the CFHR3/CFHR1 genes, this finding suggests that the absence of these proteins abrogates the benefit imparted by HORMAD2 protective alleles. It is thus noteworthy that the HORMAD2 locus encodes several cytokines (LIF, OSM) that can interact with complement factors [27].
A seven-SNP genetic risk score explained nearly 5% of IgAN variance and demonstrated co-variation with IgAN prevalence across multiple settings. The major limitations of geospatial modeling include variable sampling density and inadequate coverage of certain geographic regions. Using the most comprehensive resources presently available for geo-genetic analyses, we found that the genetic risk score strongly paralleled the well-known East-West gradient in IgAN prevalence [3], [28], [29], [30], [31], [32]. For each of these seven risk loci, the frequency of the risk alleles was highest in East Asians, lowest in African-Americans and intermediate in European populations. Accordingly, we detected co-variation of genetic risk with IgAN-ESRD incidence and prevalence among Asian-, White- and African-Americans, which share genetic background but not environment with their ancestral populations. Representative genetic data for U.S. Native Americans was not available from HGDP nor HapMap projects, precluding a direct comparison of their risk score with prevalence. However, the USRDS data and other reports indicate a high prevalence of IgAN-ESRD in US Native Americans [33], [34], [35], [36], [37], consistent with their ancestral origin from an Asian subpopulation that migrated across the Bering land bridge over 15,000 years ago [38]. In the more homogeneous population of Northern Italy, the median risk score in the Valtrompia valley was the highest among Northern Italian populations and comparable with the Northern European scores, consistent with Valtrompia's 3.5-fold higher prevalence of ESRD, which is largely attributable to IgAN [16]. Taken together, these data strongly suggested that variation in genetic risk partly explains the variation in geo-epidemiology of disease.
Because the genetic score captured general trends in IgAN epidemiology, we also tested whether the Northward gradient in genetic risk in Europe is mirrored by higher prevalence of kidney failure from IgAN. The ERA-EDTA data, which are the most unbiased source of information available, demonstrate that Nordic countries have over 2-fold higher incidence and prevalence of IgAN-ESRD compared to the Southern European countries. Although higher risk of IgAN in Northern Europe has not been previously appreciated, similar latitudinal risk gradients in prevalence and incidence have been well established for several other immune-mediated diseases, including type 1 diabetes [39], [40], multiple sclerosis [41], [42], and inflammatory bowel disease [43]. Interestingly, these disorders share risk alleles with IgAN, suggesting that variation in common genetic risk factors may mediate variation in prevalence of autoimmune disorders. Since our analysis was limited to prevalent IgAN-ESRD in countries with epidemiological data available and only a portion of IgAN cases progresses to ESRD, studies that can better estimate the population prevalence of all IgAN can confirm these findings and better delineate epidemiological connections to other immune mediated disorders.
The genetic and environmental factors leading to the observed geospatial pattern of genetic risk and disease prevalence are not clear. The pre-modern history of IgAN is not known because this disease was only first described in 1968 [44], shortly after the discovery and application of immunofluorescence in the analysis of kidney tissue. It is well known that mucosal infections can exacerbate disease, but specific environmental factors influencing the development of IgAN are not known. Based on a recently proposed pathogenesis model, the IgAN risk loci participate in sequential processes leading to the initiation and exacerbation of IgAN [45]. This may further explain the correlation of the genetic risk score with disease epidemiology. Interestingly, many of the IgAN loci are known to exhibit opposing effects on other autoimmune conditions [14]; for example, the HLA-DQB1 and HORMAD2 risk alleles are respectively protective for systemic lupus erythematosus, and inflammatory bowel disease. Thus balancing selection, in conjunction with local environmental factors, may be responsible for maintenance of risk alleles in different populations.
The current IgAN risk score captures a greater proportion of the disease variance compared to other GWAS for kidney functions, such as a recent study of 60,000 individual that reported 13 loci explaining only 1.4% of the variance for estimated glomerular filtration rate [46]. Nonetheless, the fraction of the IgAN variation explained remains modest. For example, the one standard deviation risk-score difference between Asian- and African-Americans predicts a 50% increase in risk, yet there is over 10-fold difference IgAN-ESRD occurrence between these two groups. These data suggest that additional genetic and environmental factors influence risk. Based on the effect sizes and allelic frequencies of the discovered SNPs, we estimate that doubling the GWAS sample size is likely to find up to 7 additional loci, while tripling the sample size would identify up to 11 additional loci at genome-wide significant p-values<10−8 (calculation performed as proposed by Park et al. [47]). Conditional analyses and higher-level interaction screens of more risk loci are likely to explain additional fraction of the missing heritability and better explain differences in population prevalence of this disease.
In summary, we report results of the largest collaborative genetic study of IgAN. We confirm that the IgAN risk loci discovered in GWAS explain a significant proportion of the disease risk worldwide and likely contribute to the geographic variation in disease prevalence. Our geospatial model suggests previously unrecognized northward risk gradient in Europe, which will require further confirmation by alternative sources of prevalence data, such as country specific biopsy-registry data or kidney donor-biopsy series. The approach presented in this study may serve as a blueprint for geo-genetic modeling of other complex traits that exhibit marked geographic differences in prevalence.
This investigation was conducted according to the principles expressed in the Declaration of Helsinki. All subjects provided informed consent to participate in genetic studies and the Institutional Review Board of Columbia University as well as local ethic review committees for each of the individual cohorts approved our study protocol.
The case-control cohorts analyzed in this study were contributed by clinical nephrology centers across Europe, Asia, and North America (Table S1). All cases carried a biopsy diagnosis of IgAN defined by typical light microscopy features and predominant IgA staining on kidney tissue immunofluorescence, in the absence of liver disease or other autoimmune conditions. Each individual cohort of cases was accompanied by a control cohort of similar size, matched based on self-reported ethnicity and recruited from the same clinical center. The French cohort was composed of two sub-cohorts: the St. Etienne cohort recruited in the University North Hospital of Saint Etienne (289 cases and 244 controls), and the GN-Progress cohort recruited from the nephrology departments of the Paris region (207 cases and 159 controls). The Italian cohort was also composed of two sub-cohorts: the North Italian cohort recruited in the clinical centers of Genova, Torino, Brescia, Trento, Modena, Bologna, and Trieste (410 cases and 524 controls), and the South Italian cohort recruited in Foggia (81 cases and 80 controls). The German cohorts also represent two recruitment sites: the Stop-IgAN cohort recruited among the participants of the Stop-IgAN clinical trial based in Aachen (150 cases and 293 controls), and the Hamburg-Eppendorf cohort from northern Germany (101 cases and 80 controls). The Czech and the Hungarian cohorts were recruited through the Department of Nephrology, 1st Faculty of Medicine and General University Hospital, Charles University in Prague (245 cases and 223 controls) and the Nephrology Department of the University of Pécs (139 cases and 305 controls), respectively. The Japanese participants (264 cases and 294 controls) were recruited by the nephrologists of Niigata University. The Beijing cohort (333 cases and 289 controls) was recruited by the Renal Division of the Peking University First Hospital. Finally, our African-American cohort (34 cases and 60 controls) was recruited at Columbia University (New York, NY) and at the University of Alabama (Birmingham, AL). This smaller cohort is unique, as IgAN is exceedingly rare among individuals of African ancestry. In total, 2,253 cases and 2,621 controls were available for genotyping in the replication study. The composition and recruitment of the GWAS cohorts have been discussed in detail elsewhere [14].
The genotyping was performed by KBiosciences (Hoddeston, England). and genotype calls were determined using an automated clustering algorithm the (SNP Viewer v.1.99, KBiosciences, 2008). The genotype clusters were also examined visually across all plates, to assure lack of technical artifacts. The overall genotyping rate across all samples was 98.2%. For quality control we calculated minor allele frequencies, as well as per-SNP and per-individual rates of missingness within each case-control cohort separately. Additionally, we tested for Hardy-Weinberg equilibrium among the control groups from each cohort to assure lack of bias due to genotyping artifacts or population stratification. All SNPs included in the final analyses had minor allele frequency greater than 1%, per-SNP missingness rate less than 5%, and all passed the HWE test in controls (p>1×10−2). Individuals with more than 2 missing genotypes out of the 12 loci were also excluded from the analysis.
The participants of the smaller GN-Progress study (207 cases and 159 controls) were genotyped using the Illumina HumanCNV370-duo chip at the Centre National de Génotypage (CEA, Evry, France). The analysis of intensity clusters and genotype calls were performed using the Illumina Genome Studio software. Of 366 genotyped individuals, two cases and 1.8% of SNPs were excluded based on low call rates (<95%). The overall genotyping rate was 99.6%. In total, 6 of 12 SNPs analyzed for replication were also present on the Illumina HumanCNV370-duo chip. The genotypes at the reminder loci were imputed using the phased HapMap-III CEU reference dataset (see Web Resources). The imputation was performed simultaneously for cases and controls, using MACH 1.0 software (see Web Resources). We used a standard single-step imputation approach, with 60 rounds of Markov Chain iterations to estimate the crossover maps, error rate maps, and all missing genotypes across each analyzed locus. The imputed SNPs had an excellent imputation quality, with an average estimated correlation between imputed genotypes and experimental genotypes of 0.98 (range 0.94–1.0). Consequently, association analyses using either the allelic dosage approach that accounts for imputation uncertainty, or the most likely genotype approach yielded similar results. Therefore, the most probable genotype calls were used in the downstream analyses. In the final quality control step, we compared the allelic frequencies and effect estimates between the two French cohorts (GN-Progress and St. Etienne). For each locus, we observed nearly identical frequencies among cases and controls and the odds ratios were homogenous between the two cohorts. The formal heterogeneity tests were not statistically significant for any of the tested loci (Cochrane's Q-test P>0.05, average I2 = 0). Therefore, these two cohorts were combined into a single cohort of 493 cases and 402 controls.
Similarly to the French cohorts, there was no significant heterogeneity at any of the loci for the two smaller German cohorts (STOP-IgAN and Hamburg-Eppendorf), and these were also combined into a single cohort of 249 cases and 372 controls. Analysis of the Northern and Southern Italian cohorts suggested some heterogeneity at 3 out of 12 SNPs (I2 = 40–50%). Although these observations were not statistically significant (Q-test P>0.05), we used a conservative stratified approach for all downstream analyses for these two cohorts. The final summary of all study cohorts before and after quality control is provided in Table S1.
We performed a power calculation for the final replication cohort size of 4,789 individuals (2,228 cases/2,561 controls) as a function of disease allele frequency and genotype relative risk (Table S2). The power was calculated in reference to a protective allele, with the range of allelic frequencies and effects comparable to the ones observed in the original GWAS. Assumptions included disease prevalence of 1%, log-additive model, no heterogeneity, and alpha = 0.01 (Bonferroni-adjusted considering five independent loci tested). This analysis confirmed that our study had ample power (nearly 100% for most loci) to replicate the associations observed in the initial GWAS. The power calculations were performed using QUANTO v.1.2 software [48].
The primary association analyses were performed using PLINK version 1.07 [49]. Similar to GWAS, we selected a standard 1-df Cochran-Armitage trend test as the primary association test. We also estimated the per-allele odds ratios and 95% confidence intervals for all tested SNPs within each individual cohort. The results across multiple cohorts were combined using an inverse variance-weighted method under a fixed-effects model (PLINK), as well as using a random effects model as proposed by Han and Eskin (METASOFT) [50]. We also tested for heterogeneity across cohorts by performing a formal Cochrane's Q heterogeneity test as well as by estimating the heterogeneity index (I2) [51].
The conditional association tests of the HLA loci were performed after controlling for the genotypes of the conditioning SNPs within each cohort using logistic regression (PLINK). The adjusted (conditioned) effect estimates were then combined across cohorts using a fixed effect meta-analysis considering no significant heterogeneity across these loci. For the purpose of validation of this approach, we also combined the results by adding cohort information as an additional covariate in the stratified analysis within the logistic regression framework. As expected, the results of both approaches were similar.
These analyses were carried out in PLINK v1.07 [49]. Haplotypes were first phased using EM algorithm across the HLA-DQB1, HLA-DQA1, HLA-DRB1 region. The haplotype frequencies were estimated in the cases and controls separately, as well as jointly in the entire cohort. Only common haplotypes with overall frequency >1% were included in the association tests. Global haplotype association test was performed using a χ2 test with n-1 degrees of freedom for n common haplotype groups. The ORs and the corresponding 95% confidence intervals were estimated in reference to the most common haplotype (GCAT, frequency ∼35%).
To explore the possibility of interactions between the 7 independent risk variants, we screened all possible pairwise interaction terms for association with disease within the framework of logistic regression models (R version 2.10). As a screening test, we used 1-df LRT to compare two nested models: one with main effects only and one with main effects and a multiplicative (logit-additive) interaction term. We included cohort membership as a fixed covariate in both of these models. For this analysis we selected a Bonferroni-adjusted significance of 2.4×10−3, a conservative threshold that accounts for all 21 pairwise interaction terms tested. Significant interactions from this analysis were also tested using a 4-df genotypic interaction test. In this test, we compared a model with allelic effects, dominant effects, and their interaction terms with a reduced model with no interaction terms. We followed the coding proposed by Cordell and Clayton: for each SNP i we modeled its allelic effect xia by coding the genotypes AA, AB, and BB as xia = −1, 0, 1; we modeled dominance effects as xid = −0.5, 0.5, −0.5 for the genotypes AA, AB, and BB, respectively [52].
Each study participant was scored for the number of risk alleles and the distributions of protective alleles were compared between cohorts of different ethnicity. Only individuals with complete genotype information at the 7 scored loci (14 alleles) were included in this analysis. The distributions were analyzed separately for cases and controls. A χ2 goodness-of-fit test was used to derive p-values for comparison of distributions. Because of a relatively small number of individuals at the tails of the distributions, for the purpose of statistical testing the tails of the distributions were binned into single-bin categories to achieve expected cell counts >5.
To confirm the results of conditional analyses and refine the genetic risk score proposed in the original GWAS, we subjected the genotype data from the entire cohort to a stepwise regression algorithm that selects significant covariates for the best predictive regression model based on Bayesian Information Criterion (the step function, R version 2.10). At model entry, we included all 12 genotyped SNPs, all 21 tested interactions, as well as cohort membership as a fixed covariate. Consistent with the results of our conditional analysis, the stepwise algorithm retained only the 7 SNPs exhibiting an independent effect along with the rs6677604*rs2412971 interaction term. All other terms were automatically dropped from the regression model.
The risk score was calculated as a weighted sum of the number of protective alleles at each locus multiplied by the log of the OR for each of the individual loci from the final fully adjusted model. Only individuals with non-missing genotypes for all 14 alleles were included in this analysis. The risk score was standardized across all populations using a z-score transformation, thus the standardized score represented the distance between the raw score and the population mean in units of standard deviation. The percentage of the total variance in disease state explained by the risk score was estimated by Nagelkerke's pseudo R2 from the logistic regression model with the risk score as a quantitative predictor and disease state as an outcome. The C-statistic was estimated as an area under the receiver operating characteristic curve provided by the above logistic model. These analyses were carried out with SPSS Statistics version 19.0.
For this purpose, we used publicly available genotype data of the Human Genome Diversity Panel (HGDP; 1,050 individuals representative of 52 worldwide populations), HapMap III (1,184 individuals representative of 11 populations), along with healthy controls genotyped as part of this study (4,547 individuals representative of 25 recruitment sites). The HGDP individuals have been previously genotyped for 660,918 markers using Illumina 650Y arrays (Stanford University). First, SNPs with genotyping rate<95% and samples with an overall call rate<98.5% were removed from the genome-wide data. Only 1,042 individuals with all 14 non-missing alleles at the 7 analyzed risk score loci were included in the final analysis. The geographic coordinates for the HGDP populations were downloaded from the CEPH website (see Web Resources). The HapMap III genotype data have been generated using two platforms: the Illumina Human1M (Wellcome Trust Sanger Institute) and the Affymetrix SNP 6.0 (Broad Institute). These files were merged into a single dataset of 1,440,616 markers, from which we removed (1) SNPs with genotyping rate<95%, (2) samples with an overall call rate<98.5%, (3) all non-founders from mother-father-child trios, and (4) individuals with missing genotypes at any of the 7 SNP loci used for risk scoring. In the global geospatial analyses, we excluded US-recruited individuals of African American (ASW), European (CEU), and Asian (CHD) ancestry considering non-specific geographic origin of these populations. However, the population of Guajarti Indians recruited in Houston (GIH) was mapped to the northwestern part of the Indian subcontinent, as these individuals reported having at least three out of four Gujarati grandparents, speak the Gujarati language, and trace their ancestry to the region of Gujarat. In total, 730 HapMap III individuals representative of 8 populations met our selection criteria and were included in the final analysis.
Because many European populations are underrepresented in HGDP and HapMap III datasets, we also included a total of 4,462 healthy controls from the GWAS and replication studies that were collected across 25 recruitment centers participating in our studies. Similar to the above criteria, only individuals with non-missing genotypes at all 7 scored SNPs were included in this analysis. The geographic coordinates for our populations were based on the location of recruitment centers and determined with Google Earth (see Web Resources). This resulted in a final dataset of 6,319 individuals sampled across 85 worldwide populations for geospatial analysis.
We fitted a 3rd degree polynomial trend surface based on the latitude, longitude, and median standardized risk score for each of the 85 populations using least squares approach (Spatial package version 7.3-2, R version 2.10). For higher resolution maps, we used kriging technique and accounted for the possibility of spatial correlation of errors among more densely sampled populations by modeling the covariance function in an exponential form. The estimated risk surfaces were projected over the major continents using Maps package version 2.1–6 (R version 2.10).
We obtained case counts of prevalent and incident ESRD stratified by primary renal diagnosis and by ethnicity from the United States Renal Data Systems (2011 USRDS Data Atlas, see Web Resources). For Europe, we obtained prevalent and incident ESRD case counts from the European Renal Association and European Dialysis and Transplant Association (ERA-EDTA Renal Registry, see Web Resources). Comprehensive data were available for a total of 13 European countries participating in this registry. We calculated the prevalence of ESRD due to IgAN using three definitions: (1) proportion of all ESRD cases attributable to IgAN, (2) proportion of all ESRD cases from primary glomerulonephritis attributable to IgAN, and (3) total number of ESRD cases due to IgAN per million population (PMP). The prevalence data for both USRDS and ERA-EDTA datasets were calculated for the same timepoint of December 31st, 2009. The incidence of ESRD due to IgAN was estimated using all the available data over a 3-year period for the ERA-EDTA registry (2007–2009), and a 5-year period for the USRDS registry (2005–2009). For correlation of genetic risk score with disease prevalence in the US, we scored representative samples of the three major US ethnic groups: 303 US Caucasians (CEU founders from HapMap-3 and healthy US controls from our original GWAS), 103 African-Americans (ASW founders from HapMap-3 and healthy controls from this study), and 74 Asian-Americans (CHD founders from HapMap-3). For correlation of genetic risk with disease prevalence in Europe, we calculated median standardized risk scores at a country level for 13 European countries for which we obtained genotype data. We confirmed the South-North disease gradient by regressing the prevalence and risk score data against each country's average latitude. The correlation and regression analyses were conducted in SPSS Statistics version 19.0.
HAPMAP PHASE III Data: http://hapmap.ncbi.nlm.nih.gov/downloads/phasing/2009-02_phaseIII
HGDP Genotype Data: http://hagsc.org/hgdp
HGDP Population Data: http://www.cephb.fr/en/hgdp
MACH: http://www.sph.umich.edu/csg/abecasis/MaCH
PLINK: http://pngu.mgh.harvard.edu/~purcell/plink
METASOFT: http://genetics.cs.ucla.edu/meta
CRAN: http://cran.r-project.org
GOOGLE EARTH: http://www.google.com/earth
SPATIAL: http://cran.r-project.org/web/packages/spatial
MAPS: http://cran.r-project.org/web/packages/maps
USRDS Data Atlas 2011: http://www.usrds.org/atlas.aspx
ERA-EDTA Registry Annual Report 2009: http://www.era-edta-reg.org
|
10.1371/journal.pntd.0005692 | Caprine brucellosis: A historically neglected disease with significant impact on public health | Caprine brucellosis is a chronic infectious disease caused by the gram-negative cocci-bacillus Brucella melitensis. Middle- to late-term abortion, stillbirths, and the delivery of weak offspring are the characteristic clinical signs of the disease that is associated with an extensive negative impact in a flock’s productivity. B. melitensis is also the most virulent Brucella species for humans, responsible for a severely debilitating and disabling illness that results in high morbidity with intermittent fever, chills, sweats, weakness, myalgia, abortion, osteoarticular complications, endocarditis, depression, anorexia, and low mortality. Historical observations indicate that goats have been the hosts of B. melitensis for centuries; but around 1905, the Greek physician Themistokles Zammit was able to build the epidemiological link between “Malta fever” and the consumption of goat milk. While the disease has been successfully managed in most industrialized countries, it remains a significant burden on goat and human health in the Mediterranean region, the Middle East, Central and Southeast Asia (including India and China), sub-Saharan Africa, and certain areas in Latin America, where approximately 3.5 billion people live at risk. In this review, we describe a historical evolution of the disease, highlight the current worldwide distribution, and estimate (by simple formula) the approximate costs of brucellosis outbreaks to meat- and milk-producing farms and the economic losses associated with the disease in humans. Successful control leading to eradication of caprine brucellosis in the developing world will require a coordinated Global One Health approach involving active involvement of human and animal health efforts to enhance public health and improve livestock productivity.
| Human brucellosis is an ancient disease that has had different names throughout time based on the main clinical symptom (fever) and the geographical location: Malta fever, Mediterranean fever, Undulant fever, Gibraltar fever, Rock fever, and Neapolitan fever, among others. Retrospective studies have demonstrated that goats have been the hosts of B. melitensis for centuries, with evidence of its zoonotic potential early in evolution. Since domestication of goats (and also sheep), the incidence of human brucellosis has been on the rise, becoming endemic in resource-limited settings. Today, millions of goats and approximately half of the human population worldwide live at risk. For that reason, more effective prevention and control measures, such as affordable vaccines, more sensitive and specific diagnostic techniques, and the control livestock movement, among others, are desperately needed to control and eradicate brucellosis in goats and to prevent human brucellosis.
| Brucella melitensis is the etiological agent of caprine brucellosis, an infectious zoonotic disease with significant economic impact on both the livestock industry and public health. Worldwide, there are approximately 1 billion goats, with an increase of the population by more than 20% in the last 10 years. Approximately 90% of goats are located in the developing world, where they are considered one of the most important sources of protein for humans [1]. Caprine brucellosis has been controlled in most industrialized countries; however, this disease remains endemic in resource-limited settings, where small ruminants are the major livestock species and the main economical livelihood, such as the Mediterranean region, the Middle East, Central Asia, sub-Saharan Africa, and parts of Latin America [2]. Among the different Brucella spp. capable of causing disease in humans (B. abortus, B. melitensis, B. canis and B. suis), B. melitensis is the most virulent [3]. Human brucellosis has had different names throughout time based on the main clinical symptom (fever) and the geographical location: Malta fever, Mediterranean fever, Undulant fever, Gibraltar fever, Rock fever, and Neapolitan fever, among others [4]. Brucellosis is considered a severely debilitating and disabling illness that results in high morbidity with intermittent fever, chills, sweats, weakness, myalgia, abortion, osteoarticular complications, endocarditis, depression anorexia, and low mortality. Due to causing a protracted, incapacitating disease with minimal mortality, the low infectivity dose required to cause infection (10–100 colony-forming units), and the potential for aerosol dissemination, B. melitensis was considered a potential bioterrorist agent early in the 20th century [5]. Gradually, biological warfare moved on, and Brucella’s perceived status as a potential agent of bioterrorism declined, until the World Trade Center attack in 2001 brought bioterrorism back to the public’s attention. Nowadays, B. melitensis possession and use is still strictly regulated in the United States of America, Canada, and some European countries. Conversely, more than half a million new brucellosis cases per year occur naturally in the populations of developing areas of the world, a number which is thought to be severely underestimated [6].
A MEDLINE (via Pubmed) and SCOPUS online databases search for articles with “Brucella melitensis” or “brucellosis” and “goats” or “small ruminants” as keywords with no date limit and published up to December 31st, 2016, was performed. An additional internet search was done in Google without language restriction, using those terms and including country names. Early reports were obtained from original printouts from the reference list of selected articles and printed books.
Despite the first scientific evidence that goats were the reservoir host of B. melitensis in 1905, several observations would indicate that goats have been the host of B. melitensis for centuries [7]. Phylogenetic studies suggest that brucellosis in goats emerged in the past 86,000 to 296,000 years through contact with infected sheep [8]. Interestingly, to support this observation, a recent study found lesions in vertebral bodies of an Australopithecus africanus (who lived 2.5 million years ago) consistent with brucellosis, where the source of infection could be the consumption of infected tissues from wild animals [9]. Subsequently, the closer association of humans with goats (and also sheep) due to domestication around 10,000 years ago [10] favored an increase in the incidence of human brucellosis. As essential resources for human survival, goat and sheep herds moved along with human communities from the Fertile Crescent in Southwestern Asia to lands around the Mediterranean Sea [11], where Phoenician traders might have contributed to the spread of B. melitensis infection throughout the Mediterranean littoral and islands during the first millennium B.C. [12]; it was then introduced to the Americas around the 16th century by Spanish and Portuguese conquerors [11,13].
The first written evidence of goat brucellosis could be inferred from the first description of 2 human cases of brucellosis. In the 4th century B.C., in his Epidemics book, Hippocrates II described 2 cases of a 120-day fever in people living in the Mediterranean littoral, most likely associated with the consumption of raw milk or derivatives of B. melitensis-infected sheep or goats [14]. Another testimony of the ancient presence of caprine brucellosis comes from preserved evidence from the volcanic eruption of Mount Vesuvius in Italy on August 25th in the year 79 A.D. Scanning electron microscopy examination of remnants of carbonized cheeses revealed cocci-like forms consistent with B. melitensis, while an anthropological examination of human skeletal remains from that incident revealed an arthritic condition consistent with brucellosis [7]. References to and vivid descriptions of clinical cases compatible with human brucellosis were continuously reported in histories of military campaigns and hospital reports [15]. However, the identification of the etiological agent, the reservoir, and the epidemiology of the disease was not unraveled until the second half of the 19th century, when the British government decided to find a solution for their troops stationed on the island of Malta that annually suffered substantial losses caused by the so-called “Malta fever.” In 1859, British Army surgeon Jeffery Marston contracted what he called “Mediterranean remittent fever” [16]. After recovering, he described his own case in great detail, being the first author to clinically and pathologically differentiate human brucellosis from typhus, typhoid, and other prevalent fevers [15]. In 1884, the Australian-born British physician David Bruce was deployed to Malta to investigate the cause of “Malta fever” (later called brucellosis in his honor). Late in 1886, using a microscope, he observed a great number of micrococci in a fresh preparation of the splenic pulp of soldiers who had died from the disease [17]. One year later, Sir Bruce isolated the causative agent of “Malta fever” (which he called Micrococcus melitensis and then renamed Brucella melitensis) from samples of spleens of 4 patients inoculated into Koch’s nutrient agar and was able to reproduce the disease in monkeys following Koch’s postulates [18]. A few years later, Professor Almroth Edward Wright developed a serum agglutination test and demonstrated the presence of specific agglutinins in the blood of infected patients, which helped differentiate those who suffered “brucellosis” from those with typhoid (cholera) or malarial fever [19]. The use of this serological test in goats provided the first insights into the epidemiology of the disease. In 1904, a Public Health Officer of Malta discovered that the blood of goats that supplied milk to people that had contracted “Malta fever” had agglutinins against M. melitensis, and a posterior survey indicated that around 50% of Malta’s goats’ blood reacted to this microorganism. This observation suggested that goats were susceptible to natural infection with M. melitensis. Based on all knowledge available on brucellosis, the Greek physician Themistokles Zammit hypothesized that goats were susceptible to Malta fever and that the disease spread from goats to human. To test his hypothesis, Zammit fed seronegative, healthy goats with agar cultures of M. melitensis mixed into their food. Goats became seropositive to M. melitensis after 20 days or more, and Brucella was isolated from the milk, blood, and urine of infected animals without any clinical manifestation of the disease [20]. This simple assay demonstrated that goat milk was the disseminating vehicle of the bacteria, rather than an insect vector, and helped to build the epidemiological link of “Malta fever” to the consumption of this product. This observation was further confirmed after its ban from the diet of the Malta garrison significantly reduced the incidence of brucellosis in the army and naval forces compared to the general population of Malta that continued to consume contaminated dairy products. Later on, in 1918, Alice Evans demonstrated the similar characteristics between the M. melitensis and the etiological agent of bovine epizootic abortion, the “abortus bacillus” (now Brucella abortus), isolated by a Danish veterinarian Bernhard Bang in 1896, and based on that, both agents were included under the same bacterial genus (Brucella) in honor of David Bruce, in 1920. Major events of caprine brucellosis and its relationship with public health throughout history are summarized in Table 1.
Caprine brucellosis refers to goat herds infected with B. melitensis. Goats can be susceptible to B. abortus infection [22,23,24] under particular epidemiological situations (for instance, when goats live in close contact with B. abortus-infected cattle); however, these individuals don’t sustain the infection in the herd. Similarly, B. suis isolations from goats have seldom been reported, but in recent times, they have not been further documented [25]. B. melitensis comprises 3 biovars (1–3), distinguished solely by their immunochemical reactions with monospecific anti-lipopolysaccharide (LPS) A- and M-determinant sera [26,27]. Available information indicates that most infections are caused by biovars 1 and 3 [28], both of which seem to have similar virulence for goats and humans.
Prevalence of caprine brucellosis around the world has been reported and referenced by others [28,29,30,31,32]. The disease is present in 5 out of the 7 continents (South and North America, Europe, Asia, and Africa). Despite being under control in most industrialized countries, it remains a major problem in the Mediterranean region, the Middle East, Central and Southeast Asia, sub-Saharan Africa, and parts of Latin America (Fig 1).
As expected, prevalence of human brucellosis is also high in those regions where goat brucellosis occurs [6]. The disease has been historically underreported, probably because low-income countries prioritize other diseases or lack facilities, human capabilities, and/or specific tests that would otherwise underpin diagnoses and research. Over the last 15 years, the infection has re-emerged, in particular in Eastern Europe, the Balkans, and Eurasia [2]. Table 2 shows those countries where caprine brucellosis (i.e., presence of anti-Brucella antibodies, B. melitensis isolation, or Brucella DNA detection from goat samples) or brucellosis in humans, sheep, or cattle due to B. melitensis infection have been reported in recent years (2005–present). Historically, B. melitensis biovar 1 is predominant in Latin America [28,33], while biovar 2 is predominant in the Middle East together with biovar 3, which is also more common in European and African Mediterranean countries, Eurasia, and China [2,28,31,34,35]; biovars 1 and 3 seem to be equally present in India [36,37]. Unfortunately, there are few studies addressing the characterization of isolates from sub-Saharan countries.
In the Americas, Brucella melitensis was most likely introduced around the 16th century via the infected goats and sheep of Spanish and Portuguese conquerors [11]. Today, B. melitensis is endemic in some regions of Mexico, Peru, and Argentina [28] and has also been reported in Ecuador and Venezuela [41,48]. Caprine brucellosis is apparently absent in Central America, Bolivia, Paraguay, and Brazil, although this epidemiological situation is not confirmed [129]. Goat herds from the USA, Canada, Colombia, Chile, and Uruguay are free from B. melitensis infection, and human cases in these countries are clearly associated with international travelers or infected food imported from endemic regions [6].
Despite intense joint efforts to eliminate B. melitensis from goat flocks in Europe, the disease still occurs in Portugal, Spain, France, Italy, the Balkans, Bulgaria, and Greece. Northern and Central European countries like the United Kingdom, Belgium, the Netherlands, Denmark, Germany, Austria, Switzerland, the Czech Republic, Hungary, Poland, Romania, Sweden, Norway, and Finland, among others, are officially free of the disease [129].
In Asia, brucellosis is broadly distributed. Except for Japan and the Republic of Korea (South Korea), where the disease has never been reported, caprine brucellosis is officially recognized in several countries on the continent, such as Turkey, Israel, Jordan, Iraq, Iran, Armenia, Georgia, Afghanistan, Russia, and Mongolia, among others (see references in Table 2), and is also known to be endemic in countries like Syria, Lebanon, India, China, Indonesia, Myanmar, etc., where no public information is available or the distribution of the information is restricted [28,29,102,129].
In Africa, caprine brucellosis is endemic in Mediterranean countries like Morocco, Algeria, Tunisia, Libya, and Egypt, and also in those countries located in the eastern part of the continent, such as Sudan, Eritrea, Ethiopia, Somalia, Kenya, Uganda, and Tanzania (see Table 2 for references). Unfortunately, there is no information available from Central and West African countries like Chad, Congo, Angola, Zambia, Cameroon, Mali, Cote d’ Ivoire, Guinea, and Senegal, among others, where goats are abundant [130]. Altogether, the information above indicates that the knowledge regarding distribution of caprine brucellosis as well as the presence of B. melitensis around the world is sparse, especially in some areas of the Americas, Asia and Africa. The lack of useful epidemiological data must stimulate official veterinary services and public health officers to collect and share data for designing control and eradication plans.
Since brucellosis is considered a neglected disease that significantly affects countries where resources are limited, there are only a few studies that measure the economic impact of brucellosis in small ruminants. Sulima and Venkataraman (2010) and Singh et al. (2015) estimated the annual loss in India at Rs. 2,121 per goat (around US$39) and at US$71 million total, respectively [131,132]. Brisibe et al. (1996) calculated a loss of US$3.2 million per annum in 2 states of Nigeria [133], and more recently, Bamaiyi et al. (2015) reported the annual economic impact in Malaysia due to caprine brucellosis at almost US$2.6 million [134]. However, every publication utilizes different criteria, which makes comparisons difficult. A simple analysis of economic impact of caprine brucellosis on meat goat farmers can be calculated by taking into consideration the culling of animals serologically positive for Brucella, the abortions and stillbirths, the cost of veterinary services, and miscellaneous factors arising from brucellosis on farms. The economic loss for culling 1 reactor animal is equal to the market price of a healthy goat purchased for its replacement, minus the amount perceived for selling the positive reactor to a slaughterhouse. An abortion or stillbirth must be considered as loss of profit and its value calculated as the market value of a 6-month-old kid (which weighs around 10 kg). Veterinary services include visits to the farm, professional assistance, and serological surveys, while miscellaneous factors—such as man hours for taking care of ill flocks, reduced weight gain, the increased morbidity of weak offspring and low birth weight kids, any interest paid on money borrowed from banks, etc.—are variable and, therefore, difficult to predict and calculate. Based on these premises, it is possible to roughly estimate the economic impact of a brucellosis outbreak in a meat goat herd. For instance, in Argentina, the impact of a brucellosis outbreak in a flock of 100 goats, in which 25 does abort and 10 others become serologically positive, would be:
There are some differences if the analysis is done for a dairy goat farm. For instance, the market price of healthy milking goats (Anglo Nubian, Saanen, Toggenburg) is higher than for meat goats, and the loss of milk yield due to culled does has to be taken into account as well. Thus, a conservative impact of a brucellosis outbreak in a herd of 100 milk goats, in which 25 does abort and 10 others become serologically positive, would be:
The estimated cost will vary with the location, production system, facilities, and miscellaneous factors included. The calculations need to include additional losses due to the socioeconomic and indirect health effects of the disease in humans. Still today, human brucellosis is an underreported disease, often mistaken for malaria and typhoid fever (Halliday et al., 2015). WHO estimates around 500,000 new cases reported and an equal number of nonreported cases of human brucellosis each year, a high proportion of them caused by B. melitensis. Vulnerable populations include not only dairy goat and sheep farmers, small ruminant ranchers (especially in marginalized goat-keeping communities), and veterinarians and abattoir workers, but also lab personnel and consumers of unpasteurized dairy products.
Economic losses caused by the disease in humans arise from the cost of hospital treatment, medicines, patient out-of-pocket treatment expenses, and loss of work days and income due to illness. In Spain, losses by brucellosis were estimated at 790,000 pesetas per patient (US$5,030) [135], while in New Zealand, the approximated cost per patient was NZ$3,200 (US$2,250) [136]. In Africa, the cost of treating a patient ranges from 9 EUR in Tanzania to 200 EUR in Morocco and as much as 650 EUR in Algeria [128]. In Argentina, the annual treatment cost of brucellosis was estimated to be US$4,000 [137].
Traditional recommended antibiotic treatment for human brucellosis consists of 100 mg of doxycycline twice a day per os for 45 days combined with 1 g of streptomycin daily intramuscular (IM), 15 to 21 days; gentamycin 5 mg/kg/day (300–350 mg) IM, 7 to 10 days; or, alternatively, rifampicin 15mg/kg/day (600–900 mg) per os for 45 days [3]. Today, in Argentina, the cost for antibiotic treatment for a single patient is approximately US$200–US$300. This value does not include lost profit, laboratory analysis and X-ray images, medical expenses, and other miscellaneous expenses. Considering a complete health treatment, the cost for every brucellosis-infected person is up to US$1,000.
Brucellosis in small ruminants remains a significant burden on animal and human health in the developing world. Small ruminant owners and governments where brucellosis is endemic do not usually have enough economic resources nor technical expertise or facilities to afford control or eradication campaigns. On the other hand, B. melitensis is the Brucella species with the highest zoonotic potential, and in humans, it frequently presents nonspecific clinical symptoms similar to other infectious diseases that are also present in brucellosis-endemic areas [138]. Thus, the challenge of clinical–differential diagnosis adds to the inequality of accessible healthcare facilities in most developing countries. These cumulative issues contribute to brucellosis remaining endemic and neglected in resource-limited regions of the world.
The future major challenges include developing a more effective and affordable DIVA (differentiating infected from vaccinated animals) vaccine against small ruminant brucellosis for massive protection in endemic areas. Undoubtedly, this goal must be accompanied by an integrated control strategy with a massive vaccination campaign, strict epidemiological surveillance, and controlled movement of animals. Meanwhile, current efforts must focus on controlling new outbreaks using available tools to prevent B. melitensis transmission to humans.
|
10.1371/journal.pcbi.1007206 | PrediTALE: A novel model learned from quantitative data allows for new perspectives on TALE targeting | Plant-pathogenic Xanthomonas bacteria secrete transcription activator-like effectors (TALEs) into host cells, where they act as transcriptional activators on plant target genes to support bacterial virulence. TALEs have a unique modular DNA-binding domain composed of tandem repeats. Two amino acids within each tandem repeat, termed repeat-variable diresidues, bind to contiguous nucleotides on the DNA sequence and determine target specificity. In this paper, we propose a novel approach for TALE target prediction to identify potential virulence targets. Our approach accounts for recent findings concerning TALE targeting, including frame-shift binding by repeats of aberrant lengths, and the flexible strand orientation of target boxes relative to the transcription start of the downstream target gene. The computational model can account for dependencies between adjacent RVD positions. Model parameters are learned from the wealth of quantitative data that have been generated over the last years. We benchmark the novel approach, termed PrediTALE, using RNA-seq data after Xanthomonas infection in rice, and find an overall improvement of prediction performance compared with previous approaches. Using PrediTALE, we are able to predict several novel putative virulence targets. However, we also observe that no target genes are predicted by any prediction tool for several TALEs, which we term orphan TALEs for this reason. We postulate that one explanation for orphan TALEs are incomplete gene annotations and, hence, propose to replace promoterome-wide by genome-wide scans for target boxes. We demonstrate that known targets from promoterome-wide scans may be recovered by genome-wide scans, whereas the latter, combined with RNA-seq data, are able to detect putative targets independent of existing gene annotations.
| Diseases caused by plant-pathogenic Xanthomonas bacteria are a serious threat for many important crop plants including rice. Efficiently protecting plants from these pathogens requires a deeper understanding of infection strategies. For many Xanthomonas strains, such infection strategies depend on a special class of effector proteins, termed transcription activator-like effectors (TALEs). TALEs may specifically activate genes of the host plant and, by this means, re-program the plant cell for the benefit of the pathogen. Target sequences and, consequently, target genes of a specific TALE may be predicted computationally from its amino acids. Here, we propose a novel approach for TALE target prediction that makes use of several insights into TALE biology but also of broad experimental data gained over the last years. We demonstrate that this approach yields a higher prediction accuracy than previous approaches. We further postulate that a strategy change from a restricted search only considering promoters of annotated genes to a broad genome-wide search is feasible and yields novel targets including previously neglected protein-coding genes but also non-coding RNAs of possibly regulatory function.
| Many crop plants including rice can be infected by Xanthomonas bacteria causing disease in the affected plants, which results in substantial yield losses. Many strains of Xanthomonas oryzae pv. oryzae (Xoo) and Xanthomonas oryzae pv. oryzicola (Xoc) express a specific type of effector protein called transcription activator-like effectors (TALEs). TALE proteins function as transcription factors in infected host cells [1], and contain a nuclear localization signal, a DNA-binding domain, and an activation domain. The DNA-binding domain consists of tandem repeats that bind to the promoter of plant target genes. Each repeat consists of approximately 34 highly conserved amino acids (AAs), except for the amino acids at position 12 and 13, which are termed repeat variable diresdue (RVD) and are responsible for DNA specificity. The repeat domain forms right-handed superhelical structure, while the RVD is situated within a loop accessing the DNA [2, 3]. Each RVD binds to one nucleotide of the target box [4, 5], where amino acid 13 binds to the sense strand and amino acid 12 stabilizes the repeat structure. Hence, the specificity of each TALE is determined by its RVD sequence. In addition, most known target boxes are directly preceeded by a ‘T’, while ‘C’ and ‘A’ occur with decreasing frequencies, which is also referred to as “position 0” of the target box.
Some repeats deviate from the common length of 34 AAs and have, for this reason, been termed aberrant repeats. Aberrant repeats may loop out of the repeat array when a TALE binds to its DNA target box and by this means allow for increased flexibility, also binding to frame-shifted target boxes [6].
Different Xoo and Xoc strains express different repertoires of TALEs, where a single strain may host up to 27 TALEs [7–10].
Naturally occurring TALEs may activate susceptibility (S) genes that are responsible for bacterial growth, proliferation and disease development, but also disease resistance (R) genes [1].
The names of TALEs and TALE classes are based on the nomenclature introduced by the tool AnnoTALE [11]. TALEs are clustered according to the similarity of their RVD sequence and divided into classes.
Target boxes upstream of all known major virulence targets are located in forward orientation relative to the transcription start site (TSS). Recently, target boxes of TALEs have been reported to be also functional in reverse orientation relative to the transcription start site (TSS) of their target gene [12, 13]. However, reverse binding seems to be rather an exception than a general rule [13]. Accurate predictions of target boxes of TALEs are important for studying naturally occurring TALEs and determining their virulence targets, but also for the identification of target and off-target sequences of artificially designed TALEs. Over the last years, several tools have been designed for the in-silico prediction of TALE target boxes based on the RVD sequence of a given TALE and, subsequently, for the identification of target genes.
The TALE-NT suite includes “Target Finder”, a tool for predicting target boxes of TALEs based on their RVD sequence. It is available as online or command line application (http://tale-nt.cac.cornell.edu/) [14, 15]. In Target Finder, predictions are based on a position weight matrix calculated from frequencies of naturally occurring RVD-nucleotide associations. The user can choose whether the target box should start with nucleotide T or C.
Talvez is another prediction tool that uses PWMs to model RVD-nucleotide interactions [16]. It differs from Target Finder in deriving specificities of rare RVDs from those of common RVDs with the same 13th amino acid. Target sequences may only begin with nucleotide T or C, with a lower score assigned in the case of cytosine. In addition, Talvez may explicitly model that mismatches are tolerated to a larger degree if these are located near the C terminus [17]. Users of Talvez can choose between web-based and command line applications.
TALgetter [18] uses a local mixture model to predict TAL target sequences. The specificities were learned from 267 pairs of TALEs and target sites with qualitative information whether the pair is functional or not. According to Streubel et al. [19], the efficiencies of different RVDs are non-identical. The TALgetter model adapts a similar concept using an importance term, which is learned independently from the specificity of each RVD. TALgetter is implemented within the Java framework Jstacs [20], and is available as online and command line program.
In the web tool SIFTED [21], specificity data from a large-scale study using protein-binding microarrays (PBMs) were used for training model parameters. For this purpose, 21 TALEs constructed exclusively from the most common four RVDs (NI, HD, NN, NG) were designed and their binding specificity measured on ≈ 5,000-20,000 DNA sequences per protein using PBMs. However, we will not consider SIFTED in the remainder of this manuscripts, as the SIFTED web server is currently unavailable and the limited set of RVDs included into SIFTED does not cover the entire spectrum of those occurring in natural TALEs.
Predictions of all of these approaches still comprise a substantial number of false positive predictions, whereas some of the known target genes cannot be detected by these approaches. During the last years, several quantitative studies of TALE binding and transcriptional activation have been published. The studies included quantitative analyses of target gene activation by TALEs spanning naturally occurring RVDs [19, 22], specificities at position 0 of target boxes [23], complete exploration of all possible combinations of amino acids at RVD positions [24, 25], and systematic analyses of those RVDs frequently used in designer TALEs [21].
In this paper, we aim at developing a novel approach for modelling TALE target specificities based on these quantitative data. This approach, called PrediTALE, explicitly captures putative dependencies between adjacent RVDs, dependencies between the first RVD and position 0 of the target box, and also includes positional effects of mismatch tolerance. In contrast to previous approaches, model parameters are adapted by minimizing the difference between prediction scores and quantitative measurements for pairs of TALEs and target boxes. Like previous approaches, PrediTALE also predicts target boxes in reverse strand orientation relative to the TSS, but applies a small penalty term in this case, following the assumption that functional reverse target boxes are rather rare in planta. PrediTALE is the first approach to account for aberrant repeats when predicting TALE targets.
Pairs of TALEs and putative target boxes were collected from systematic, quantitative experiments reported in [19, 22–25]. Data were further processed as detailed in S1 Text. Data were grouped by TALE, and the global weight was computed as the maximum assay value for the current TALE divided by the maximum assay value reported for all TALEs with the same 13th AA at any position in the current assay. Target values were computed as the assay value of the current pair of TALE and target box divided by maximum assay value over all tested target boxes for the current TALE.
While the normalization of target values has a mostly technical background as it simplified the selection of initial values during numerical optimization of our model (see below), the definition of global weights influences the optimization result. The choice of global weights has been motivated by the observation that some TALE architectures (e.g., those with long successions of identical RVDs, or 12th AAs not occurring in nature) show a generally lower activity than others, which also affects the influence of measurement noise and, hence, the reliability of assay values. With the choice of global weights proposed here, the influence of such TALEs on the final optimization result is reduced, while such TALEs do not need to be completely removed from the training set.
As detailed in S1 Text, PBM experiments from [21] were filtered for apparent data quality, normalized log-intensities were used as target values, and global weights were defined uniformly for all putative target boxes from a common PBM experiment.
Xanthomonas oryzae pv. oryzae (Xoo) strains PXO83, PXO142 and ICMP 3125T were cultivated in PSA medium at 28°C.
Oryza sativa ssp. japonica cv. Nipponbare was grown under glasshouse conditions at 28°C (day) and 25°C (night) at 70% relative humidity (RH). Leaves of 4-week-old plants were infiltrated with a needleless syringe and a bacterial suspension with an OD600 of 0.5 in 10 mM MgCl2 as previously described [26].
Rice cultivar Nipponbare leaves were inoculated with Xoo strains PXO83, PXO142, ICMP 3125T, or MgCl2 as mock control in five spots in an area of approx. 5 cm using a needleless syringe. Two leaves of three rice plants each were inoculated for each strain and control, respectively. 24h later, samples were taken, frozen in liquid nitrogen, and RNA prepared. Three replicates of this experiment were done on separate days and subjected to RNAseq analysis, separately. Stranded libraries were sequenced on an Illumina HiSeq 2500 instrument (Eurofins Genomics) as 100 bp single-end reads
RNA-seq data 48h after inoculation with different Xoc strains (BLS256, BLS279, CFBP2286, B8-12, L8, RS105, BXOR1, CFBP7331, CFBP7341, CFBP7342), and mock controls [9] were downloaded from Gene Expression Omnibus available under accession number GSE67588.
RNA-seq data were adapter clipped using cutadapt (v1.15) [27] and quality trimmed using trimmomatic (v0.33) [28] with parameters “SLIDINGWINDOW:4:28 MINLEN:50”. Transcript abundances were computed by kallisto [29] using parameters “–single -b 10 -l 200 -s 40” and the cDNA sequences available from http://rice.plantbiology.msu.edu/pub/data/Eukaryotic_Projects/o_sativa/annotation_dbs/pseudomolecules/version_7.0/all.dir/all.cdna. Differentially expressed genes relative to the respective control samples were determined by the R-package sleuth [30].
For the Xoo strains and the respective mock control, replicates have been paired during library preparation and sequencing. Hence, the replicate was considered as an additional factor when computing p-values of differential expression for the Xoo samples but not for the Xoc samples. Differential expression was aggregated on the level of genes using the parameter target_mapping of the sleuth function sleuth_prep(), and b-value, p-value, and Benjamini–Hochberg-corrected q-value were recorded. The b-value reported by sleuth when applying a Wald test is actually a biased estimator of the log-fold change. However, as this is a more commonly understood term, we refer to the b-value as “log-fold change” in the remainder of this manuscript. Gene abundances, and sleuth outputs with regard to differential expression are provided as S1 and S2 Tables, respectively. RNA-seq reads were also mapped to the rice genome (MSU7) to obtain detailed information about transcript coverage. To this end, adapter clipped and quality trimmed reads were mapped using TopHat2 v2.1.0 [31], and the resulting BAM output files were processed in further analyes described below.
Let r = r1r2…rL denote the RVD sequence of length L of a TALE, where rℓ ∈ {AA, …, YY, A*, …, Y*} denotes a single RVD, and rℓ,12 and rℓ,13 denote the 12th and 13th AA of that RVD, respectively. Let x = x0x1…xL denote a putative target box of length L + 1 of that TALE, where xℓ ∈ {A, C, G, T} and x0 denotes the nucleotide bound by the zero-th, cryptic repeat.
The general idea of the model proposed here is to model the total binding score of a putative target box x given the RVD sequence r of a TALE as a sum of contributions of i) binding to the zero-th repeat, ii) binding to the first RVD, and iii) binding to the remaining RVDs, where the latter two terms may be weighted by an additional, position-dependent but sequence-independent term.
s ( x | r , θ ) = m 0 ( x 0 | r 1 , θ 0 ) + m 1 ( x 1 | r 1 , θ 1 , θ m ) · p ( 1 | θ p ) + ∑ ℓ = 2 L m ( x ℓ | r ℓ - 1 , r ℓ , θ m ) · p ( ℓ | θ p ) (1)
Here, θ = (θ0, θ1, θm, θp) denote the sets of real-valued parameters of the term for binding to the zero-th, first, and remaining repeats, and the position-dependent term, respectively.
The term m0(x0|r1, θ0) for binding to the zero-th repeat may depend on the first RVD on the TALE, since dependencies between zero-th and first repeat have been observed before [23]. However, our knowledge about such dependencies is limited to the data presently available and, hence, we limit the RVDs for which a dependency is considered to a set R 0. Our data regarding systematic, quantitative analyses of the base preference of the zero-th repeat is limited in general, although it is widely assumed that position 0 in target boxes of natural TALEs is preferentially T and less frequently C. We include this prior knowledge into a-priori parameters π x 0.
m 0 ( x 0 | r 1 , θ 0 ) = π x 0 + θ 0 , x 0 + δ ( r 1 ∈ R 0 ) · θ 0 , x 0 | r 1 (2)
In this paper, we set R 0 = { H D , N N , N G , N I , N S } and πT = log(0.6), πC = log(0.3), πA = πG = log(0.05).
The term m1(x1|r1, θ1, θm) for binding to the first repeat depends on the 13th AA r1,13 of the first RVD r1, but may be extended by additional terms that either model a general dependency on the complete first RVD (including the 12th AA), and/or a separate base preference for a given 13th AA at the first position. Again, this modularity allows us to adapt the model to the resolution of data available, since a substantial part of RVDs is only covered by the systematic but limited data reported in [24, 25].
m 1 ( x 1 | r 1 , θ 1 , θ m ) = θ m , x 1 | r 1 , 13 + δ ( r 1 ∈ R 1 ) · θ m , x 1 | r 1 + δ ( r 1 , 13 ∈ R 2 ) · θ 1 , x 1 | r 1 , 13 (3)
In this paper, we set R 1 = { H D , N N , N G , H G , N I , N K } and R 2 = { D , N , G , I }.
The term m(xℓ|rℓ−1, rℓ, θm) for binding to the remaining repeats again depends on the 13th AA rℓ,13 of the current RVD rℓ, but may be extened by additional terms that either model a dependency on the complete RVD (with parameters shared with the correponding term used for the first RVD), and/or the complete RVD rℓ at the current repeat and the 12th AA rℓ−1,12 at the previous repeat:
m ( x ℓ | r ℓ - 1 , r ℓ , θ m ) = θ m , x ℓ | r ℓ , 13 + δ ( r ℓ ∈ R 1 ) · θ m , x ℓ | r ℓ + δ ( r ℓ , r ℓ - 1 ∈ R 3 ) · θ m , x ℓ | r ℓ , r ℓ - 1 , 12 (4)
In this paper, we set R 3 = { H D , N N , N G , N I }.
Finally, we define the position-dependent term as a mixture of two logistic functions and a constant term, where the logistic functions depend on the relative distance of ℓ from the start and end of the putative target box, respectively:
p ( ℓ | θ p ) = e θ p , 1 ∑ j = 1 3 e θ p , j 1 1 + e - θ p , a , 1 ( ℓ L + θ p , b , 1 ) + e θ p , 2 ∑ j = 1 3 e θ p , j 1 1 + e - θ p , a , 2 ( L - ℓ L + θ p , b , 2 ) + e θ p , 3 ∑ j = 1 3 e θ p , j (5)
The parameters θp,a,1 and θp,a,2 denote the slopes, and θp,b,1 and θp,b,2 denote the location parameters of the logistic functions.
The implementation of this model is available from the Jstacs github repository (cf. section “Availability”) in package projects.tals.linear.
The training data D = ( t 1 ,…, t N ) comprise tuples ti = (ri, xi, vi, wi, gi) of TALE RVD sequence ri, target box xi, target value vi, global weight wi and group gi (cf. sections “Data” and “Model”). Given the current parameter values θ, we may further compute for each pair of TALE and target box, the corresponding model score si = s(x|ri, θi). The goal of the learning process is to adapt the parameter values θ such that the differences between computed scores si and target values vi becomes minimal. However, despite the normalization of target values described in section “Data”, target values from different experimental setups (represented by the groups gi) may live on different scales. Hence, we allow the learning process to linearly transform the computed scores si before comparing them to the target values. The total error between target value and prediction score is defined as
E ( θ ; D , β ) ≔ ∑ i = 1 N w i · ( f ( s ( x i | r i , θ ) | g i , β ) - v i ) 2 (6)
where
f ( s i | g i , β ) = exp ( β a , g i ) · s i + β b , g i , (7) β = (βa,1, βb,1, …, βa,G, βb,G), β a , g i and β b , g i are group-specific scale and shift parameters, respectively, and G is the total number of groups in the data set D.
In addition, we use an L2 regularization term on the model parameters θ to avoid overfitting and explosion of parameter values:
L 2 ( θ ) ≔ λ · | | θ | | 2 (8)
where the regularization parameter λ is set to 0.1 in this paper.
The number of model parameters for the different terms varies greatly, depending on the number of conditions (e.g., 12th AA of previous RVD, separate parameters for individual RVDs). This regularization also has the effect that more complex dependency parameters assume values considerably different from 0 only if the modeled specificity cannot be captured by the less complex sets of parameters.
The final objective function is then to minimize sum of the error term E ( θ ; D , β ) and the regularization term L2(θ) with respect to the parameter values:
( θ * , β * ) = argmax( θ , β )E ( θ ; D , β ) + L 2 ( θ ) (9)
This objective function is implemented in class MSDFunction in package projects.tals.linear. Parameter optimization is performed by a gradient-based quasi-Newton method as implemented in class de.jstacs.algorithms.optimization.Optimizer of the Jstacs library [20]. As the objective function is not convex, we start the optimization from 50 independent, random initializations and finally choose the set of locally optimized parameters that achieves the minimum value of the objective function.
The final parameters θ* of the trained model may then be used to determine prediction scores of previously unseen pairs of TALEs and putative target boxes, whereas the value of β* is discarded after optimization.
For predicting putative TALE target boxes for a given TALE with RVD sequence r of length L, we follow a sliding window approach scanning input sequences x1, …, xN. Input sequences could, for instance, be promoter sequences of annotated genes but also complete chromosomes. Each sub-sequence xi,ℓ, …, xi,ℓ+L then serves as input of the model to compute the corresponding score s(xi,ℓ, …, xi,ℓ+L|r, θ*). To allow for a rough comparison of scores, even between TALEs of different lengths, we normalize this score to the length of the input sequence, i.e., we compute a normalized score as s′(xi,ℓ, …, xi,ℓ+L|r, θ*) ≔ s(xi,ℓ, …, xi,ℓ+L|r, θ*)/(L + 1).
For scanning promoter sequences, we also provide an option for penalizing predictions of the reverse complementary strand, relative to the orientation of the downstream gene. Specifically, a small constant c is subtracted from all prediction scores s′ on the reverse complementary strand. Throughout this paper, we use c = 0.01.
The scanning process explicitly accounts for aberrant repeats, which may loop out of the repeat array [6]. To this end, we search for putative target boxes with all repeats present in the repeat array, but also all combinations of aberrant repeats removed from the RVD sequence. Due to the normalization of scores by the number of repeats, predictions based on these modified RVD sequences can still be ranked in a common list.
In addition, we provide a box-specific p-value as a statistical measure for the significance of target box predictions. Those p-values may either be computed from a dedicated background set of sequences or from a random sub-sample of the scanned input sequences. In either case, scores are computed for the sub-sequences given the current RVD sequences, then a Gaussian distribution is fitted to those score values, and the p-value for a given score is determined from that Gaussian distribution. While the Gaussian distribution does not perfectly fit the true distribution of score values, it allows for computing p-values with high resolution (as opposed to just using percentages of the scores themselves) and even for score values larger than any of the scores in the random sample. Using this procedure, the mapping from scores to p-values is monotonic, i.e., a larger prediction score results in a smaller p-value. Scanning promoters of a large number of genes for putative target boxes results in a multiple testing problem, and users may choose to apply a correction method of their choice controlling for family-wise error rate or false discovery rate. As a rough guideline under the assumption that promoters of tens of thousands genes are scanned for target boxes, p-values below 10−6 may be promising candidates for further inspection.
We use PrediTALE for genome-wide prediction in the genome of Oryza sativa Nipponbare (MSU7, http://rice.plantbiology.msu.edu/pub/data/Eukaryotic_Projects/o_sativa/annotation_dbs/pseudomolecules/version_7.0/all.dir/all.chrs.con). We make predictions for each TALE of 3 Xoo strains and 10 Xoc strains. In order to confirm that the predicted target boxes might indeed be bound by the respective TALE, we use the above-mentioned RNA-seq data to determine if there are differentially transcribed regions around a putative target box. For each of the top 100 predictions, we search ± 3000 bp around the predicted site for regions of at least 400 bp that are differentially expressed. Specifically, we count the number of mapped reads for each 400 bp window in replicates of treatment and control. Counts are then normalized relative to the total number of reads within each library, and replicates are averaged separately for treatment and control. Here, we consider a region as differentially expressed if the mean normalized number of reads after infection (treatment) is at least 2-fold larger than the mean normalized number of reads in the control experiment. If several, adjacent 400 bp regions meet this criterion, those are joined to a common, longer region.
This procedure is implemented in a tool called DerTALE. As input, DerTALE expects genomic positions, i.e., the position of predicted target boxes, and BAM files of mapped reads for replicates of treatment and control. Region width, thresholds and averaging methods may be adjusted by user parameters.
For each predicted target box, a profile output is generated if there is at least one differential expressed region with a minimum length of 400 bp that does not overlap the target box, or if it overlaps, the differential region starts or ends at most 50 bp upstream or downstream of the target box.
The obtained profiles may be visualized using an auxiliary R script. In addition to the profile data, this R script requires annotations data of already known transcripts in gff3 format. By this means, users may then investigate whether the predicted binding site may activate the transcription of a gene that has not been annotated yet. Here, we use the MSU7 annotation (http://rice.plantbiology.msu.edu/pub/data/Eukaryotic_Projects/o_sativa/annotation_dbs/pseudomolecules/version_7.0/all.dir/all.gff3).
For differentially expressed regions without annotated MSU7 transcript, we searched for similar sequences using blastx of NCBI BLAST+ version 2.7.1 ftp://ftp.ncbi.nlm.nih.gov/blast/executables/blast+/LATEST/ and choose the non-redundant protein sequence (nr) database. In cases, where we did not receive a convincing hit, we additionally compared sequences with blastn against the reference RNA sequences (refseq_rna) database.
For scanning large input sequences, e.g., complete genomes of host plant species, an acceptible runtime is essential. Since the parameters at each position of the proposed model depend on the RVD sequence of the TALE of interest but do not include dependencies between different nucleotides of a putative target box, we may convert the model given a fixed TALE RVD sequence into an position weight matrix (PWM) [32, 33]. This allows for a quick computation of prediction scores that may be formulated as the position-wise sum of values stored in the TALE-specific PWM model. We further speed-up the scanning process by pre-computing indexes of overlapping k-mers in the same manner as proposed for the TALENoffer application earlier [34].
We compare the performance of the approach presented in this paper to those of established tools for predicting TALE target sites, namely Target Finder [14], Talvez [16], and TALgetter [18], based on RNA-seq data after inoculation with different Xoo and Xoc strains described above.
To this end, we collect the promoter sequences of all transcripts based on the MSU7 assembly and gene models [35] available from http://rice.plantbiology.msu.edu/pub/data/Eukaryotic_Projects/o_sativa/annotation_dbs/pseudomolecules/version_7.0/all.dir/. We consider as promoter the sequence spanning from 300 bp upstream of the transcription start site to 200 bp downstream of the transcription start site or the start codon, whichever comes first, as proposed before [18]. We then run each of the tools using default parameters on the extracted promoter sequence providing the RVD sequences of the TALEs present in the respective Xanthomonas strain (cf. S1 Data). Predictions in promoters of different transcripts belonging to the same gene are merged by considering only the prediction yielding the best prediction score.
Assessment of prediction performance based on in-planta inoculation experiments with Xanthomonas strains harboring multiple TALEs has the inherent complications that i) putative target genes cannot be attributed to one specific TALE based on the RNA-seq data alone and ii) genes showing increased expression after inoculation may either be regulated directly by a TALE binding to their promoter or indirectly via other, regulatory target genes. Hence, we define true positives as those genes that have a predicted target box in their promoter and are also up-regulated after inoculation with the respective Xanthomonas strain relative to control as derived from RNA-seq data. By contrast, we cannot clearly define false negatives, since genes that are up-regulated after inoculation but do not contain a predicted target box in their promoter could be indirect target genes. False positives, in turn, would be genes with a predicted target box in their promoter that are not up-regulated after Xanthomonas inoculation.
A further issue hampering performance assessment by standard methods like receiver operating characteristic (ROC) [36] or precision-recall (PR) curves [37, 38] is that for two of the tools considered (Target Finder and Talvez), none of the reported prediction scores is comparable between different TALEs, especially TALEs of different lengths. Hence, we decide to use varying cutoffs on the number of predicted target genes per TALE to establish a common ground for comparing all four approaches.
Following these considerations, we collect for each of the four approaches the number of true positive predictions (TPs) for cutoffs on the number of predictions per TALE from 1 (i.e., the top prediction) to 50. We then plot for each approach the number of true positives against this cutoff to obtain a continuous picture of its prediction performance. In addition, we collect for the same cutoffs the number of TALEs with at least one predicted target gene among the true positives.
The area under these curves may serve as a further measure of general prediction performance in analogy to, for instance, the area under the ROC curve.
Finally, we compare the TPs at distinct cutoffs (1, 10, 20, 50) between the four tools. For a specific cutoff, we collect the TPs (or, in analogy, number of TALEs with at least one predicted target) for each of the four tools. Statistical significance of the differences in observed TPs is then assessed by a Quade test [39] using the quade.test function in R [40] and pairwise comparisons are performed by the post-hoc test implemented in function quadeAllPairsTest of the PMCMRplus R-package [41].
In addition, we obtain promoter sequences of five plant species to test PrediTALE for pathosystems beyond Xanthomonas oryzae—rice. To this end, we download genome sequences and gene annotations from phytozome (https://phytozome.jgi.doe.gov) for cassava (Manihot esculenta, v7.0, [42]), sweet orange (Citrus sinensis, v1.1, [43]), cotton (Gossypium raimondii, v2.1, [44]), and from solgenomics (https://solgenomics.net) for tomato (Solanum lycopersicum, ITAG3.20, [45]) and pepper (Capsicum annuum CM334, v1.55, [46]). For these plant species, we consider as promoter the sequence from 300 bp upstream of the annotated transcription start site to the start codon to be less dependent on the exact annotation of transcription start sites.
PrediTALE is available as a web-application based on Galaxy at http://galaxy.informatik.uni-halle.de. Both PrediTALE and DerTALE are available as command line application from http://jstacs.de/index.php/PrediTALE and have also been integrated in AnnoTALE 1.4. Source code is available from https://github.com/Jstacs/Jstacs in packages projects.tals.linear, projects.tals.prediction, projects.tals.training, and projects.tals.rnaseq, where also provide an XML representation of the trained model at projects.tals.prediction.preditale_quantitative_PBM.xml. The parameters of the PrediTALE model will be adapted as additional training data become available in the future, while we will preserve a history of PrediTALE models to assure reproducibility. PrediTALE and DerTALE will also be maintained as part of the AnnoTALE suite.
In this section, we benchmark the predictions of PrediTALE against those made by one of the previous approaches, namely Target Finder [14], Talvez [16], and TALgetter [18].
To this end, we consider different Xanthomonas oryzae pv. oryzae (Xoo) and Xanthomonas oryzae pv. oryzicola (Xoc) strains for which we have an experimental support of up-regulated genes in Oryza sativa after infection based on RNA-seq data. Specifically, we consider the Xoo strains ICMP 3125T, PXO142 and PXO83 with in-house RNA-seq data available, and the Xoc strains B8-12, BLS256, BLS279, BXOR1, CFBP2286, CFBP7331, CFBP7341, CFBP7342, L8 and RS105 based on public RNA-seq data [9]. For the TALEs from the repertoires of these three Xoo and ten Xoc strains, we determine target gene predictions for each of the previous approaches and for PrediTALE. Predicted target genes are ranked by the corresponding prediction scores of the different approaches per TALE.
First, we study the overlaps between the sets of predicted target genes per approach to investigate how strongly predictions are affected by conceptual differences of these approaches. In Fig 1A, we show Venn diagrams of predicted target genes for the three Xoo strains based on the top 20 predictions per TALE, while the corresponding diagrams for the ten Xoc strains are available as S1 Fig. In general, we observe a substantial number of unique predictions for each of the four approaches, but especially for Talvez and PrediTALE. By contrast, the overlapping predictions between all four approaches amount to less than a quarter of the total predictions per approach. This demonstrates that prediction results strongly depend on the employed approach. However, prediction accuracy cannot be assessed without an experimental knowledge about genes that are up-regulated in planta upon Xanthomonas infection.
RNA-seq data for the three Xoo strains including previously unpublished data for PXO83, have been collected 24 hours after infection. Collection at this early time point has the advantage that the number of secondary targets, i.e., genes that are up-regulated as a secondary effect of direct TALE targets with regulatory function, should still be low. However, as the infection might not be fully established, yet, the variation between replicates and, hence, the number of significantly differentially expressed genes based on standard FDR-based criteria is rather low (cf. Table A in S2 Text). As we aim at sensitivity for the benchmark study, i.e., we want to avoid predictions to be erroneously counted as false positives, we consider genes as differentially up-regulated if they obtain an uncorrected p-value below 0.05 and are at least 2-fold up-regulated in this case, which results in 43 (PXO142) to 107 (ICMP 3125T) differentially up-regulated genes.
In case of the ten Xoc strains, RNA-seq data have been recorded 48 hours after infection. Here, infection should be fully established, but we expect a substantial number of secondary targets to be up-regulated already. Hence, we resort to rather standard thresholds with a FDR-corrected q − value < 0.01 and log fold change greater than 2 in this case. Notably, this still results in a larger number of differentially up-regulated genes (cf. Table B in S2 Text) than for the Xoo strains with numbers between 202 (CFBP2286) and 672 (L8).
Given these up-regulated genes as a ground truth, we may now count predictions of TALE target boxes in promoters of up-regulated genes as true positives, and predictions without observed up-regulation as false positives. In Fig 1B, we plot Venn diagrams of the true positives among the top 20 predictions of all four approaches. Notably, we find that the intersection of the predictions of all four approaches constitutes (one of) the largest set(s) in each of the three Venn diagrams. Among the predictions that are unique to one of the four approaches, we consistently find the largest number of true positive predictions for PrediTALE, which indicates the utility of our novel approach. Turning to the ten Xoc strains (S2 Fig), we again find the same tendency with regard to the predictions overlapping among all four approaches. However, the number of true positives among the unique predictions shows a less clear picture with a slight advantage towards Talvez, while predictions of PrediTALE often overlap with TALgetter and/or Target Finder. Together, the Venn diagrams for the Xoo and Xoc strains also illustrate why it is generally beneficial to complement in silico TALE target predictions with experimental data about gene regulation.
The results presented so far strongly depend on the thresholds of the ranks of the target predictions but also on the thresholds applied to the RNA-seq data. To address the former problem, we aim at an assessment of target predictions over all rank thresholds, while we will handle the latter by separate evaluations applying different criteria to the RNA-seq data.
As detailed in section “Evaluation of prediction results”, standard performance measures like the area under the ROC curve [36] or the area under the precision-recall curve [37, 38] are inappropriate under this setting. Briefly, we cannot attribute an up-regulated gene to a specific TALE from the TALE repertoire of the strain under study. In addition, genes that are up-regulated in the RNA-seq experiment might also be due to secondary effects of TALE targets, due to general plant response to the bacteria, or due to other classes of effector proteins. Thus, we may not consider up-regulated genes without a matching prediction of a TALE target box in their promoter as false negatives. Hence, we decide to compare the performance of different approaches by means of the number of true positive predictions at different rank cutoffs, i.e., considering the top N predicted target genes of each approach.
In Fig 2, we plot the number of true positives for the three Xoo strains and each of the four approaches against the total number of predictions per TALE, considering only the highest-ranking prediction up to 50 target predictions per TALE, which we consider a reasonable cutoff under the scenario of manual inspection. In addition, we compute the area under this curve as an overall performance statistic across all rank cutoffs. For all three Xoo strains, we find that PrediTALE dominates the other three tools for rank cutoffs of 5 and above. For lower rank cutoffs, the ranking of tools is less clear, but PrediTALE still yields—for instance—the largest number of true positive predictions on rank 1 for two of the three strains. In the ranking with regard to the area under the curve (AUC), we find that PrediTALE again yields the best overall performance among all four approaches.
We take a different perspective on prediction results by assessing prediction performance on the level of TALEs. Specifically, we count the number of TALEs with at least one true positive target prediction for the same rank cutoffs as before. Again, PrediTALE identifies targets for a larger number of TALEs than the other approaches for the majority of rank cutoffs (Fig 3). However, we see notable differences between the different Xoo strains, where PrediTALE is able to identify putative targets for 10 of the 17 TALEs of ICMP 3125T, but only for 7 out of 19 TALEs for PXO142 and for 7 out of 18 TALEs for PXO83. As ICMP 3125T has also been the strain with the largest number of differentially up-regulated genes (cf. Table A in S2 Text), the lower number of TALEs in PXO142 and PXO83 with a predicted target might be due to a different progression of the Xanthomonas infection.
We further summarize the data behind Figs 4 and 3 in Tables C and D in S2 Text, where we also report the average ranks of the four approaches across all three Xoo strains.
For sake of completeness, we also evaluate the four approaches for differentially up-regulated genes after Xoo infection based on the same FDR-based thresholds as for the Xoc experiments (S3 and S4 Figs).
Although it has been shown that TALEs may activate transcription in both strand orientations relative to the transcription start site (TSS) of target genes [12, 13], a preference for the forward orientation has been postulated [13]. This is reflected by the strand penalty of PrediTALE, but no similar parameter exists for the previous approaches. Hence, above comparison might be perceived as partially unfair in favor of PrediTALE. For this reason, we repeat the benchmarking after restricting the predictions of all four approaches to a forward orientation relative to the TSS (S5 and S6 Figs). While the restriction to the forward strand has an effect on the number of target genes and TALEs with at least one true positive target, PrediTALE still yields an improved performance compared with the previous approaches over a wide range of rank cutoffs and, hence, achieves the largest AUC value of the four approaches in all cases.
For the ten Xoc strains, we find an improved prediction performance for PrediTALE as well. On the level of true positive target genes (Fig 4), PrediTALE yields the largest number of true positives for a rank cutoff of 1 for seven of the ten Xoc strains (cf. Table I in S2 Text). We also find an improved performance for the majority of the remaining rank cutoffs and Xoc strains. This improvement is especially pronounced for strains Xoc BLS279, CFBP7331, CFBP7341, and L8, whereas PrediTALE performs similar to or slightly worse than at least one of the previous approaches for Xoc CFBP7342 and RS105. For the remaining strains (B8-12, BLS256, BXOR1, CFBP2286), the improvement by PrediTALE is either rather small or mostly restricted to rank cutoffs of 20 or larger. This is also reflected by the areas under the curves, where PrediTALE yields the largest areas for B8-12, BLS256, BLS279, BXOR1, CFBP2286, CFBP7331, CFBP7341, L8, and also RS105, but nor for CFBP7342. Results are largely similar on the level of TALEs with at least one true positive predicted target (S7 Fig), where PrediTALE yields the largest area under the curve for the same strains.
To obtain a more condensed overview on the results for the Xoc strains, we finally compute the average performance ranks across all ten Xoc strains for each of the four approaches and fixed rank cutoffs of 1, 10, 20, and 50, and for the area under the curve both on the level of target genes and on the level of TALEs (Table 1 and Table I and J in S2 Text). For all rank cutoffs and the area under the curve, we observe that PrediTALE yields the best average rank with values betwen 1.1 and 1.5. We further assess the statistical significance of differences between the different tools by a Quade test, and the pairwise differences between tools by the associated post-hoc test (see Methods). This assessment is partly limited by the fact that pairs of Xoc strains may have identical TALEs in their TALEomes, which also means that the performance values of those strains are not truly independent. However, we did not find a clear relationship between the similarity of performance values obtained for the different strains and the similarity of the corresponding TALEomes. For this reason, we consider this dependency rather mild and favor this limited statistical assessment over the complete lack of it.
Consistent with the previous observations, we find that PrediTALE never performs significantly worse then any of the three previous approaches, whereas in many cases it performs significantly better, often with p-values below 0.001 in the post-hoc test. Notable exceptions are a rank cutoff of 1, where PrediTALE does not perform significantly different from Target Finder, a rank cutoff of 10, where PrediTALE does not perform significantly different from Talvez, and on the level of TALEs, a rank cutoff of 20, where PrediTALE does not perform significantly different from TALgetter.
Repeating the same analysis for varied q-value threshold (S8 and S9 Figs, Table K, L, and M in S2 Text), for varied log fold change threshold (S10 and S11 Figs, Table N, O, and P in S2 Text), and for predictions restricted to the forward strand relative to the TSS (S12 and S13 Figs, Table Q, R, and S in S2 Text), benchmarking results are essentially similar to our previous findings. One notable exception is the Quade test for rank 1 predictions restricted to the forward strand (Table S in S2 Text), which is no longer significant. This means that none of the approaches studied yields significantly better rank 1 predictions than any other under this scenario.
Although the focus of this manuscript is on target predictions for TALEs from X. oryzae strains, PrediTALE may as well be applied to TALEs from other Xanthomonas species. To illustrate this, we perform promoterome-wide scans for putative target boxes of TALEs from five additional Xanthomonas species and corresponding host plants for which virulence targets have been published previously. We find the known targets of these five TALEs on rank 1 or 2 of the corresponding PrediTALE predictions (Table 2 and S6 Table). Interestingly, the top prediction of PrediTALE for AvrBs3 in pepper is a different target (transcription factor bHLH137, CA06g21040) than the well described target (transcription factor UPA20, CA03g22700) [47].
Summarizing the benchmark studies, we find i) that PrediTALE produces several unique predictions that might not have been considered based on previous approaches, ii) although low in absolute terms, the number of true positives among these predictions is often larger than for the previous aproaches, and iii) an assessment of the performance of PrediTALE across a wide range of rank cutoffs demonstrates that in most of the cases the application of PrediTALE yields a larger number of true positive target predictions than any of the three previous approaches. However, we also observe true positive predictions of one of the previous approaches that would be missed by PrediTALE. A general recommendation would be to use the union of the predictions of all four tools when aiming for sensitivity, i.e., to recognize as many true positives as possible. Aiming at precision instead, i.e., maximizing the fraction of true positives in the predictions considered, our results indicate that using either only PrediTALE predictions or predictions in the intersection of all four approaches would be recommended.
Having established that PrediTALE often yields an improved performance compared with previous approaches, we investigate in the following, which aspects of the PrediTALE model contribute to which extent to the performance of the full PrediTALE model. To this end, we first consider a baseline model for which we define the sets R 0, R 1, R 2 and R 3 as empty sets, and set the position-dependent term p(ℓ|θp) to a uniform distribution. Starting from this baseline model, we then individually restore each individual set and the position-dependent term to its original value, and record the difference in the observed performance. Reciprocally, we consider the full model and determine the difference of its performance to a model where only one of the individual sets is defined as empty or the position-dependent term is set to uniform. In Fig 5, we present the results of this analysis, again considering the number of TALEs with at least one true positive prediction based on the top 20 predictions per TALE, while the respective results with regard to the total number of true positive target genes are shown in S14 Fig. As a reference, we also include the difference in performance of the full model compared with the baseline model.
We find that the results of the two perspectives (adding a feature to the baseline model vs. completing the full models) are contradictory. While some of the features even reduce performance when added to the baseline model (separate specificities for position 1, R 2; specificities for individual RVDs, R 1), all features increase performance either on the level of TALEs or target genes when completing the full model. The specificities at position 0 dependending on the first RVD (R 0) are a notable exception. Here we observe an improvement of performance in either case, which is also substantially greater than for any of the other features.
However, this effect may not only be attributed to the specificities at position 0 being modeled depending on the first RVD. Inspecting the specificity parameters of the full model (Fig 5C) and comparing these to those of the baseline model, the baseline model with R 0 restored, and the full model with R 0 set to the empty set (S15 Fig), we find complex interactions among the specificity parameters. As this study has been conducted with a large number of independent restarts of the procedure optimizing model parameters, this is unlikely an effect of the optimization getting stuck in local optima. Rather the objective function (difference between observed quantities and prediction scores) appears to skew some of the remaining model parameters to achieve its optimum if the model is lacking the conditional specificities at position 0. Nonetheless, these results indicate that the inclusion of specificities at position 0 depending on the first RVD is an essential ingredient of the PrediTALE model. Currently, this aspect is limited by the corresponding training data from [23] and, hence, it might be a worthwhile perspective to quantitatively investigate this dependency for further RVDs in the future.
In addition, the specificity parameters of the PrediTALE model may also contain interesting patterns per se. For instance, we find that base preference at position 0 given RVD “NS” at position 1 is less clear than for other RVDs, where the known target box of TalC (TalBS1) harboring “NS” at position 1 is preceeded by a ‘C’ in the promoter of Os11N3 [52]. The UPA target box of AvrBs3 in pepper has an ‘A’ at position 1 [4], although the first RVD of AvrBs3 is “HD”, which complies with the specificity of ‘D’ at position 1 being shifted towards base ‘A’ relative to the general preference of RVDs with AA 13 equal to ‘D’ being ‘C’.
Finally, we consider the position-dependent term of the full model (Fig 5B), and find that it is much simpler than allowed by the mixture of two logistic functions, corresponding to a straight, slightly decreasing line. In contrast to the specificity parameters, the position-dependent term seems to be largely independent of the specificity features (cf. S15 Fig).
As all features contribute, at least slightly, to the performance of the full PrediTALE model, we consider this model in the remainder of this manuscript.
As we have seen from Fig 1B, putative target genes with up-regulation after Xoo infection are often found in the intersection of the predictions of all four approaches. In addition, PrediTALE predicts several putative target genes of TALEs from the three Xoo strains that might have been neglected using one of the previous tools. In the following, we scrutinize the predictions for the Xoo strains with a focus on novel predictions, while we give a complete list of top 20 predictions of all four approaches including the ten Xoc strains in S3 Table.
In Table 3, we collect further information about those target genes including the corresponding log fold change and prediction ranks for all four approaches.
The target genes in the intersections of the predictions of all four approaches comprise several well known targets. For instance, Os09g29820 (OsTFX1), a bZIP transcription factor, is targeted by TALEs from class TalAR with members in all three Xoo strains (S16 Fig) and has been proposed as a TALE target early [5, 53].
Os01g73890 (TFIIAγ) [5], that has been shown to promote TALE function [54], is targeted by TalBM2 in ICMP 3125T. In concordance to TalBM class members missing in PXO142 and PXO83, Os01g73890 shows no up-regulation in these two strains. Os07g06970 (HEN1) has also been among the first TALE target genes proposed [5] and is targeted by TalAP members present in all three Xoo strains, but falls below the threshold on the log fold change by a small margin in ICMP 3125T (S17 Fig). Os01g40290 [5], an expressed protein without annotated function, Os06g29790 [18], a phosphate transporter, and Os11g26790 [16] (RAB21), a dehydrin that has been shown to play a role in drought tolerance related to pathogen infection [55], have also been predicted in previous studies.
In addition, we find putative target genes in the intersection that have not been reported before: Os02g06670, a retrotransposon protein, is predicted as a target of TalBA8 and TalBA2 in ICMP 3125T and PXO83, respectively, whereas PXO142 lacks a TalBA member. Nonetheless, Os02g06670 is up-regulated after PXO142 infection, although to a lesser degree than in the other two strains (cf. S17 Fig). Os02g49350, a plastocyanin-like protein, is strongly up-regulated only in PXO142 and predicted as a target of TalBH2, where class TalBH is exclusive to PXO142 among the strains studied.
Finally, we find several putative target genes that have been predicted only by a subset of approaches: For ICMP 3125T, Os04g43730 [56] (OsWAK51) is among the top 20 predictions for TalES1 only for Target Finder and Talvez. In turn, PrediTALE predicts Os06g03710 (DELLA protein SLR1) as a TalES1 target on rank 19, which appears on later ranks for the other approaches. Os04g43730 is induced more strongly than Os06g03710 and exclusively in ICMP 3125T, which renders this the more likely target. Os03g51760 [16] (OsFBX109) is among the top 20 predictions for TalAD members only for PrediTALE. Due to variations in their RVD sequence, TALgetter has this in the top 20 predictions only for TalAD22 in ICMP 3125T, but not for the other strains. As Os03g51760 is clearly up-regulated after infection with any of the three Xoo strains (S17 Fig), this is likely a true TalAD target.
Talvez and TALgetter have Os03g03034, annotated as a flavonol synthase, among their top 20 predictions for TalAQ members in ICMP 3125T and PXO83, while this gene is among the top 20 predictions of PrediTALE only for TalAQ3 in PXO83 due to differences in RVD sequence. In PXO142, TalAQ15 is annotated as a pseudo gene and this pattern is also reflected by the RNA-seq data. Os03g03034 has been proposed to be a TALE target before [5, 56].
Os04g05050 [16, 56], annotated as a pectate lyase, is only among the top 20 predictions of PrediTALE in ICMP 3125T (TalAB16) and PXO83 (TalAB5), whereas this gene is ranked substantially lower (rank 83) for TalAB8 from PXO142 by PrediTALE as well. From the RNA-seq data, we find that Os04g05050 is up-regulated in all three Xoo strains, although the level of up-regulation is lower for PXO142 than for the other two strains.
Os05g45070, annotated as hairpin-induced protein 1, is predicted only by PrediTALE as an alternative target of TalAO15 in ICMP 3125T and shows clear up-regulation only after infection with this Xoo strain. Os10g28240 [16], a calcium transporting ATPase, is predicted by TALgetter and PrediTALE as target of TalAR13 of ICMP 3125T but, on later ranks, also by the other two approaches, and is up-regulated exclusively after ICMP 3125T infection. Os09g07460 [16], a kelch repeat protein, is only among the top 20 predictions of Talvez for TalBA and on later ranks for the other approaches. This gene is up-regulated only in ICMP 3125T, although not strongly.
For PXO142, we find two further putative targets of TalBH2 that are predicted exclusively by PrediTALE: Os03g09150 (pumilio-family RNA binding) is up-regulated in PXO142 but also in PXO83, for which it does not appear among the top 20 predictions of any approach. Os03g09150 has been predicted before as a target of class TalAC [16]. However, PXO142 is lacking members of class TalAC, while Os03g09150 only appears at later ranks for TalAC5 of Xoo PXO83. Os11g31190 (Os11N3, OsSWEET14) is a well known target [52, 57], which is predicted here also for TalBH exclusively by PrediTALE due to its ability to adequately handle the aberrant repeat [6] of TalBH2. Os11g31190 is also known to be targeted by TalAC members (previously termed AvrXa7) [53] including TalAC5 in PXO83 and, hence, is strongly up-regulated after PXO83 infection as well. However, in this case all approaches fail to predict this target due to the large number of mis-matches in the target box [6], even accounting for the aberrant repeat in TalAC5.
Instead, another retrotransposon protein (Os04g19960 [58]) is the top prediction of PrediTALE for TalAC5 from PXO83, which is confirmed by RNA-seq data as this gene is strongly up-regulated after PXO83 infection but not after infection with one of the other strains.
In summary, we find several novel putative target genes of which three are highly promising (Os02g49350, Os05g45070, Os03g09150), where two of these (Os05g45070, Os03g09150) are predicted as targets of the respective TALE classes on high ranks exclusively by PrediTALE. Recently, we could experimentally validate the targets Os04g43730 (OsWAK51), Os06g29790 (phosphate transporter), Os03g51760 (OsFBX109), Os03g03034 (flavonol synthase), and Os04g05050 (pectate lyase) by qRT-PCR using a TALE-less strain (Roth X1-8) complemented with individual TALEs [56].
We also observe from Fig 3 and S7 Fig that for many strains, neither of the approaches considered is able to identify a putative target genes for all TALEs present in their TALEome. We term such TALEs without reasonable target prediction orphan TALEs, and we will discuss these in more detail in the following.
More precisely, we call a TALE or a TALE class orphan if there is no up-regulated gene among the top 50 predictions of any of the four approaches. Furthermore, we check if this pattern is consistent for the TALEs from a common TALE class across almost all Xoo and Xoc strains studied.
We find as orphan the TALE classes present in all three Xoo strains TalAF, TalAI and TalAN. In addition, TalAG (PXO142, PXO83), TalAL (PXO142), TalAS (PXO142, PXO83), TalBJ (PXO83), TalCA (PXO83), TalET (ICMP 3125T), and TalDR (PXO142) are orphan TALE classes in individual Xoo strains. The TALEs from class TalAI and TalDR are truncTALEs that are lacking large parts of the C-terminus including the activation domain and, for this reason, do not act as transcriptional activators. TruncTALEs have been found to function as suppressors of resistance mediated by an immune receptor [59].
In the Xoc strains, however, TalAF is not orphan as we find putative target genes among the top 50 predictions for the class members present in B8-12 and L8. For TalAZ, we find a target for TalAZ7 from Xoc L8, but not for the other 7 Xoc strains harboring TalAZ TALEs. In addition, we consider TalCQ1 from BXOR1 and TalCR1 (CFBP7331) and TalCR2 (CFBP7341) as orphan.
Reasons for orphan TALEs could be manifold. First of all, we cannot be sure that these TALEs are indeed expressed by the bacteria and are secreted into the host plant cells. Second, some TALEs might activate target genes slower or to a lesser degree than others and, for this reason, target gene activation might not be detectable, yet, in the RNA-seq experiments, especially at the 24h timepoint chosen for Xoo. Third, these TALEs might target specific variants of boxes in promoters of rice lines that are not represented by the O. sativa Nipponbare reference genome, or might even target genes in alternative host plants, e.g., grasses in the vicinity of fields where rice is grown. Fourth, these TALEs might target genes that are missing from the current gene annotations of rice. Such targets would be neglected by the current approach to specifically scan promoter sequences of annotated genes for putative TALE boxes. To address the latter issue, we switch to an alternative approach in the following. Here, we perform genome-wide scans for putative target boxes instead, and search for differentially expressed regions in the vicinity of putative target boxes predicted anywhere in the reference genome.
We perform genome-wide predictions of TALE target boxes in Oryza sativa Nipponbare (MSU7) for the 256 Xoc TALEs from 10 strains and 54 Xoo TALEs from 3 strains and check for differentially expressed regions near the predicted target boxes. Differential expression is based on the mapped RNA-seq data after infection with the respective Xoo and Xoc strains. Performing genome-wide scans is facilitated by the runtime optimization of the PrediTALE scanning process described in section “Implementation & scanning speed-up”, and we provide a comparison of exemplary running times of genome-wide scans for target boxes of all 28 TALEs of strain Xoc BLS256 in Table T in S2 Text.
After infection with Xoo strains, 14 TALEs are found to have differentially expressed regions near at least one predicted target box. Table 4 lists the total number of 19 TALE target boxes together with MSU7 gene annotations overlapping the differentially expressed regions. Notably, 15 of these targets have already been reported in subsection “PrediTALE predicts novel putative target genes” when restricting the search to promoter regions of annotated genes. However, for two genes, target boxes from other TALs were predicted in case of genome-wide scan. The expression of the pectate lyase precursor (Os04g05050) was up-regulated by TalAB5 according to promotor prediction, but the genome-wide prediction contains the same gene up-regulated by TalAD22. The same scenario for the phosphate transporter 1 (Os06g29790), which according to promotor predictions is up-regulated by TalAO16 and TalAP15. However, in the genome-wide scans, a target box of TalAH11 was predicted. The genome-wide scan i) does not make use of gene annotations, and ii) could be expected to be more prone to false positive predictions than the restricted search in promoters. Hence, the fact that many predictions re-occur in the genome-wide scan demonstrates the general utility of this approach.
In addition to those targets reported previously, we find three novel target boxes in the vicinity of differentially expressed regions that overlap annotated genes, including a wound-induced protein and an oxidoreductase. For TalAO16 from PXO142, we find a differentially expressed region next to a predicted target box on chromosome 7 with no annotation in MSU7 (S18 Fig; complete list in S4 Table). For this reason, we extracted the sequence under the differentially expressed region, and first compared it against the NCBI protein database ‘nr’ using blastx but received no matching result. We additionally compared this sequence against the NCBI reference RNA sequences (refseq_rna) using blastn, which resulted in a highly significant hit for XR_001547425.2, a predicted long non-coding RNA.
Upon infection of rice with Xoc strains, differentially expressed regions near at least one predicted target box were found for 26 of 28 (B8-12), 28 of 28 (BLS256), 25 of 26 (BLS279), 26 of 27 (BXOR1), 22 of 28 (CFBP2286), 19 of 22 (CFBP7331), 19 of 21 (CFBP7341), 18 of 23 (CFBP7342), 27 of 29 (L8) and 19 of 24 (RS105) TALEs. S5 Table lists all genome-wide predicted targets in the vicinity of differentially expressed regions of these Xoc strains.
In the following, we will discuss two example regions in detail. As discussed in the previous section, TalAZ appears to be an orphan TALE based on the promoterome-wide scans for target boxes. However, based on genome-wide scans, we find a differentially expressed region, which could constitute a target gene of TalAZ, on Chr4 (Fig 6). Only 8 of the 10 Xoc strains studied have a TalAZ member in their TALEome. The profile plots clearly show that the region of interest is only differentially expressed after infection with these 8 strains harbouring TalAZ members. Performing blast searches of the differentially expressed sequences, we received a hit for XP_015634381.1, a sulfated surface glycoprotein 185 [Oryza sativa Japonica Group], which has been added to the IRGSP-1.0 annotation at NCBI but was not present in MSU7.
As a second example, we consider a putative TalBD target on Chr6. The profile plots (Fig 7) show differentially expressed regions in all 10 strains. However, a blastx search of the respective sequences, spanning two larger differentially expression regions, provides no clear result. Matches include an Auxin-responsive protein IAA22 (Q69TU6.1) and different bromodomain-containing factors (XP_006659043.1, XP_025882131.1 XP_015650662.1). As drops in the coverage profiles and split reads in the mapping indicate the existence of introns within the differentially expressed regions, we additionally compare the spliced sequence using blastn against the NCBI reference RNA sequences. The result contains a predicted non-conding RNA (XR_003242961.1) and different transcript variants of a predicted mRNA, coding for bromodomain-containing factors (XM_015840709.1, XM_015840708.1, XM_006658980.2, XM_026026346.1, XM_015795177.2, XM_015795176.2).
In summary, our results demonstrate that genome-wide prediction of target boxes using PrediTALE enables us to identify novel targets independently of existing gene annotations including previously missing non-coding RNAs.
Accurate computational predictions of TALE target boxes are required for elucidating virulence targets of TALEs that support bacterial infection of host plants. In this paper, we present PrediTALE, a novel approach for predicting target boxes based on a TALE’s RVD sequence. Since the publication of all previous approaches [14, 16, 18], our understanding of mechanisms and principles of TALE targeting has increased substantially. Specifically, it has been shown that repeats of aberrant lengths may compensate for frame shifts in target boxes [6], that activation of gene expression by TALEs binding to the reverse strand is possible, but rare [13]. In addition, quantitative data about virtually all combinations of AAs at RVD positions have been collected [19, 21–25]. All these insights have been integrated into PrediTALE either as part of the model or as training data that are used to adapt model parameters. Here, we demonstrate that PrediTALE predicts TALE targets with improved accuracy compared with previous approaches, where ground truth is derived from in-house and public RNA-seq data after Xoo and Xoc infection. However, our results also confirm that any of the current computational approaches suffers from false positive predictions and, hence, experimental support of predicted targets is essential.
PrediTALE predicts several unique target genes, several of which are highly promising for further experimental validation. While RNA-seq data supports that these are activated by TALEs in planta, their importance for the infection process still needs to be investigated.
Previously, predictions have been mostly limited to putative promoter regions of annotated genes. Here, we consider genome-wide predictions instead, which are feasible due to the acceptable runtime of PrediTALE, the improved accuracy of target box predictions, and the filtering steps based on RNA-seq data as implemented in DerTALE. We demonstrate that targets reported from promoterome-wide predictions are also recovered in genome-wide scans, but we also find differentially expressed regions at loci that do not overlap with annotated genes. These could be either protein-coding genes that are missing from the current annotation, but also include putative non-coding RNAs, which might have regulatory activity or other functions that foster bacterial infection.
To promote future research in plant-pathogen interactions related to TALEs, we make our methods available to the scientific community as open-source software tools.
|
10.1371/journal.pntd.0002241 | Association of Eumycetoma and Schistosomiasis | Eumycetoma is a morbid chronic granulomatous subcutaneous fungal disease. Despite high environmental exposure to this fungus in certain regions of the world, only few develop eumycetoma for yet unknown reasons. Animal studies suggest that co-infections skewing the immune system to a Th2-type response enhance eumycetoma susceptibility. Since chronic schistosomiasis results in a strong Th2-type response and since endemic areas for eumycetoma and schistosomiasis do regionally overlap, we performed a serological case-control study to identify an association between eumycetoma and schistosomiasis. Compared to endemic controls, eumycetoma patients were significantly more often sero-positive for schistosomiasis (p = 0.03; odds ratio 3.2, 95% CI 1.18–8.46), but not for toxoplasmosis, an infection inducing a Th1-type response (p = 0.6; odds ratio 1.5, 95% CI 0.58–3.83). Here, we show that schistosomiasis is correlated to susceptibility for a fungal disease for the first time.
| Eumycetoma is a mutilating fungal disease of mainly the foot and is found in (sub)tropical regions such as Sudan. At the moment it is not understood why some people develop eumycetoma and others not. In the regions were eumycetoma is prevalent many other infections are also found. These infections could alter the immune system which makes people more or less susceptible in obtaining another infection. One of the infections with such an effect is Schistosomiasis. In Africa, eumycetoma is found in regions were schistosomiasis is prevalent. In this study we show that eumycetoma patients more often have antibodies against Schistosoma species, than healthy controls from the same region. In contrast, eumycetoma patients did not have more often antibodies against Toxoplasma species. This might implicate that schistosomiasis predisposes eumycetoma development. If schistosomiasis indeed predisposes eumycetoma development, eradicating Schistosoma in a population could also lower the number of eumycetoma cases in that area, which in the end could lead to intervention strategies not only for schistosomiasis but also for eumycetoma.
| Eumycetoma is a chronic granulomatous subcutaneous infectious disease endemic in many tropical and sub-tropical regions in the so-called mycetoma belt between 30°N and 15°S of the equator [1]. Sudan is a country with the highest country-wide prevalence of eumycetoma (Figure 1). In a recent survey conducted by the Mycetoma Research Centre, it appeared that in the endemic villages in the Gezira area of Sudan 2% of the population has eumycetoma (Prof. A. Fahal, personal communication) [2], [3]. Although mycetoma can be caused by a variety of bacterial and fungal micro-organisms, most mycetoma cases in Sudan (ca. 70%) are caused by the fungus Madurella mycetomatis (eumycetoma) [4]. Based on antibody measurements in earlier studies it was noted that although most people living in endemic areas in the Sudan have developed antibodies against M. mycetomatis, and thus have been exposed to M. mycetomatis, only few of them actually developed eumycetoma [5], [6]. To date, it is unknown why some people are predisposed to develop eumycetoma.
Multiple explanations can be considered for the scanty susceptibility to eumycetoma. Firstly, genetic differences in the pathogen might exist that could lead to pathogenic and non-pathogenic variants of M. mycetomatis. Secondly, genetic polymorphisms in the host involved in sex hormone synthesis and neutrophil function have already been associated with eumycetoma, indicating that the genetic make-up of the host is a crucial factor in susceptibility to eumycetoma development [7]–[9]. Furthermore, a combination of specific genetic requirements and capabilities in both the pathogen and the host could lead to an even more sporadic development of eumycetoma. Thirdly, temporal conditions influencing the host immune response, such as co-infections, nutritional status, use of antibiotics and/or immune suppression or skewing may also play a role in susceptibility to M. mycetomatis. This possibility is supported by the observation that M. mycetomatis could only induce eumycetoma in animals in the presence of an adjuvant predisposing towards a Th2-response [10] but not a Th1-response [11], [12].
Skewing of the immune response is highly affected by invasive pathogens [13], and therefore, co-infections could play a critical role in eumycetoma [14]. In this respect, infections inducing a strong and long-lasting Th2-type of immune response could favour the development of eumycetoma disease most. Schistosomiasis seems to meet such requirements for the following reasons. Firstly, schistosomiasis induces a long-lasting Th2-type immune response that is strong enough to even convert an already established Th1-response [15], [16]. Secondly, in endemic countries schistosomiasis is often a chronic life-long disease. Even when patients are regularly treated for schistosomiasis, their continuous exposure to the parasite during fresh water contacts and the lack of the development of immunity against schistosomes will rapidly result in a re-infection with a persistent Th2-response.
Based upon the above mentioned observations, and the fact that we recently have shown that eumycetoma patients have increased concentrations of circulating IL-10 [7], we hypothesize that schistosomiasis, which induces a Th2-type response with elevated levels of IL-10, might increase the susceptibility to eumycetoma, whereas toxoplasmosis which induces a Th1-type response [16], [17], should not be associated with eumycetoma.
A total of 84 serum samples was taken from 53 eumycetoma patients and 31 controls, matched for age and gender, in the endemic areas of Sudan between 2001 and 2008 (Table 1). Serum samples were stored at −80°C until assay. The patients' demographic characteristics were recorded and that included gender, duration of disease, lesion size and site of infection. Eumycetoma was confirmed by culture and molecular identification based on sequencing the Internal Transcribed Spacer [18]. Written informed consent was obtained from all participants and ethical clearance was obtained from Soba University Hospital Ethical Committee, Khartoum, Sudan.
Specific IgG antibodies against Toxoplasma gondii were determined with the commercially available Toxo IgG II assay on the automated Liaison serology platform according to the manufacturer's protocol (Diasorin, Saluggia, Italy). Antibody levels against Schistosoma species were determined as described before by a combination of a commercial indirect hemagglutination test with Schistosoma mansoni adult worm antigens (IHA; Fumouze Laboratories, Levallois-Perret Cedex, France) and an enzyme-linked immunosorbent assay with homemade S. mansoni Soluble Egg Antigens (SEA) [19]. The IHA was considered positive when the titre was ≥1∶80 and the SEA ELISA was considered positive when the Optical Density (O.D.) at 492 nm was ≥0.15. For optimal specificity, Schistosoma spp. serology was only considered positive when a positive result was obtained in both the IHA and SEA-ELISA tests.
Antibody levels against Madurella mycetomatis Translationally Controlled Tumour Protein (TCTP) were measured with Luminex Technology as described before [5].
Difference in positive and negative serology for schistosomiasis and toxoplasmosis between eumycetoma patients and endemic controls were calculated with the Chi square test (GraphPad Instat 3.00) by determining both the two-sided p-value and the Odds Ratio using Yates correction. The 95% confidence interval of the Odds Ratio was calculated using the approximation of Woolf. The Mann-Whitney U test was used to compare differences between IgG levels raised against the MmTCTP antigen in the study populations (GraphPad Instat 3.00). The Kruskal-Wallis test (SPSS Inc 17) was used to test if concentrations of antibodies against Schistosoma spp. differed significantly between patients with larger lesions compared to patients with smaller lesions, by including size (small, moderate, large) as the grouping variable. A value of p<0.05 was considered significant.
Between 2001 and 2008, 53 patients and 31 endemic controls were included in the study. Most patients were male and had eumycetoma of the foot (Table 1). All patients and controls came from the same area mainly from Central Sudan as indicated in Figure 1.
As shown in Figure 2, eumycetoma patients were significantly more often sero-positive for Schistosoma infections as compared to endemic controls (Chi square, p = 0.03). In other words, an association exists between the infections caused by Madurella mycetomatis and Schistosoma spp. since the odds ratio for co-occurring schistosomiasis in eumycetoma patients is 3.2 (95% Confidence Interval 1.18–8.46). In contrast, eumycetoma patients were not significantly more often sero-positive for Toxoplasma gondii infections (Figure 2, Chi square, p = 0.5) thus the risk for developing eumycetoma does not seem to be increased in case of concurrent toxoplasmosis (Odds ratio 1.5, 95% Confidence Interval: 0.60–3.75). No correlation was found between sero-positivity for schistosomiasis and sero-positivity for toxoplasmosis (Chi square, p = 0.1, Odds ratio 0.44, 95% Confidence Interval: 0.14–1.32).
Positive serology for either schistosomiasis or toxoplasmosis was not correlated to the size of the eumycetoma lesion (Chi square, p = 0.8 and p = 0.2, respectively, data not shown). Since schistosomiasis is known to reduce the humoral immune response against co-infecting pathogens [20], [21], we investigated the antibody response against M. mycetomatis antigen TCTP. The antibody levels against TCTP did not differ between the eumycetoma patients with positive schistosomiasis serology and those with negative schistosomiasis serology, nor did they differ between matched endemic controls with positive schistosomiasis serology and healthy endemic controls with negative schistosomiasis serology (Figure 3). This suggests that schistosomiasis does not influence the humoral immune response against the TCTP antigen of M. mycetomatis.
Schistosomiasis is a chronic disease with an estimated 200 million people infected in subtropical countries [22]. Therefore, almost all schistosomiasis patients will subsequently be infected by one or more additional pathogens. Although a prior infection with schistosomes often has an effect on the subsequent infection by a virus, bacterium, protozoan or other helminth, schistosomiasis can cause both an increase or a decrease in the severity of the subsequent infection for yet unknown reasons (reviewed in Abruzzi and Fried [23]. Decreased subsequent disease severity was observed for co-infections with Helicobacter pylori, Fasciola hepatica, Echinostoma and with Plasmodium in case of S. haematobium schistosomiasis [23]. In addition, a worsened outcome of a subsequent co-infection has been described for HIV (reduced viral clearance) [24], Leishmania donovani [25], Toxoplasma gondii, Entamoeba histolytica infections and Plasmodium in case of S. mansoni schistosomiasis [23], [26]. The effect of subsequent fungal infections has not been addressed yet.
Since eumycetoma infections can only be established in animals with adjuvants inducing a strong Th2-type immune response, we hypothesized that co-infections inducing a Th2-type immune response would predispose to eumycetoma disease. This study compared eumycetoma patients with matched endemic controls without eumycetoma for co-infections with Schistosoma spp.. As a control, Toxoplasma gondii infections were monitored since this infection is also endemic in Sudan and results in a Th1-type of immune response in immune-competent hosts [17].
Although we only studied a limited number of people, which might not represent the full population of the study area, we did find an overall sero-prevalence of schistosomiasis in our study population of 51%, which is consistent with the earlier reported variable sero-prevalences for schistosomiasis in Sudan in the New Halfa and Um Zukra villages in the Gezira and Kassala regions (Figure 1) (16% and 70%, respectively) [27]–[29]. Large variations in prevalence of schistosomiasis even occur among villages in close proximity and depend on multiple factors, such as the environmental conditions for the intermediate snail host and the hygiene and bathing habits of the inhabitants [22]. The overall sero-prevalence of Toxoplasma gondii in our study population was 42%, which was exactly the same as found by Abdel-Hameed et al. in Gezira in 1991 [30]. This study now showed that eumycetoma patients were significantly more often sero-positive for schistosomiasis when compared to matched, endemic controls. The correlation is strengthened by the fact that antibody levels against Toxoplasma, another prevalent infection in Sudan, did not correlate with eumycetoma disease. Furthermore no correlation was found between seropositivity of schistosomiasis and toxoplasmosis.
The main underlying cause may be the strong Th2-response induced by schistosomiasis and the high expression of interleukin-10 (IL-10) by Th2 cells [16], [31]. IL-10 is capable of inhibiting synthesis of pro-inflammatory cytokines. Another anti-inflammatory trait of IL-10 is its potent ability to suppress the antigen-presentation capacity of antigen presenting cells. Increased IL-10 cytokine levels have also been detected in mycetoma lesions [7], [32] as well as in animal models of actinomycetoma caused by Nocardia brasiliensis [33]. Moreover, IL-10 levels were significantly elevated in serum of M. mycetomatis eumycetoma patients in Sudan [7], suggesting that IL-10 concentrations also play an important role in development or maintenance of eumycetoma.
In conclusion, even among this relatively small number of patients, eumycetoma was significantly associated with schistosomiasis and not with toxoplasmosis. Since this correlation is only based on serological data, animal studies are currently performed to investigate the precise role of schistosomiasis and Th2-predisposition in the development of fungal eumycetoma.
|
10.1371/journal.pgen.1000476 | Microdissection of Shoot Meristem Functional Domains | The shoot apical meristem (SAM) maintains a pool of indeterminate cells within the SAM proper, while lateral organs are initiated from the SAM periphery. Laser microdissection–microarray technology was used to compare transcriptional profiles within these SAM domains to identify novel maize genes that function during leaf development. Nine hundred and sixty-two differentially expressed maize genes were detected; control genes known to be upregulated in the initiating leaf (P0/P1) or in the SAM proper verified the precision of the microdissections. Genes involved in cell division/growth, cell wall biosynthesis, chromatin remodeling, RNA binding, and translation are especially upregulated in initiating leaves, whereas genes functioning during protein fate and DNA repair are more abundant in the SAM proper. In situ hybridization analyses confirmed the expression patterns of six previously uncharacterized maize genes upregulated in the P0/P1. P0/P1-upregulated genes that were also shown to be downregulated in leaf-arrested shoots treated with an auxin transport inhibitor are especially implicated to function during early events in maize leaf initiation. Reverse genetic analyses of asceapen1 (asc1), a maize D4-cyclin gene upregulated in the P0/P1, revealed novel leaf phenotypes, less genetic redundancy, and expanded D4-CYCLIN function during maize shoot development as compared to Arabidopsis. These analyses generated a unique SAM domain-specific database that provides new insight into SAM function and a useful platform for reverse genetic analyses of shoot development in maize.
| All the organs of plant shoots are derived from the shoot apical meristem (SAM), a pool of plant stem cells that are both organogenic and self-sustaining. These dual SAM functions take place in distinct yet adjacent meristematic domains; leaves are derived from the peripheral zone (PZ) of the SAM whereas cells lost during organogenesis are replenished from the central zone (CZ). Deciphering the global patterns of differential gene expression within these discrete SAM functional domains is integral toward understanding the molecular-signaling networks that regulate plant development. We utilized laser-microdissection technology to isolate tissues from the SAM crown and center (SAM-proper) and from initiating leaves (P0/P1) at the SAM periphery for use in microarray comparisons of gene expression within these SAM functional domains. Nine hundred and sixty-two maize genes were differentially expressed, confirming that the distinct functions of these meristematic domains involve widespread differences in gene expression. Genes involved in cell division, cell wall biosynthesis, chromatin structure, and RNA binding are especially upregulated in initiating leaves, whereas genes regulating protein stability and DNA repair are upregulated in the SAM proper. Mutations in a D-cyclin gene that was upregulated in the P0/P1 render narrow-leafed mutant plants with defective stomatal patterning, providing functional genetic data for a previously uncharacterized maize gene.
| The maize shoot apical meristem (SAM) is a complex signaling network of distinct structural and functional domains that performs two essential developmental functions during plant shoot development: (1) self-maintenance and (2) organogenesis. Responsible for the development of all above ground organs in the plant, the SAM must maintain a precise equilibrium during which cells lost to newly-initiated leaves are replenished to maintain the SAM proper. Comprised of two tissue layers, the single-celled tunica (L1) and a multilayered corpus (L2), the maize SAM displays histological zonation that is correlated with its functions (Figure 1A). Determinate lateral organs arise from the peripheral zone (PZ) whereas the central zone (CZ) is comprised of more slowly dividing meristem initial cells that replenish the SAM. Although Caspar Wolff first recognized the SAM as the organogenic center of the plant shoot almost 250 years ago [1], detailed mechanisms of SAM function remain a fundamental question in plant biology.
Molecular genetic analyses have identified a growing number of genes contributing to the complexity of SAM function in maize. The homeobox gene knotted1 (kn1) is required for meristem indeterminacy; null kn1 mutants fail to maintain the SAM [2],[3]. Down-regulation of KN1 accumulation in the PZ precedes lateral organ initiation, and is correlated with auxin transport and expression of the knotted1-homeobox (KNOX) regulator rough sheath2 (rs2) in the PZ [4]–[7]. SAM size is also controlled by the cytokinin-inducible RESPONSE REGULATOR ABPHYL1, in which mutations increase SAM size and lead to disrupted phyllotaxy in the maize shoot [8],[9]. Maize leaves are formed via recruitment of ∼200 leaf founder cells from the PZ of the SAM [10], followed by differentiation along three developmental axes (proximo-distal/medio-lateral/adaxial-abaxial). Genetic analyses have identified several maize genes involved in these SAM functions, including those required for leaf initiation and phyllotaxy (terminal ear1 and aberrant phyllotaxy1), proximodistal patterning (rs2 and semaphore1), mediolateral development (narrow sheath1&2, ragged seedling2, wavy auricle in blade1), and adaxial-abaxial patterning (rolled1, miR166, leafbladeless1, milkweed pod1) [11], [8], [12]–[19]. Elucidation of the regulatory networks that coordinate these intersecting developmental functions will be bolstered by the use of genomic approaches to generate testable models for the SAM interactome, followed by comprehensive genetic and biochemical analyses to test and extend these hypotheses.
The complementary expression domains of the molecular markers rs2 and kn1 clearly illustrate that indeterminate cells of the SAM proper are immediately juxtaposed to leaf founder cells within the maize shoot apex (Figure 1B). The close proximity of these distinct functional domains presents technical barriers to comparative analyses of these discrete SAM functions. However recent technical advances have enabled a genomics approach toward the molecular dissection of SAM function. The relatively large size of the maize SAM, 50–250 founder-cells are recruited into the incipient leaf versus 25–30 in Arabidopsis [10],[20], renders the maize plant especially tractable to laser-microdissection technologies. Laser-microdissection permits the precise isolation of specific tissues, organs, or cells from fixed and sectioned plant tissues adhered to microscope slides [21]. Nanogram quantities of RNA extracted from less than 1 mm2 of microdissected tissue (comprising five to ten whole SAMs) can be linearly amplified using T7 RNA polymerase to generate microgram quantities of RNA sufficient for transcriptional profiling using microarray technology [22]–[26]. Owing to its unique ability to sample discrete microdomains in plant tissues, laser-microdissection eliminates the transcriptional noise contributed by adjacent or contaminating unrelated tissues and thereby enables transcriptional profiling that is focused on the cells and tissues of interest.
Laser microdissection-microarray technology was utilized in comparative transcriptional analyses of functional domains in the maize SAM. Gene expression within SAM microdomains encompassing the initiating maize leaf (P0/P1) and the stem cells of the SAM-proper was analyzed; 962 maize genes were differentially expressed in this comparison. Control genes of known expression domain confirmed the accuracy of the laser microdissections and validate the dataset. Genes predicted to function during cell division/growth, chromatin remodeling, RNA-binding, cell wall biosynthesis and translation are especially upregulated in initiating leaves, whereas genes involved in protein fate and DNA repair are prevalently expressed within the SAM-proper. In situ hybridization analyses, and qRT-PCR analyses of apices that are arrested in leaf initiation identified twelve maize genes predicted to function during leaf initiation. Reverse genetic analyses of the maize D-cyclin gene asceapen1 (asc1) confirmed its predicted function during maize leaf and shoot development; novel mutant phenotypes revealed differing levels of genetic redundancy and divergent patterns of subfunctionalization among cyclin paralogues in maize and Arabidopsis. Our data provide a unique database that provides insight into SAM function and a useful platform for reverse genetic and biochemical analyses of maize shoot development.
Maize seedlings were grown under controlled conditions and processed for laser microdissection of SAM domains (see Materials and Methods). Although KNOX immunohistolocalization analyses clearly delineate the leaf/non-leaf boundary in the maize shoot apex (Figure 1B), these treatments require crosslinking fixatives that preclude the extraction of RNA from microdissected tissues. Therefore, in lieu of molecular makers, two distinct SAM domains were captured using morphological/anatomical cues (Figure 1 C–D). The “SAM-proper” comprised the apical crown and central stem of the shoot apex, and is estimated to include the CZ. Tissue extracted from the “P0/P1” domain included the PZ and the newest-initiated lateral organ that formed a protruding buttress on the SAM flank. Care was taken to avoid the SAM peripheral zone during captures of the SAM-proper domain; likewise the P0/P1 samples were harvested from a depth of no more than three cell-layers in order to avoid tissues that typically accumulate KNOX proteins (Figure 1B), markers of meristematic identity [2]. Tissues derived from ten total SAMs were pooled into domain-specific samples comprising a single biological replicate.
Following RNA extraction and amplification (see Materials and Methods), six such biological replicates were utilized in microarray hybridizations to 29,600 total elements (including approximately 23,000 unique maize genes) contained on the customized maize cDNA microarrays SAM1.1 and SAM3.0 [described in 25],[26]. Replete with genes identified from meristematic tissues, SAM 1.1 contains over 7,500 cDNAs derived from maize inflorescences and SAM 3.0 contains over 10,500 cDNAs derived from vegetative apices (i.e. SAM plus four leaf primordia). For each array platform, three of the six cDNA pairs were labeled with Cy3 from the SAM-proper and Cy5 for the P0/P1. Dye assignments were reversed for the other three replications. Normalized Cy5 and Cy3 signals were used to test for evidence of differential expression among the SAM domains using a linear model analysis for each gene (see Materials and Methods).
A total of 1,312 array elements were differentially expressed in these SAM domains utilizing cut-off parameters of P-value<2.93 E-4 and fold change>2.0. Alignment of these 1,312 cDNA sequences to predicted genes within the sequenced maize genome (see Materials and Methods) identified 962 maize gene contigs (MGC) that were differentially expressed in the SAM-proper and P0/P1 leaf primordia (Table S1). These included 542 genes upregulated in the P0/P1 leaf and 420 genes upregulated in the SAM-proper (Figure 2). Notably, 48 (i.e. 3.6%) of the 1,312 differentially expressed array elements did not align to any sequenced MGC; the EST accession numbers of these unaligned genes are thus listed among the 962 genes contained in Table S1. None of these unaligned EST sequences are predicted to comprise repetitive retrotransposons, but presumably correspond to a portion of the maize gene space that is as yet unrepresented in the sequenced portion of maize genomic DNA. Approximately 31.8% of the cDNA array dataset aligned with equal affinity to multiple MGCs, and are likely to comprise gene family members for which the available cDNA sequence does not distinguish between close paralogs (Table S1).
The estimated false discovery rates were 1.1% for SAM1.1 and 0.5% for SAM3.0. Bioinformatic predictions of function were performed for all differentially expressed genes as described [27], and are presented at the SAM-The Maize Shoot Apical Meristem Project database created during this project (http://sam.truman.edu/geneva/geneva.cgi). A total of eighteen different Gene Ontology (GO) functional categories are identified as detailed below (Figure 2), including: DNA repair; photosynthesis related; RNAi; transposable element; other; cytoskeletal; extracellular matrix/cell wall; signal transduction; cell division/growth; protein fate; RNA binding; stress related/defense; vesicle trafficking/transport; transcription; chromatin; metabolism; translation; and unknown.
Control genes whose expression in either the SAM-proper or the P0/P1 is described previously attested to the precision and accuracy of the SAM domain microdissections (Figure 1D; Table 1 and references therein). For example, the meristem maintenance gene knotted11 (kn1; [2], see Dataset S1 for a list of the MGC accession numbers corresponding to superscripted numerals in this text), the phyllotaxy regulator terminal ear12 (te1; [11]), the trans-acting siRNA (tasiRNA) biogenesis gene leafbladeless13 (lbl1; [28]) and the maize homolog of the sterol biosynthetic gene fackel4 [29] were all identified in our microarrays as up-regulated in the SAM-proper, in agreement with published expression analyses. Likewise, the knox-regulatory gene rough sheath25 (rs2; [6],[7]), a maize homolog of growth-regulating factor16 (grf1; [30]), a maize auxin response factor5/monoteros17 (arf5/monopteros; [31]), and several members of the yabby gene family of transcription factors (yab158, yab109; Zm-drooping leaf-like10; [16],[26]) were all up-regulated in the P0/P1 domain, as predicted from previous studies. Control genes also up-regulated in the initiating leaf included maize orthologues of the auxin transporters pinformed111 (pin1) and auxin insensitive112 (aux1), as well as the cell wall-loosening gene beta expansin813 (expb8), all of which are known to be expressed during leaf initiation in maize and/or Arabidopsis [32]–[35].
Microarray analyses of the SAM-proper and P0/P1 apical domains reveal discrete GO functional categories of preferentially expressed genes (Figure 2; Table S1). For example, significantly more (P<0.036) genes involved in protein fate/ubiquination were found to be upregulated in the SAM-proper (42) as compared to the P0/P1 (24), including multiple paralogues encoding a predicted E2 UBIQUITIN CONJUGATING ENZYME-LIKE14. Also, three-fold more DNA-repair genes were upregulated in the SAM-proper than in the P0/P1, including maize orthologues of rad2315, radA16, mus117 and the SNF2 domain/helicase protein18.
Genes comprising five predicted functional categories were significantly upregulated in the P0/P1, including those functioning in the extracellular matrix/cell wall (P<6.10E-05), cell division/growth (P<0.005), RNA binding (P<0.009), chromatin (P<2.64E-04), and translation (P<4.76E-07). Fifteen genes involved in cell wall biology were upregulated in the P0/P1, whereas none were upregulated in the SAM-proper. These included genes encoding cell wall GLYCOPROTEINs19 and GLYCOSYLASES20, an ALPHA-EXPANSIN21, a BETA-EXPANSIN13, and a CELL WALL-ANCHORED PROTEIN22.
Differentially expressed genes encoding proteins involved in cell division and growth in the P0/P1 outnumbered those identified in the SAM-proper twenty-three to seven (Figure 2). Reflecting the increased mitotic activity found in the peripheral zone and initiating leaf as compared to the SAM central zone, these included genes encoding various CYCLINs23–27 and a putative maize homolog of mammalian growth regulating factor16. Also identified are at least four maize paralogs of the TRANSLATIONALLY-CONTROLLED TUMOR PROTEIN (TCTP28–31), guanine exchange factors that control organ size in Drosophila and mammals by regulating a specific dRheb-GTPase within the target of rapamycin (TOR) signaling pathway [36]. Recent analyses of a TCTP gene in Arabidopsis revealed increased expression in rapidly growing tissues; reverse genetic mutant plants exhibited a range of developmental defects including reduced cell size and leaf expansion, and decreased sensitivity to auxin [37].
Three distinct argonaute1-like maize paralogues32–34 were identified in our microarray data, all of which were upregulated in the P0/P1 (Table S1). Arabidopsis contains two close paralogues, argonaute1 (ago1) and pinhead1/zwille1 (pnh1/zll1), which encode components of the multi-subunit RNA-induced silencing complex (RISC; [38]). In keeping with their partially overlapping roles in the miRNA-regulated control of leaf polarity and of SAM maintenance, ago1 is evenly expressed throughout the Arabidopsis SAM and young leaf primordia [39]–[42], whereas pnh1/zll1 transcripts accumulate preferentially in leaf primordia and in the vasculature [43]–[45]. Owing to the nearly identical amino acid sequences of AGO1 and PNH1, it is not possible to predict which of maize ago-like genes are ago1 orthologues and which are pnh1/zll1 orthologues. However, in situ hybridization analysis of a maize ago1-like gene that was upregulated more than seven-fold in the P0/P1 confirmed our microarray data, and revealed a pnh1/zll-like expression pattern (Figure 3A). Although transcripts are indeed detected in the SAM crown and center, more abundant transcript accumulation is observed in the leaf founder cells, the SAM periphery, and in young leaf primordia (Figure 3A).
In contrast to the miRNA regulatory ago1 genes identified in the P0/P1, the tasi-RNA gene lbl13 [28] and the siRNA effector protein gene argonaute435 (ago4; [46],[47]) were upregulated in the SAM-proper. Moreover, significantly more RNA-binding genes were preferentially expressed in initiating maize leaves compared to the SAM-proper, including four genes predicted to encode RIBONUCLEOPROTEINs36–39, numerous GLYCINE-RICH RNA-BINDING PROTEIN40–44 paralogs, and a maize homolog of the Arabidopsis flowering-time regulator flowering locus K45 gene (flk; [48]; Table S1).
A preponderance of gene elements predicted to function during chromatin structure and remodeling were upregulated in the P0/P1 versus the SAM-proper (45 versus 16, respectively; Figure 2). For example, a cytosine-5-methyltransferases46 and three methyl-CpG DNA binding domain47–49 genes are specifically upregulated in the P0/P1. Likewise, whereas hdt3–like50 and sir2-like51 histone deacetylase gene are upregulated in the leaf initials, a hdt2-like histone deacetylase52 is highly expressed in the SAM-proper and three swib-domain56–58 gene paralogs are detected only in the SAM-proper.
Although the number of putative transcription factors (TFs) preferentially expressed in either the SAM-proper or the P0/P1 is exactly equal at thirty-three each (Figure 2), each SAM domain exhibited upregulation of various distinct TF genes not identified in the other. For example, the founding member of the knotted-like homeobox (knox) gene family kn11 is differentially expressed in the SAM-proper, while the related knox gene gnarley159 (gn1) is upregulated in the P0/P1. Although a previous report detected gn1 expression in the shoot apex [49], the SAM domain specificity of gn1 expression was not described previously. Likewise, a maize homologue of the Arabidopsis leunig co-repressor60 [50] gene is identified in the SAM-proper, as were three paralogs encoding B3-domain61–63 TFs. Developmental regulators of embryo and meristem development, many B3-domain TFs are shown to function via interaction with auxin or ABA signaling pathways [51],[52, and references therein]. In contrast, rs25, auxin response factor264 (arf2), and multiple members of the yabby8–10 gene family were identified in the P0/P1. RS2 represses knox gene expression in developing leaves [6],[7], whereas arf2 accumulates in Arabidopsis lateral organs [53] and maize yabby genes are transcribed in the P0 and leaf primordia (Figure 3B; [16],[17],[26].
Lastly, the largest single gene category identified in our microarray analyses comprised genes of unknown predicted function, which contained 142 genes upregulated in the SAM-proper and 133 genes in the P0/P1 (Figure 2).
Distinct gene paralogs of the histone-methylating SET DOMAIN-encoding gene53–55 family are upregulated in the immediately adjacent apical domains that comprise the P0/P1 and SAM-proper. In addition, paralogs of six other maize gene families including auxin response factor165, 66 (arf1), histone367, 68, histone469, 70, ubiquitin-conjugating enzyme E271, 72, ADP-ribosylation factor/Secretion-associated and Ras-related73, 74 protein (ARF/SAR), and the ubiquitin-ligase subunit gene S-phase kinase-associated protein175, 76 (skp1) exhibit preferential expression within the SAM-proper and the P0/P1, and thus provide intriguing evidence for subfunctionalization of these gene families within discrete functional zones of the maize SAM.
Focusing on previously uncharacterized maize genes implicated during leaf development, six genes upregulated in the P0/P1were subjected to in situ hybridization analyses in order to verify the domain-specific transcript accumulation predicted from our microarray data and identify novel patterns of gene expression. In addition to the maize ago1-like gene described in Figure 3A, five genes whose functions are yet to be demonstrated in maize were analyzed. These included a putative oligopeptide transporter77, a yabby gene drooping leaf110, a predicted growth-regulating factor6, a lipid-transfer protein78, and a D4-class cyclin25. In all cases the pattern of transcript accumulation observed in the in situ hybridizations correlated with the microarray data. Stronger signals were observed in the SAM periphery, P0, and small leaf primordia as compared to the SAM crown and center (Figure 3), which verified the P0/P1 upregulated expression observed in our microarray analyses.
Auxin transport is the earliest-demonstrated prerequisite to KNOX downregulation and leaf initiation from the SAM flank; disruption of auxin transport by the chemical inhibitor N-1-naphthylphthalamic acid (NPA) leads to the arrest of lateral organogenesis in plant shoots [54],[4],[5]. Therefore, NPA-induced inhibition of leaf initiation provides a compelling experimental system with which to monitor SAM gene expression during very early events in leaf development. Toward this end, 14-day-after-germination seedling shoot apices were dissected to remove all organs except the SAM and the six youngest leaf primordia and placed in tissue culture with or without 30 mM NPA as described [4]. As shown in Figure 4A–B, NPA-cultured SAMs became greatly elongated but failed to initiate any new leaves, whereas equivalent sibling apices generated 6–7 new leaf primordia in NPA-free culture. After 14 days in culture, samples were processed for SAM laser-microdissection mediated qRT-PCR analyses as described [55]. Genes found to be upregulated during leaf initiation but down-regulated during NPA-induced arrest of organogenesis are especially implicated to function during early stages of maize leaf development.
Transcript accumulation analyses were performed for nine genes that were significantly upregulated in the P0/P1 in our microarray analyses (Table S1). As shown in Figure 4D, qRT-PCR analyses revealed that transcripts of six of these nine genes were also down-regulated in leaf-arrested apices, including a second maize ago1-like32 gene, a putative brassinosteroid response factor79 gene, an E2 ubiquitin-conjugating-like71 paralog, the aux112 auxin transporter gene, a yabby 158 gene, and the growth-regulating factor16. One gene (a putative seven-in-absentia-like ubiquitin ligase80) was weakly down-regulated in NPA-treated apices, whereas two genes (a tctp-like31 gene paralog and a maize AMP-dependent synthetase81) were unchanged in NPA-treated versus untreated shoots. Thus, six genes identified as upregulated in the P0/P1 are downregulated in shoot apices that are arrested in leaf initiation. We speculate that the three genes whose expression levels were unchanged following NPA treatment may mark a domain within the PZ that functions upstream or independent of auxin transport during leaf initiation, since accumulation of some PZ markers has been shown to persist in Arabidopsis pin1 mutants and in tomato apices treated with NPA [54]. Alternatively, these NPA-unaffected genes may not be preferentially expressed during early leaf initiation.
The differential gene expression data presented in this study identify genes implicated in SAM domain-specific functions during maize shoot development. Validation of these predicted functions, however, requires biochemical or genetic analyses. A reverse genetic strategy was implemented (see Materials and Methods) to investigate the function of a D4-class cyclin25 gene that was identified as upregulated in the leaf primordia, which we have named asceapen1 (asc1). In situ hybridization of seedling shoot apices verified the P0/P1 upregulated asc1 expression observed in the microarray analyses (Figure 3E). The asc1 gene contains six exons (Figure 5A) and is located at position 39,743–41,597 of contig 45 on maize chromosome 7. The 1068 bp open reading frame is predicted to encode a protein of 355 amino acids, which contains the canonical LxCxEx RETINOBLASTOMA-interaction domain that is characteristic of D-CYCLINS (Figure 5B;Wang et al., 2004). In addition, ASC1 contains the conserved amino cyclin box and a CYCLIN recognition motif. Limited expression profiles are described for four related maize D-cyclins (including a D2-cyclin, a D4-cyclin, and two D5-cyclins [56], although expression within the vegetative SAM was not examined. No genetic analyses of D-CYCLIN function have been performed previously in maize.
D-CYCLINS perform an evolutionarily conserved growth-regulatory function to regulate progression through the G1 phase of the cell cycle [57]. Genetic analyses in Arabidopsis suggest that D-CYCLINS function as important regulators of asymmetric cell division, a process that is critical to developmental differentiation and has played a pivotal role in the evolution of multicellularity [reviewed in 58],[59]. Arabidopsis has 10 CYCD genes comprised of six subgroups (CYCD1, CYCD2, CYCD4 (2 genes), CYCD3 (3 genes), CYCD5, CYCD6, and CYCD7 [60]. Overexpression analyses suggest that as a group, D-CYCLINS may regulate the developmental progression from cell proliferation to differentiation [61],[62],[63]. Genetic analyses reveal redundant functions for the three CYCD3 genes in Arabidopsis [64]. Single CYCD3 mutations yield non-mutant phenotypes; triple mutations condition small yet fertile plants with narrow leaves, a small SAM, and decreased cytokinin response. CYCD4;1 is expressed in both shoot and root apices, although CYCD4;2 is not detected in the SAM. Single mutations in CYCD4;1 and CYCD4;2 render no macrophenotype, although reduced numbers of anatomically normal stomata develop in mutant hypocotyls [65].
A phylogenetic analysis was performed on the maize ASC1 protein and thirteen additional plant D-CYCLINS for which transcriptional analyses and/or genetic analyses are documented [66],[56],[67], including four additional maize D-CYCLINS, ten Arabidopsis D-CYCLINs, and three D-CYCLINs from Antirrhinum majus. Utilizing the D1-CYCLIN from the moss Physcomitrella patens as an outgroup. ASC1 was placed on a well-supported clade together with the D4-CYCLINS and D2-CYCLINS from Arabidopsis (Figure 5C). All the other D-CYCLIN proteins were placed on separate clades; the D3-CYCLINS from Arabidopsis and Antirrhinum comprise a well-supported separate clade from ASC1. Reverse genetic analyses of asc1 were instigated in order to investigate the function of this D4-CYCLIN in maize.
F2 seedlings were obtained from self-pollination of over 3,000 maize plants with Mutator (Mu) transposon activity, a maize transposon with an unusually high forward mutation rate [68]. A PCR-based reverse genetic strategy similar to previously published protocols ([69],[70]; see Materials and Methods) identified two independently-segregating Mu-insertion alleles of the asc1 gene, designated asc1-M1 and asc1-M2 (Figure 5). The asc1-M1 allele harbors a Mu4 insertion in position 47 of the 129 bp intron 1, whereas the predicted null asc1-M2 allele harbors a Mu1 at position 31 of the 87 bp second exon (Figure 5A). RNA gel-blot hybridization analyses reveal that asc1 transcript accumulation is greatly diminished in asc1-M1 homozygotes and is virtually absent in asc1-M2 homozygous plants, relative to non-mutant siblings (Figure S1).
F2 progeny of plants heterozygous for asc1-M1 or asc1-M2 each segregate for short, infertile plants with very narrow leaves (Figure 6A–B). These mutant phenotypes co-segregate with homozygosity for asc1 mutations, and interallelic crosses of plants heterozygous for asc1-M1 and asc1-M2 fail to complement (Figure 6N). No female inflorescences (ears) are observed in homozygous asc1 mutant plants, and male inflorescences form only rudimentary tassels with sterile branches and no floral morphogenesis (Figure 6B–D). Histological examinations of asc1 mutant seedlings reveal extremely narrow leaves and small vascular bundles with reduced numbers of xylem and phloem vessels, as well as reduced SAM size (Figure 6E–J). Both mutant alleles conditioned equivalent phenotypes, although the range of phenotypes is more severe in plants homozygous for the exon-insertion allele asc1-M2.
Especially striking are the effects of asc1 mutations on stomatal patterning and anatomy. Comprised of two subsidiary cells that surround and appress two smaller guard cells, interspaced stomatal complexes are formed via a series of ordered, asymmetric cell divisions in the leaf epidermis(Figure 6K; reviewed in [71]). Analyses of the asc1 mutant leaf epidermis reveal irregular stomatal patterning (Figure 6L–N). Two mutant stomatal complexes often form immediately adjacent to one another, a pattern not observed in non-mutant leaves. Other abnormalities include enlarged, distorted, and supernumery subsidiary cells and guard cells, which often develop immediately adjacent to completely normal stomatal complexes.
Novel transcriptomic comparison of the functionally distinct microdomains within the maize SAM are presented, an analysis that was enabled by the relatively large size (∼120 µm) of the maize SAM as compared to Arabidopsis. The differential expression of 275 unknown genes within a particular SAM domain thereby provides a first suggestion of their potential function. Moreover, the documentation of seven instances wherein gene paralogues exhibited subfunctionalized preferential expression within distinct SAM microdomains provides insight into the evolution of specific gene families in maize. The entirety of this unique expression database is publicly available (http://sam.truman.edu/geneva/geneva.cgi) and represents a starting point for subsequent reverse genetic analyses of SAM function in maize. In addition, these genomic analyses are likely to uncover genes whose functions that are not amenable to traditional genetic analyses, owing to the embryo/seedling lethality that may result from mutations in genes required for early events in SAM ontogeny and/or leaf initiation.
Genes whose SAM domain-specific expression are previously described in maize or Arabidopsis (Table 1) served as experimental controls for the analyses presented here, and attest to the power and precision of laser-microdissection for analysis of transcript accumulation within plant microdomains. Three distinct paralogs of ago132–34 are identified, each of which was upregulated in the P0/P1. Arabidopsis contains two ago1-like genes, one of which (pnh1/zll1) is preferentially expressed in leaf primordia whereas the other (ago1) is evenly expressed throughout the SAM and young leaves [39]–[45]. Although the extreme amino acid conservation observed in PNH1 and AGO1 precludes the identification of specific maize orthologues from homology alone, in situ hybridization analysis verified the leaf-preferential expression of one maize ago gene (Figure 3A) whereas a separate ago1 paralog was downregulated in NPA-treated apices that are arrested in leaf initiation (Figure 4D). Our results are analogous to the reported leaf-upregulated expression of the rice pinhead/zwille orthologue OsPNH1 [72]. We speculate that maize co-orthologs whose expression domains mirror that of the Arabidopsis ago1 gene would not be identified in our microarray analyses, owing to the relatively equivalent transcript accumulation in the SAM-proper and P0/P1domains. In support of this hypothesis, the SAM1.1 and SAM3.0 gene chips contain additional ago1 co-orthologs that were not detected as differentially expressed in this analysis. It appears likely that as in Arabidopsis, the maize ago1 gene family has expanded and paralogs became subfunctionalized to perform specialized tasks during leaf development and/or shoot meristem maintenance. Plant miRNAs are described that function in the ARGONAUTE1-directed regulation of leaf initiation and polarity, including miR166, and miR156 (reviewed in [73],[74]). Although these regulatory RNAs and/or their mRNA precursors are detected in both the SAM-proper and the initiating leaves of maize, mature microRNAs preferentially accumulate in the P0 and leaf primordia [75], which may be functionally correlated with the differential expression of one or more of the ago1-co-orthologs identified herein.
We speculate that the SAM-upregulated expression of a maize ago4-like35 gene, predicted to function during regulation of siRNA-induced gene silencing and maintenance of DNA methylation [46],[47], may be elicited in response to the pronounced upregulation of retrotransposon transcription that is observed in the maize SAM [25]. Moreover, six genes predicted to function during DNA repair are upregulated in the SAM-proper versus just two in the P0/P1 (Figure 2), which may reflect selective pressures to maintain a mutation-free DNA template in the indeterminate, stem cell population of the meristem. Ultimately, the SAM is the source of all the somatic cells comprising the plant shoot, as well as the germinal cells within floral organs. While DNA repair is certainly occurring in the P0/P1, spontaneous mutations in the DNA of sterile, determinate leaf primordia may be subjected to weaker selective pressure as compared to the SAM.
Of the 66 protein fate genes upregulated in our microarray analyses, 42 were identified in the SAM-proper (Figure 2). Although ubiquitination and additional mechanisms of proteolysis are widespread throughout plant tissues, these data suggest the particular importance of these proteolytic pathways during SAM function. Previous studies in Arabidopsis and rice revealed that 26S proteaosome-dependent proteolysis is required for shoot meristem maintenance and identity [76]–[79]. Our data suggest that 26S proteaosome-dependent proteolysis is also important during the function of the maize SAM, and likewise implicates ubiquitin-related proteases, serine carboxy peptidases, OTU-like cysteine proteases, CLP proteases, aspartic proteases and various SUMO proteins during SAM function (Table S1).
Multiple categories of gene function are identified as upregulated during leaf initiation (Figure 2). Genes involved in cell wall biosynthesis and cell division/growth are logically co-regulated, and both gene categories are significantly upregulated in the P0/P1. Although it is true that cell division is absolutely required in the CZ of the SAM in order to replace cells lost during organogenesis and to maintain the meristematic stem cell population, live imaging in Arabidopsis has shown that mitotic activity in the PZ during leaf initiation is more expansive and proceeds at a faster rate than in the SAM proper [80]. Therefore, our array data are in agreement with both classical (reviewed in [81]) and recent descriptions of differential cell division rates within SAM functional zones. Although the maize yabby-like8–10 genes upregulated in the P0/P1were placed in the separate GO category of transcription, the YABBYs are likewise presumed to function during expansive organ growth [17],[26],[82].
A maize homolog (Zm-grf16) of a family of transcription factors that regulate cell expansion in Arabidopsis leaves and cotyledons [30] was also identified in the P0/P1 dataset (Table 1). Bioinformatic analyses reveal that grf1 homologs in Arabidopsis and rice have complementary target sites for miR396, a relatively rare small RNA that is either expressed at very low levels or in a limited number of cells/tissues [83]. Zm-grf1 is expressed in the SAM periphery (Figure 3D) and leaf primordia and is downregulated in leaf-arrested SAMs (Figure 4), implicating a function very early in maize leaf development. Zm-grf1 also harbors the conserved miR396 recognition motif, and thus represents an intriguing candidate gene for reverse genetic analyses of microRNA-regulated leaf development. The grf genes function redundantly in Arabidopsis [30] and at least eight grf1sequence paralogues are present in maize, suggesting that reverse genetic analyses utilizing RNAi approaches or miR396-resistant transgenes may be more informative than characterization of Zm-grf1 knockout alleles.
Nearly three times as many gene elements involved in chromatin structure and remodeling were upregulated in the P0/P1 as in the SAM-proper (45 versus 16). These data may reflect the fundamental and widespread changes in chromatin that are predicted to accompany the switch from meristematic to leaf developmental programs. Alterations in chromatin structure are inherent when changing from the propagation of an extant developmental state (i.e. the SAM) to the installation of a new developmental program (i.e. leaf initiation), and may be further enhanced during the transition from an indeterminate to a determinate developmental field.
Previously uncharacterized in maize, the asc1 gene was selected for reverse genetic analysis because our microarray and in situ hybridization analyses revealed significantly upregulated expression in leaf primordia (Table S1; Figure 3F). Moreover, the related genes Zm-cycD4 and Zm-cycD2 are also contained on the SAM 3.0 gene chip used in these assays, although neither gene was identified as differentially-expressed. This failure to detect redundant, differential expression of D4-CYCLIN paralogs in the maize SAM suggested that potential mutant phenotypes conferred by asc1 mutations may not be masked by paralogous gene functions.
As shown in Figure 6, single mutations in asc1 condition infertile plants with extreme reductions in leaf width and plant height. Interestingly, these mutant phenotypes are more widespread and severe than those observed in Arabidopsis triple mutant plants homoyzygous for mutations in each of three D3-cyclin paralogs [64], which are phylogenetically distinct from the CD4/CD2 cyclins (Figure 5C). In addition, whereas mutations in each of the paralogous D4-cyclin Arabidopsis genes condition mild reductions in hypocotyl stomatal number [65], solo asc1 mutants exhibit profound abnormalities in leaf stomatal patterning (Figure 6). These asc1 mutant phenotypes suggest that ASC1 is required for normal maize leaf development, and that subfunctionalization of D-cyclin gene function has proceeded quite differently within the maize and Arabidopsis lineages. These data further demonstrate that laser microdissection-microarray analysis is a tractable approach toward the identification of important gene functions within adjacent yet distinct microdomains during maize shoot development.
Seedlings of the maize inbred B73 were raised in a growth chamber on a 15 hr light cycle. Samples were incubated at 25°C during the light cycle and 20°C during the dark cycle. Seedlings were harvested for dissection and fixation at 14 days after germination.
For use in shoot-apex culture, maize shoot apices were hand-dissected from 14-day-old seedlings to remove all except the four youngest leaf primordia as described [4]. Dissected apices were cultured on maize culture medium (MCM; described in [4]) containing 30 µmol N-1-naphthylphthalamic acid (NPA) dissolved in DMF, or in maize tissue culture media containing equal amounts of DMF but no NPA. Apices were incubated for 14 days on a 14 hour light cycle at 28°C (light period) or 24°C (dark period).
EST clone AW067338 was used to identify maize core gene AC196112.3_FG024 located at location 39,743–41,597 of contig 45 on chromosome 7 (Maize Genome Browser; http://www.maizesequence.org/index.html). As the second D4 cyclin gene characterized in maize, this locus was named leaf cyclinD42 (asc1). For reverse genetic analyses of asc1, DNA samples were prepared from pooled F2 seedling progeny obtained via self-pollination of 3,456 F1 plants containing active Mutator (Mu) transposon systems and subjected to PCR-based screens using nested asc1 gene-specific primers (lcd1-CTTGCATCCTCCACTTGAGC and lcd2-AGCAGCTGTGTCATCCAAGC) and a Mu specific primer (MuTIR-AGAGAAGCCAACGCCAWCGCCTCYATTTCGTC). To rule out false-positive results derived from multiple Mu insertions, control reactions were performed with the Mu primer only. PCR reactions with specific products only from the nested PCR amplifications were sequenced to verify the Mutator transposon insertion. Sibling seed from PCR-positive families were planted in a corn nursery in Aurora, NY, screened for developmental phenotypes and outcrossed for two generations to inbred B73. Interallelic crosses of plants heterozygous for independent Mu-insertion alleles of asc1 failed to complement, indicating the mutant phenotype observed in F2 progeny of self-pollinated plants harboring asc1-Mu insertion alleles are due to mutations in asc1.
Alignments were performed on protein sequences translated from ten Arabidopsis proteins AtCYCD1;1 (NM105689), AtCYCD2;1 (NM127815), AtCYCD3;1 (NM119579); AtCYCD3;2 (NM126126), AtCYCD3;3 (NM114867), AtCYCD4;1 (NM125940), AtCYCD4;2 (NM121082), AtCYCD5;1 (NM119926), AtCYCD6;1 (NM116565), AtCYCD7;1 (NM120289), a D1-CYCLIN from Physcomitrella patens (CAD32542), three CYCLINS from Antirrhinum majus including AmCYCD1 (AJ250396), AmCYCD3a (AJ250397) and AmCYCD3b (AJ250398), ASC1 and the maize CYCLINS ZmCYC2;1 (AF351189), ZmCYC4;1 (AF351191), ZmCYCD5;1 (AF351190) and ZmCYCD5;2 (AY954514). Sequences were aligned using CLUSTALX 2.0.4, and cladograms were generated with PAUP 4.0 using the Maximum Parsimony method and after treating gaps in the alignment as missing data. Equivalent cladograms were generated using the Neighbor Joining method and without removing gaps; bootstrapping values were calculated for 1,000 replicates.
Maize seedlings harvested at 14 days after germination were fixed in FAA, paraffin-embedded, sectioned at 10 µm, and stained in either Toluidine Blue O or Safranin-Fast Green using Johanssen's method as described [84]. Immunohistochemical analyses of KNOX protein accumulation were performed as described [13]. Epidermal images were obtained using cyanoacrylate glue surface impressions as described [85]. All micrographs were imaged on a Zeiss Z1-Apotome microscope (Thornwood, NY).
Seedling shoots were fixed by incubating in acetone, paraffin-embedded as described and sectioned at 10 µm as described [25],[26]. All laser-microdissections were performed using a P. A. L. M. Laser Microbeam (P.A.L.M. Microlaser Technologies, Bernried, Germany). SAM tissue domains were captured from 5–10 sections per sample, comprised of 0.3 mm2–2 mm2 of tissue. Six biological replicate samples were obtained. RNA was isolated from laser-microdissected tissue as described [25],[26]; RNA amplifications were performed using the RiboAmp™ HS kit (Arcturus, Mountainview, CA) according to the manufacturers protocol.
The SAM cDNA-enriched SAM1.1 and SAM3 microarrays used in these experiments were as described [26]. The MIAME guidelines utilized, hybridization protocols, and array scanning procedure were as described [25],[26]. All microarray data are available at Gene Expression Omnibus (GEO; http://www.ncbi.nlm.nih.gov/geo). Six biological replicate array hybridizations were performed. One of the six SAM1.1 slides was excluded from analysis due to poor hybridization quality and areas of very high background. Data from the other 11 slides were normalized within slides using loess normalization and across slides within each platform using scale normalization [86]. The limmaGUI R package [87] was used to conduct a linear model analysis for each gene following described methods [88]. The method of Benjamini and Hochberg [89] was used to estimate the false discovery rate associated with the identified sets of differentially expressed genes. Annotations of the predicted GO functions for all differentially expressed genes were performed as described [27]; annotated data is presented at SAM-The Maize Shoot Apical Meristem Project (http://sam.truman.edu/geneva/geneva.cgi).
EST sequences of the 1,312 microarray elements that were found to be differentially-expressed in the P0/P1 and SAM datasets were sorted by BLAST homology analyses to the contigs sequenced maize genomic DNA (i.e. maize contigs) in order to convert array elements (ESTs) into maize gene contig (MGC) groups. For multiple ESTs that hit a single MGC, the EST list was collapsed under the single MGC identity, and mean and standard error of mean (sem) were calculated for P-values and fold change of those ESTs. Sample size (n) value for mean and sem calculations is represented by the number of ESTs for a single MGC group. Multiple ESTs linked to a single MGC group were searched in both P0/P1 and SAM datasets to detect potential dual hits; such MGC groups containing P0/P1- and SAM-specific ESTs were removed from the dataset. To this end, 10 MGC groups representing 20 P0/P1- and 9 SAM-specific ESTs were removed. In some cases, a MGC could not be assigned for an EST, as indicated by the identifier ‘contig:NONE’. In other cases, single or multiple ESTs hit multiple MGCs (Table S1).
A test based on the binomial distribution was used to identify functional categories for which the discrepancy between the number of genes upregulated in SAM and the number upregulated in P0/P1 was significant. Conditioning on n = total number of genes identified in a given category, the null hypothesis that the proportion of genes upregulated in SAM was equal to 1/2 was tested against the alternative that the proportion was not 1/2. A p-value was obtained by comparing the observed number of genes upregulated in SAM to a binomial distribution with n trials and success probability 1/2. For example, given that 34 genes predicted to function in RNA binding were identified as differentially expressed, the probability that only 9 of these 34 would be up in SAM relative to P0/P1 is approximately 0.0045 according to a binomial distribution with 34 trials and success probability 1/2. This yields a two-sided p-value of 2*0.0045 = 0.009 and suggests that discrepancy (9 up in SAM vs. 25 up in P0/P1) cannot easily be explained by a simple chance mechanism.
qRT-PCR analyses of NPA-treated and untreated shoot apices (described in Plant Materials, above) were performed on cDNA prepared from tissue-cultured/laser-microdissected SAMs as described [55]. Analyses utilized three technical replications performed on pooled cDNA prepared from ten microdissected SAMs. Gene-specific primer pairs used in these analyses were as follows: AI855049 (5′-CAGAATCATCACCTACACCT-3 and 5′-GAGTAGTAGAAGATTGCTGTGAG-3′); DN220821(5′-GCTAATGAGCATAGTATGCC-3′ and 5′-CTGCTCATTACCATGTCCTG-3′); CD527823 (5′-TCCGTCTTGTACATGTGAG-3′ and 5′-TCTCGACATTCTTAAGGAGC-3′); CD670256 (5′-GGTCTCTAAAGTCACTGAAACC-3′ and 5′-GAGCTGATCCCTTAGTTAAGTC-3′); BG840831 (5′-GATCAAATCATAGACCTAGAGTCC-3′ and 5′-ATTGGTGTAGTTTCCTAGCTG-3′); AY313902 (5′ CCTCAAGAAGACCTTCAAGAC-3 and 5′-TTATTAGAATGGAGTGATGCCC-3′); CB380920 (5′-TCACCGTCAGAATTTACGTC-3′ and 5′-GCATAAACAACCACTGAACC-3′); CA998660 (5′- TTGAACTCATCCGCTTTCTC-3′ and 5′-TTGACACATTCCGTCTACAG-3′); BM073971 (5′-CCTCAAGGCATTCAGATCTC-3′ and 5′-AGATGATGTCTTCCTGTCGT-3′).
Fourteen-day-after-germination maize B73 seedlings were processed for in situ hybridization as described [90] with modifications [91]. Gene-specific probes were synthesized from the cDNA clones DN229322, BM073398, CB381076. BG840831, AW067338. For each gene-specific probe analyzed, at least six replicate samples were hybridized.
For use in RNA gel-blot hybridizations, total RNA was extracted from 14-day seedlings using the Trizol lysis method and prepared for Northern transfer as described [92]. The gene-specific asc1 probe was prepared from genomic DNA using the primer pair 5′-CGGTTTCCTGGAGTCTGAGG-3′ and 5′- CTTGCATCCTCCACTTGAGC-3′, which amplifies a 776 bp fragment spanning exons 1–4 (Figure 5A).
|
10.1371/journal.ppat.1005110 | Incomplete Neutralization and Deviation from Sigmoidal Neutralization Curves for HIV Broadly Neutralizing Monoclonal Antibodies | The broadly neutralizing HIV monoclonal antibodies (bnMAbs) PG9, PG16, PGT151, and PGT152 have been shown earlier to occasionally display an unusual virus neutralization profile with a non-sigmoidal slope and a plateau at <100% neutralization. In the current study, we were interested in determining the extent of non-sigmoidal slopes and plateaus at <100% for HIV bnMAbs more generally. Using both a 278 panel of pseudoviruses in a CD4 T-cell (U87.CCR5.CXCR4) assay and a panel of 117 viruses in the TZM-bl assay, we found that bnMAbs targeting many neutralizing epitopes of the spike had neutralization profiles for at least one virus that plateaued at <90%. Across both panels the bnMAbs targeting the V2 apex of Env and gp41 were most likely to show neutralization curves that plateaued <100%. Conversely, bnMAbs targeting the high-mannose patch epitopes were less likely to show such behavior. Two CD4 binding site (CD4bs) Abs also showed this behavior relatively infrequently. The phenomenon of incomplete neutralization was also observed in a large peripheral blood mononuclear cells (PBMC)-grown molecular virus clone panel derived from patient viral swarms. In addition, five bnMAbs were compared against an 18-virus panel of molecular clones produced in 293T cells and PBMCs and assayed in TZM-bl cells. Examples of plateaus <90% were seen with both types of virus production with no consistent patterns observed. In conclusion, incomplete neutralization and non-sigmoidal neutralization curves are possible for all HIV bnMAbs against a wide range of viruses produced and assayed in both cell lines and primary cells with implications for the use of antibodies in therapy and as tools for vaccine design.
| Antibodies that potently neutralize a broad range of circulating HIV strains have been described. These antibodies target a variety of sites on the envelope protein of HIV, three copies of which associate to form a trimer that decorate the membrane surface of the virus particle. Some of these antibodies target regions of the envelope protein close to the membrane, some bind to the top of the trimer, others bind via carbohydrates which cover the envelope protein and another subset binds to the same site as the human HIV receptor CD4. Despite effectively blocking 50% of infection at low antibody concentrations, for some particular virus/antibody combinations a proportion of virus particles are resistant to antibody neutralization, even at extremely high concentrations. This phenomenon is called incomplete neutralization and also frequently results in non-sigmoidal dose-response curves when antibody concentration is plotted against the level of virus infection. Previously, antibodies that target the apex of the trimer have been associated with incomplete neutralization and non-sigmoidal curves. In this study we show that representatives from all the groups of antibodies described above result in incomplete neutralization against at least one virus but that the phenomenon is more frequent for those binding the apex and the stalk of the trimer. Resistant populations of virus were seen whether the virus was produced in the natural target of HIV infection (human CD4+ T cells) or engineered human cells more commonly used to produce virus to test antibody function. Understanding this phenomenon is important for the future use of antibodies as therapeutics and for vaccine studies as a resistant population of viruses could result in failure to control the virus infection in patients.
| The HIV-1 envelope glycoprotein (Env) spike, the sole target of neutralizing antibodies (nAbs), is a heterotrimer of composition (gp120)3(gp41)3. The gp120 protein includes about 25 N-linked glycans that comprise almost 50% of its mass [1] and the gp41 protein typically includes four conserved N-linked glycans on the C-terminal half of the ectodomain [2]. While the virus uses glycans as a strategy to escape immune detection, there are several regions of Env that are well established as being vulnerable to broadly neutralizing antibody (bnAb) recognition [3–6]. Three regions are found on gp120: the CD4 binding site (CD4bs), an area of V2 at the apex of the Env spike that includes the glycan at N160 and an area involving V3 that includes glycans forming a high-mannose patch and most particularly a glycan at N332. Recent structural studies show that the V2 apex and high-mannose patch epitopes form a contiguous region at the top of the trimeric Env spike [7,8]. One region is found on gp41 close to the viral membrane and is known as the Membrane Proximal External Region (MPER). In addition, 3 new regions of vulnerability bridging gp120 and gp41 have recently been defined [9–12]. The recognition of each of these neutralizing epitopes by broadly neutralizing monoclonal antibodies (bnMAbs) has variable glycan dependence.
We have previously shown that certain bnMAbs display non-sigmoidal neutralization curves that plateau at <100% for some isolates, [13,14]. In these cases, the behavior has been shown to be partly due to glycan heterogeneity in the Env epitope. Therefore, it is worthwhile to briefly consider each of the bnAb epitopes in turn and, in particular, the role glycans play in each epitope before considering neutralization in more detail. The CD4bs is a conserved region on gp120 involved in receptor binding. The first antibody isolated recognizing this region, b12 [15], remains one of the best studied. However, it is less broad and potent than the more recently isolated CD4bs bnMAbs such as VRC01 [16], PGV04 [17], NIH-45-46, 12A12, 3BNC117 [18] and CH103 [19], which neutralize up to 55–90% of circulating viruses. Considering VRC01 and PGV04, crystal structures of the bnMAbs bound to the gp120 core reveal that both contact the glycan at N276 as part of their epitope but, whereas neutralization of PGV04 is dramatically decreased by its removal, neutralization by VRC01 is enhanced [16,17,20]. This glycan is also required for neutralization by another CD4bs bnMAb, HJ16 [21]. Apart from this glycan, no other glycan has been shown to be so strongly associated with neutralization of virus by CD4bs-targeting bnMAbs, although both VRC01 and PGV04 neutralization is enhanced by the removal of the glycan at N461, and furthermore removal of glycans at N301 or N386 can increase accessibility to the CD4bs [22].
PG9 and PG16 [13,23–25], PGT141-145 [26], CH01-04 [27], PGDM1400 [28] and CAP256-VRC26 [29], target a V2 quaternary site at the apex of Env that includes the glycan at N160. The crystal structure of PG9 bound to a V1/V2 scaffold revealed PG9 makes major contacts with this glycan [25], along with strand C of V1/V2 and the glycan at either N156 (CAP45) or N173 (ZM109). Subsequent studies showed that the PG9 epitope involves a Man5GlcNAc2 at N160 and high mannose-type or complex-type N-linked glycans at the secondary site (N156 or N173) [30]. Further structural studies with a stabilized Env trimer highlighted that only a single PG9 fragment antigen-binding (Fab) binds to each Env trimer [31]. In addition, molecular modeling and isothermal titration calorimetry studies suggested that PG9 can interact with the N160 glycan on an adjacent gp120 protomer within the antibody-trimer complex [31].
The third region of gp120 targeted by bnMAbs is a set of overlapping epitopes that all appear to involve the glycan at N332 as a significant contributor to antibody binding. Some of the most potent bnMAbs to HIV, PGT121-123, PGT125-7, PGT128, and 10–1074 are amongst those targeting this region as well as 2G12, PGT130-131 and PGT135-137. 2G12 binds the terminal mannose of glycans at N295, N332, N339, and N392. PGT121-123 bind gp120 V1, V3 and several surrounding glycans [7]. The central glycan is the Man8/9GlcNAc2 glycan at residue N332 but complex or hybrid glycans at N137, N156 and N301 are also involved [7,22]. PGT128 binds glycans at N332, N301 and protein sequence at the C-terminal V3 stem [32]. PGT135 binds glycans at N332, N392, and N386, and protein sequence in V3 and V4 [33].
BnMAbs 4E10, 2F5, Z13e1 and 10E8 recognize the MPER on gp41 [34–36]. While these bnMAbs are not known to make direct contacts with glycans, partial deglycosylation of gp140 has been shown to enhance binding of 4E10 and 2F5 [37]. In contrast, the recently described anti-gp41/gp120 bnMAbs PGT151 and PGT152 have been shown to bind to tri and tetra-antennary complex-type glycans at N611 and N637. These glycan interactions are essential for neutralization as the removal of both completely abrogated neutralization for all isolates tested [9]. The newly identified bnMAb 35O22 has been shown to target a site spanning gp41 and gp120 using glycans N88, N230 and N241 [10]. In addition, a previously described bnMAb 8ANC195 [18] was recently found to bind Env via contacts in gp120 and gp41 within a single protomer [11].
Several bnMAbs have been particularly associated with incomplete neutralization. Notably, the PGT150 series bnMAbs do not always achieve 100% neutralization even when able to potently neutralize the virus under investigation with an IC50 value of less than 1μg/ml. Maximum neutralization ranges from 50–100% with PGT151 and PGT152 achieving less than 80% neutralization against 26% and 20% of 117 viruses tested. Similarly, PG16, and to a lesser degree PG9, have also been shown earlier to display neutralization curves, for some isolates, that plateau at <100% and are non-sigmoidal [13,14]. Incomplete neutralization has also been described for the potent MPER bnMAb 10E8 and this has been associated in part with glycan heterogeneity, particularly at N625 [38].
Incomplete neutralization may have serious consequences for the ability of nAbs to protect against HIV exposure. If transmitted viruses are heterogeneous with respect to neutralization sensitivity, as observed for pseudoviruses and primary viruses grown in PBMCs, then the failure of an antibody to neutralize a population of viruses completely could lead to the establishment of an infection. However, it should be noted that incomplete neutralization has not been reported for serum samples from elite neutralizer donors. In the current study, we were interested in investigating the extent of non-sigmoidal slopes that can plateau at <100% for HIV bnMAbs targeting the four best-studied regions of the Env spike involved in broad neutralization. We found all bnMAbs to be subject to inhibition curves that do not reach 100%, and most bnMAbs also had neutralization curves with non-sigmoidal slopes but bnMAbs targeting certain areas of the trimer were more prone to these effects. Primary isolate and molecular cloned viruses produced in primary cells were also incompletely neutralized, suggesting this phenomenon could be significant in vivo.
A panel of bnMAbs targeting different regions on Env was tested in a highly quantitative pseudovirus neutralization assay on a panel of 278 viral clones (Fig A in S1 Text). Pseudovirus entry into U87 cells expressing CXCR4 or CCR5 was measured by luciferase activity. Viruses that were neutralized with an IC50 >1μg/ml were not further considered because we could not exclude the possibility that at/above this concentration the IC50 would be too high for 100% neutralization to be reached by a typical sigmoidal curve at the highest antibody concentration measured (50 μg/ml). For each bnMAb, we determined the maximum percent neutralization (MPN), i.e. the percent at which the neutralization curve plateaus for those viruses neutralized with an IC50 <1μg/ml. For the panel of bnMAbs tested, b12, 2G12, PGT136 and PGT137 had a relatively low number of viruses neutralized with an IC50 <1μg/ml (Table 1), which may create a small sample bias in the ranking of these two bnMAbs.
A neutralization curve that plateaus at less than 100% indicates there is a fraction of neutralization resistant viruses that likely express a proportion of Env spikes that are not recognized by the corresponding bnMAb. In Fig 1A, the bnMAbs are ordered from left to right according to decreasing median of the MPN (Fig A in S1 Text). None of the bnMAbs had median MPN values that were <90%, but all of the bnMAbs neutralized at least one individual virus with <90% MPN, except PGT122, PGT127 and PGT136 (Fig 1A). As a group, the bnMAbs that target the high-mannose patch: PGT121-123, PGT125-131, PGT135-137 and 2G12 (Fig 1A, magenta data points) had the highest median MPN values. Of these bnMAbs, PGT121-123 had the highest median MPN, neutralizing viruses to 100% on average, while the other high-mannose patch targeting bnMAbs had median MPN values of 99%. Of the bnMAbs that target the high-mannose patch, 2G12 had the lowest median MPN at 98%. However, 2G12 is an unusual bnMAb because its epitope comprises solely glycans and it is thus more dependent on several glycan sites than the other high-mannose patch bnMAbs [39–41]. PGT130 and 135 also showed less complete neutralization than other high-mannose patch bnMAbs, their median MPN values were equivalent but MPN values as low as 79 and 83% were observed for individual viruses.
The CD4bs bnMAbs also had high median MPN values of 100% and 99% for PGV04 and b12 respectively. These results show that these two CD4bs targeting bnMAbs largely display complete neutralization and, if glycan heterogeneity is the source of incomplete neutralization, this may reflect a relative insensitivity to Env glycan expression, as has been described earlier for bnMAb b12 [14]. However, it should be noted that b12 neutralizes a relatively small proportion of the virus panel, which may have skewed its median MPN value. These data do not allow us to unequivocally conclude that CD4bs bnMAbs as a class are less prone to incomplete neutralization; this would require the evaluation of a larger panel of CD4bs bnMAbs.
Of the bnMAbs targeting the V2 apex (Fig 1A, blue data points), PG9 leads the group with a median MPN value of 99%, although as seen for PGT130 and 135, many individual viruses resulted in MPN values as low as 72%, while median MPN values for PGT 141–145 and PG16 range from 95 to 98%. We have previously reported that, for PG9 and PG16, incomplete neutralization can be attributed, in part, to glycan heterogeneity. JR-FL E168K is an example of a pseudovirus that displays PG9 and PG16 <100% neutralization plateau curves. Notably, PG9 and PG16 neutralization of JR-FL E168K produced in the presence of the glycosidase inhibitor swainsonine reverts the shape of the inhibition curve to a standard curve with a maximum inhibition ∼100% [14].
The gp41 targeting bnMAbs showed the lowest median MPN values (Fig 1A, green data points) at 93% for 4E10 and 96% for 2F5. The inability of 4E10 and 2F5 to reach 100% neutralization for their neutralization curves is, at first glance, inconsistent with the association of glycan heterogeneity with incomplete neutralization since these bnMAbs do not target a glycan as part of their known epitopes. In fact, 2F5 and 4E10 neutralization of the HIV-1 CRF07_BC pseudovirus named FE, increased when single glycans were removed at N197, N301, N355 (gp120) and N625 (gp41) [42]. In addition, deglycosylation of gp140 Env oligomers with PNGase F increased 2F5 and 4E10 binding to the protein [37]. These results could be explained if glycans on gp120 and/or on gp41 affect the accessibility of the MPER in the context of the viral membrane.
Notably, the median MPN value for each bnMAb does not fully describe its propensity for incomplete neutralization. Another way to assess the efficiency of neutralization is to consider the interquartile range of the individual Ab-virus MPN values from which the median arises. Across all bnMAbs tested, regardless of epitope, there was a general trend of increasing interquartile range as the median of MPN decreased (as shown in Fig 1A), in which the interquartile range of each MPN is defined as the scatter of the population and delineates where the middle 50% of the population is located (whiskered bars). PG16, PGT144 and 4E10 showed the largest scatter, indicating these bnMAbs showed the most variability in their capacity to bind heterogeneous envelopes of different viruses.
The percentage of total viruses neutralized at different maximum levels was determined for all viruses that resulted in an IC50 <1μg/ml (Fig 1B and Fig B in S1 Text) and the data suggested a natural division of the bnMAbs into three groups. The bnMAbs in Group 1 neutralized >90% of susceptible viruses with MPN values of >95% and include PGT121-123, PGT125-28, PGT136 and PGV04. Group 2 bnMAbs were less effective, resulting in neutralization of between 60–84% of susceptible viruses with MPN values of >95%. Group 2 bnMAbs, include b12, PGT125-127, 130–131, 135,137, 141–143, 145, 2G12 and PG9. The Group 3 bnMAbs, PG16, PGT144, 2F5 and 4E10, were the least able to completely neutralize viruses for which they have an IC50 of <1μg/ml. For this group only 36–60% of susceptible viruses were neutralized with MPN values of >95%. The percent of viruses neutralized at >98% was generally relatively similar for somatic variants of a given family although the neutralization activity of these variants against individual viruses can differ substantially[13,26]. PGT128 and PGT145, the broadest and most potent somatic variants of their respective families neutralized the most viruses with MPN values of >98% when compared to their sister clones. In contrast, PGT136, the least broad and potent of the PGT135 family, neutralized the largest proportion of viruses to >98% inhibition when compared to its sister clones. However, due to its lower potency, only a small number of viruses were neutralized by PGT136 with an IC50 <1ug/ml so this analysis considered only a restricted subset of viruses, which may explain the relative efficiency of PGT136 at producing complete neutralization.
While incomplete neutralization suggests the possibility of a mixed population of wild-type envelopes, a percentage of which are not neutralized by a given antibody leading to resistance of a fraction of virions, a non-sigmoidal neutralization curve with a plateau <100% suggests populations of viruses with varying sensitivity to antibody neutralization. A perfect sigmoidal dose-response curve has a slope of 1. The slope of a neutralization curve for a bnMAb that binds all functional Env spikes on a population of wild-type viruses equally well, and neutralizes the corresponding viruses equally well should be close to 1 (Fig 2A and Fig C in S1 Text). A bnMAb that neutralizes a subset of viruses in a wild-type population of viruses less effectively than other viruses in that population may cause a neutralization curve that plateaus at <100% and will have a slope <1 (Fig 2A). A steeper curve with a slope >1 (Fig 2A), suggests either irreversibility of neutralization, cooperativity in binding to trimers or a favorable kinetic neutralization profile which then affects the rate of neutralization of the corresponding viruses. We determined the median curve slopes for the bnMAbs (Fig 2B and Fig D in S1 Text) and found the same general trend was seen for the median slopes as was seen for the median MPN values, and the two had a moderate correlation (Spearman r = 0.745 and a P-value of <0.0001) (Fig 2C and Fig E and F in S1 Text). Again, the high-mannose patch- and CD4bs-targeting bnMAbs conformed most to ideal behavior with median slopes ~1, and V2 apex- and MPER-targeting bnMAbs had median slopes that deviated the most from 1. An exception to this was PG9, which generally had sigmoidal neutralization curves like the high-mannose patch and CD4bs targeting bnMAbs. The median slope of 4E10, 0.6, was the furthest away from a standard slope. The V2 apex-targeting bnMAbs ranged from 0.6 to 0.9. The high-mannose patch bnMAbs ranged from 0.8 to 1.1. None of the bnMAbs had median curve slopes <0.6 and all the bnMAbs had at least one virus for which the slope was <0.72. PGT130 and PGT131 had the largest spread in slope behavior suggesting that neutralization by these bnMAbs was the most sensitive to the isolate context of Env heterogeneity. The median MPN values from each bnMAb-virus pair were also stratified according to viral clade to assess whether certain groups of viral strains are incompletely neutralized by all bnMAbs (Fig J in S1 Text). No clear clade-dependence was observed although some individual viruses are less completely neutralized by all bnMAbs.
To investigate a large number of viruses and reliably determine whether complete neutralization had been achieved, the study reported above used a high throughput quantitative pseudovirus assay system and U87 cells expressing CXCR4 or CCR5 as the target cells. However, many more recently described bnMAbs (including the PGT151 family) have been characterized with a TZM-bl cell based assay using a different 117-pseudovirus panel. Therefore, we determined the MPN of bnMAbs using this system. In Fig 3A, the bnMAbs are ordered from left to right according to decreasing median MPN as shown for the U87 cell assay in Fig 1A. The overall level of incomplete neutralization was similar to that seen for the U87 cell assay with median MPN values of the bnMAbs ranging between 94 and 100% and all of the bnMAbs neutralizing at least one virus with an MPN of <90% (Fig G in S1 Text). Furthermore, as previously described for the U87 cell assay, most of the bnMAbs that bind the high-mannose patch resulted in the most complete neutralization across the 117-virus panel with median MPN values close to 100% and the majority of viruses neutralized to >90%. Again, as seen in the U87 assay, when the bnMAbs were divided into groups based on their median MPN of susceptible viruses (Fig 3B and Fig H in S1 Text). PGT121 and 123 were among the bnMAbs with the most efficient neutralization, with most viruses neutralized to between 95–100% (Fig 3B), although some outliers had lower MPN values (Fig 3A). However, individual viruses were neutralized by certain bnMAbs to a much lesser degree (e.g. 56%, 60%, and 75% for PGT135, 128 and 123 respectively) despite an IC50 <1μg/ml. Of note, bnMAbs PGT130 and 135 were less able than other high-mannose patch bnAbs to produce high levels of virus neutralization in agreement with the U87 cell assay results shown in Fig 1A. PGT130 resulted in MPN values >90% for many viruses (Fig 3B) but also resulted in a substantial number of less completely neutralized viruses (Fig 3A). PGT135 on the other hand resulted in a disparate pattern of MPN with a very wide range of values (Fig 3A). It should be noted that PGT135 neutralized a lower number of viruses with an IC50 <1μg/ml in this 117-virus panel than any of the other bnMAbs except PGT144.
Together with the high-mannose patch bnAbs, the CD4bs antibodies showed a very high tendency to complete neutralization against the 278-virus panel in the U87 cell assay. In line with this, the CD4bs antibody 12A12 resulted in a MPN values of >90% of all but three viruses in the TZM-bl assay (Fig 3A), with a median MPN value of 100%. In contrast, another CD4bs antibody PGV04 resulted in notably less complete neutralization. While the median MPN values for this bnMAb in both assays was close to 100%, more individual viruses were incompletely neutralized in the TZM-bl than in the U87 cell assay (Fig 3A).
The V2 apex bnMAbs were shown in the U87 cell system to have a greater propensity to result in incomplete neutralization than the CD4bs or high-mannose patch binding bnMAbs. Similarly, in the TZM-bl assay PG9 and PGT143-144 resulted in a greater spread of MPN values than PGT121, 123, 125, 127, 128 and 12A12, with many viruses showing between 90–95% neutralization and a number that were neutralized to only between 75–90% (Fig 3B). However, the level of incomplete neutralization by PG9 and PGT143-144 is comparable in this assay to that seen with PGT130, 135 and PGV04 (Fig 3A).
Finally, the two related specificities PGT151 and PGT152, which target a newly described cleavage-dependent gp120/gp41 glycan epitope [9], showed the least complete neutralization with a wide range of MPN values from 57–100% and 51–100% respectively (Fig 3A). Both showed a considerable proportion of viruses with MPN values <90% (Fig 3B). This is consistent with previous findings for these bnMAbs [9].
Although the U87 and TZM-bl studies reported here use different target cells and different virus panels, the trends observed for the bnMAbs tested in both assays are similar (compare Figs 1A and 3A). A direct comparison of results is provided for the bnMAbs PGT121, PG9 and PGV04 in Fig 3C (Fig I in S1 Text). PGT121, which showed among the most complete neutralization in both assays relative to other bnMAbs results in a somewhat greater range of MPN values in the U87-cell/Monogram panel assay than in the TZM-bl/CAVD panel assay. Similarly, PG9, which results in less complete neutralization than PGT121 in both assays, showed a somewhat greater range of values in the U87 cell assay. In contrast, PGV04 results in more complete neutralization in the U87 assay and a smaller range of values as compared to the TZM-bl assay. However, the U87-cell/monogram assay involved 278 viruses and the TZM-bl/CAVD assay used 117. This difference in number of viruses and the divergent viruses used suggest some caution in some of the more detailed results from the two assays.
Previously a subset of bnMAbs targeting the high-mannose patch, V2 apex and the CD4 binding site were characterized for neutralization of molecular clones of virus in a PBMC neutralization assay [43,44]. The clones were isolated from subtype B viral swarms from patients who participated in the Amsterdam Cohort Studies on HIV Infection and AIDS and produced in PBMCs. Here we examine the subset of bnMAbs described above for the ability to neutralize patient-derived virus to 100%. The greater variability inherent in the PBMC neutralization assay renders impossible the definition of MPN with the high degree of precision achievable in the pseudovirus assays and makes reproducibility difficult. Furthermore, for replication competent virus there are multiple rounds of infection during the neutralization assay as compared to a pseudovirus assay where no new viral progeny can be produced. However, with these caveats, we assayed the ability of bnMAbs to neutralize viruses in the PBMC assay (Fig 4A and Fig K in S1 Text). Regardless of epitope, each bnMAb showed a wide variety of MPN values for different viruses. PGT121 produced the least incomplete neutralization with a median MPN of 90% but with values ranging from 64 to 100%. The CD4 binding site antibodies VRC01 and NIH45-46 54W had median MPN values of 74% and 87% respectively. Both PGT121 and NIH45-46 54W, the bnMAbs with the most overall complete neutralization, showed a large group of viruses that were well neutralized (MPN>85%) and a smaller group of viruses forming a “tail” in Fig 4A of dots below 85%. The other bnMAbs neutralized more viruses inefficiently (MPN <85%) and as a result there are more dots in Fig 4A below 85% resulting in a more even distribution of MPN values for these bnMAbs in comparison to PGT121 and NIH45-46W (Fig 4A).
Given our previous work [9,14] describing the influence of glycan heterogeneity on incomplete neutralization by PG9 and PG16, we hypothesized the production of virus in the 293T cell line or primary PBMCs could alter the plateau of a bnMAb neutralization curve against a given virus. One obstacle to investigating this hypothesis is that most of the viruses used in the neutralization panels described in Figs 1–3 are pseudoviruses and therefore cannot be grown in primary cells. Therefore, to study the effects of producer cell type on incomplete neutralization, we used an 18-virus panel of molecular clones, which, unlike the pseudoviruses (Figs 1, 2 and 3) or the patient-derived panels (Fig 4), can be produced either in 293T cells or PBMC cultures and assayed in the TZM-bl neutralization assay. PGT121 was compared to PG9, PG16, PGDM1400 and PGT151 (Fig 4B). The overall trends observed were similar for viruses grown in 293T cells or PBMCs although somewhat greater incomplete neutralization was observed for PBMC-grown viruses. For PGT121, the median MPN was 100% for viruses grown in 293T and 96% for viruses grown in PBMCs; for PG9 the corresponding values were 98% and 88%. Values for the other bnMAbs are summarized in Fig L in S1 Text. As for the large panel of PBMC-grown viruses, overall the median MPN values were lower for PBMC-grown virus than 293T cell-grown virus for the molecular clone panel, although for some individual viruses the opposite was noted i.e. greater MPN values were observed for PBMC-grown molecular clones than the 293T cell-grown virus (Fig M in S1 Text).
We have previously shown that, for a minority of isolates, both PG9 and PG16 have neutralization curves that reach <100% and non-sigmoidal slopes (Walker, 2009) [14]. In the current study, we were interested in determining more on the extent of this phenomenon and therefore we analyzed the neutralization profiles of bnMAbs targeting four major regions of the Env using a large panel of pseudoviruses in a quantitative high-throughput U87 neutralization assay. We found that bnMAbs that target epitopes in the region incorporating the high-mannose patch and two CD4bs bnMAbs neutralized the majority of viruses with MPN values close to 100%. In contrast, bnMAbs targeting the V2 apex region and the MPER more often neutralized viruses with MPN values of <100%. Frequently, neutralization to less than about 95% was associated with non-sigmoidal neutralization curves (Fig 2). The high-mannose patch and CD4bs targeting bnMAbs most frequently showed standard neutralization curves and the V2 apex and MPER targeting bnMAbs showed a higher frequency of non-sigmoidal curves. However, it should be noted that all of the bnMAbs showed less than 95% neutralization and non-sigmoidal neutralization curves for some isolates studied.
To extend these findings, the neutralization profiles of bnMAbs were evaluated with a large panel of pseudoviruses in the TZM-bl assay (Fig 3). Again, the PGT121 and 128 families targeting the high-mannose patch neutralized the majority of viruses with an MPN value close to 100%. The CD4bs bnMAbs 12A12 and PGV04 also neutralized many viruses with MPN values approaching 100%, although PGV04 also resulted in incomplete neutralization with MPN <90% for a sizeable fraction of viruses. Similarly, the V2 apex binding bnMAbs PG9, PGT143-145 more often neutralized viruses with a suboptimal MPN value of <90%. Notably, this study included the addition of two newly described glycan-dependent gp120/gp41 bnMAbs, PGT151-152, which exhibited variable MPN values with a substantial proportion of viruses neutralized to less than 90%. Overall, the results from the two neutralization assays were similar with some indication of a greater tendency to incomplete neutralization in the U87 cell assay than the TZM-bl assay.
The mechanism(s) of incomplete neutralization and non-sigmoidal neutralization curves were beyond the scope of the present investigation but an earlier study on the bnMAbs PG9 and PG16 showed the importance of glycosylation heterogeneity for these features [14], and this heterogeneity correlated with the critical involvement of glycans in the epitopes bound by PG9 and PG16 [7,25,31]. Incomplete neutralization and non-sigmoidal curves are shown here to occur also for other bnMAbs targeting the V2 apex and notably for gp41 bnMAbs. Of these, the PGT151 family depends on tri and tetra-antennary complex-type glycans at N611 and N637 for neutralization function [9], and resulted in the least complete neutralization in the TZM-bl assay, with MPN values as low as 57 and 51% for PGT151 and 152 respectively. In contrast, the MPER Abs have not generally been associated with glycan dependence of binding but this has not been thoroughly investigated [42] and glycan status can influence the accessibility of the MPER to Abs following CD4 engagement, which is when MPER Abs are thought to act [45]. Furthermore, incomplete neutralization by bnMAb 10E8 has been associated in part with glycan heterogeneity [38]. The high MPN values for the bnMAbs targeting the high-mannose patch at first seems inconsistent with their reliance on glycan binding and previous findings showing glycan heterogeneity can result in incomplete neutralization by apex-specific bnMAbs. However, structural studies have shown that PGT128 contacts the N332 glycan on an engineered gp120 outer domain by interacting with the terminal mannose residues of the D1 and D3 arms [32]. This mannose presentation only exists on Man8/9GlcNAc2 glycans, which suggests that the glycan at N332 may be expressed as relatively homogenous Man8GlcNAc2 and/or Man9GlcNAc2 in pseudoviruses [32] and gp120 Env protein [46]. Similarly, PGT122 in complex with a stabilized Env gp140 trimer shows the central N332 glycan is again Man8GlcNAc2. This homogeneity may be a significant contributor to the ability of bnMAbs that are dependent on the glycan on N332 to bind a high proportion of the envelopes of an isolate effectively and therefore neutralize the vast majority of virions. Furthermore, some bnMAbs recognizing this region can tolerate some heterogeneity in glycan usage [44] and/or use alternate glycan sites [28] that may facilitate tolerance of glycan heterogeneity. Notably, the number of glycan sites on which the high-mannose patch bnMAbs depend for neutralization varies. Consequently those bnMAbs that require more glycan sites such as PGT135/136 had lower median MPN values. The CD4bs bnMAbs are primarily sensitive to a glycan at N276 [20,21] and heterogeneity at this position is a candidate for incomplete neutralization of a virus population. It should be emphasized that there may be mechanisms other than glycan heterogeneity for incomplete neutralization and non-sigmoidal neutralization curves that have yet to be understood; an example would be conformational heterogeneity of the Env trimer [47,48].
While the incomplete neutralization by HIV bnMAbs has herein been clearly demonstrated, it is important to establish how incomplete neutralization in pseudovirus assays relates to anti-viral activity of bnMAbs in vivo. As a first step, we investigated the effects on primary viruses in vitro by study of the incidence of incomplete neutralization in a large panel of molecular cloned viruses generated from patient swarms from the Amsterdam Cohort Studies on HIV Infection and AIDS. These clones were propagated and assayed in PBMC cultures, which increases the level of variability inherent in the assay and decreases the confidence with which we can define complete neutralization. Furthermore, Abs were titrated from a lower concentration than in the other assay systems which means, although a 1μg/ml cut-off was used as previously described, some Abs may not have reached saturating concentrations in this assay. However, among a large primary virus panel and the five bnMAbs tested there were clear examples of both incomplete neutralization (<90%) and complete neutralization (100%) (Fig 4). Thus, we can conclude that this phenomenon is not an artifact of using 293T cell-produced pseudoviruses for neutralization assays. These data did not directly address whether virus produced in 293T cells or primary cells is more or less susceptible to incomplete neutralization by bnMAbs. To address this we tested a 18-virus panel of molecular clones produced in both cell types and showed that production in PBMCs does not abrogate the incomplete neutralization phenotype (Fig 4 and Fig M in S1 Text) on an individual virus basis. However, the median MPN values for each bnMAb across the 18-virus panel were lower for the PBMC-grown viruses than those for the 293T cell-grown clones. That incomplete neutralization was more frequently observed using PBMC-produced virus on average is a potentially important limitation to the use of 293T cell-grown pseudoviruses for characterizing the likely in vivo efficacy of anti-HIV bnMAbs.
As described, a key question for further study is how incomplete neutralization impacts the ability of bnMAbs to mediate protective effects. Protection has been clearly demonstrated in macaques via passive transfer of many bnMAbs when the challenge virus was completely neutralized [49–55] and as beneficial treatment for pre-existing infection in macaques and humanized mice [56–60]. No systematic studies have been carried out on protection when the challenge virus shows incomplete neutralization by the bnMAb under investigation. However, a recent study found that PG9 provided adequate and typical protection against a challenge virus that showed incomplete neutralization in a TZM-bl assay but essentially complete neutralization in a PBMC assay [61]. Further protection studies will be needed to firmly establish the significance of incomplete neutralization for prophylaxis by bnMAbs and in considering vaccine strategies.
Human blood samples from healthy donors were obtained form The Normal Blood Donor service at The Scripps Research Institute. The collection of human blood samples for isolation of PBMCs and subsequent propagation of HIV-1 was approved by the institutional Review Board at The Scripps Research Institute (protocol number HSC-06-4604), all samples were analyzed anonymously. The Amsterdam Cohort Studies on HIV infection and AIDS (ACS) are conducted in accordance with the ethical principles set out in the declaration of Helsinki, and written informed consent was obtained prior to data collection. The study was approved by the Academic Medical Center’s Institutional Medical Ethics Committee.
2G12, 4E10 and 2F5 (Polymun Scientific, Vienna, Austria) were procured by the IAVI Neutralizing Antibody Consortium. The recombinant bnMAb IgG1 b12 was expressed by the Center for Antibody Development and Production (The Scripps Research Institute, La Jolla, CA) in Chinese hamster ovary (CHO-K1) cells, purified using affinity chromatography (GammaBind G Sepharose, GE Healthcare) and the purity and integrity was checked by SDS-PAGE. PGT 121–123, PGT 125–131, PGT 135–137, PGT 141–145, PGV04, PDGM1400, PGT151-153 and PG9/16, were transiently expressed with the FreeStyle 293 Expression System (Invitrogen). Antibodies were purified using affinity chromatography (Protein A Sepharose Fast Flow, GE Healthcare) and the purity and integrity was checked by SDS–PAGE.
Monogram Biosciences performed neutralization assays using pseudovirus that undergoes a single round of replication as previously described [62]. Briefly, pseudoviruses capable of a single round of infection were produced by co-transfection of HEK293 cells with a subgenomic plasmid, pHIV-1lucu3, that incorporated a firefly luciferase gene and a second plasmid, pCXAS, which expressed an HIV-1 Env clone. Pseudoviruses were harvested 3-days post-transfection and used to infect a U87 cell line expressing co-receptors CCR5 or CXCR4. Pseudovirus neutralization assay using TZM-bl target cells was previously described [63]. For neutralization assays using PBMCs: human PBMCs were obtained from healthy individuals, isolated and stimulated as previously described [44,64]. HIV-1 infectious clone virus stocks were grown and titered on CD8+-depleted PBMCS [65]. Virus production was monitored by p24 ELISA (Aalto Bioreagents, Dublin, Eire). For PBMC neutralization assay as previously described [44].
Statistical analyses were done with Prism 5.0c for Mac (GraphPad).
|
10.1371/journal.pcbi.1000516 | Statistical Use of Argonaute Expression and RISC Assembly in microRNA Target Identification | MicroRNAs (miRNAs) posttranscriptionally regulate targeted messenger RNAs (mRNAs) by inducing cleavage or otherwise repressing their translation. We address the problem of detecting m/miRNA targeting relationships in homo sapiens from microarray data by developing statistical models that are motivated by the biological mechanisms used by miRNAs. The focus of our modeling is the construction, activity, and mediation of RNA-induced silencing complexes (RISCs) competent for targeted mRNA cleavage. We demonstrate that regression models accommodating RISC abundance and controlling for other mediating factors fit the expression profiles of known target pairs substantially better than models based on m/miRNA expressions alone, and lead to verifications of computational target pair predictions that are more sensitive than those based on marginal expression levels. Because our models are fully independent of exogenous results from sequence-based computational methods, they are appropriate for use as either a primary or secondary source of information regarding m/miRNA target pair relationships, especially in conjunction with high-throughput expression studies.
| MicroRNAs are a family of small RNAs that play important roles in the development, physiological function and stress responses of a wide variety of organisms, and if abnormally expressed are associated with multiple types of cancer in humans. Rather than being translated into proteins, members of the family of microRNAs operate by preventing the translation of messenger RNAs to which they have some degree of sequence complementarity. Although sequence-based bioinformatics techniques have yielded large numbers of predicted messenger- and microRNA targeting relationships, verifying these as bona fide has proven practically difficult. We have developed a novel statistical approach based on the system biology of microRNAs in humans to detect such targeting relationships using high-throughput RNA expression data. Because our approach is not based on information from external target pair predictions, it can play a fully independent role in verifying such predictions as well as be used to obtain de novo target pair predictions. Using two separate data studies, we show that our approach is capable of both reproducing previously observed target pairs and verifying putative target pairs predicted from sequence data, at rates substantially better than marginal comparisons of messenger- and microRNA expression levels.
| Micro RNAs (miRNAs) are small (20–22 bp) RNAs transcribed by a wide variety of organisms, from viruses [1], to plants [2],[3], to animals such as C. elegans, Drosophila and humans [4]–[6]. While most RNAs function in ribosomes or splicesomes, or are translated into proteins necessary for cellular function, miRNAs instead serve as negative regulators of gene expression by preventing the translation of messenger RNAs (mRNAs). Through their regulatory activities, miRNAs have been shown to affect organismal development, physiological function and stress responses. Abnormal miRNA production has also been associated with the development of several types of cancer [7]–[10].
Posttranscriptional gene silencing through miRNA activity occurs through a multistep process (Figure 1) [11]–[17] with an overall structure that has been remarkably conserved across organisms. This process begins with primary miRNA transcripts (pri-miRNAs) being either transcribed from “miRNA genes” or spliced from the intronic regions of mRNAs. In the nucleus, pri-miRNAs fold into hairpin structures from which trailing 3′ and 5′ ends are cleaved away by the RNase Drosha. The resulting precursors to mature miRNAs (pre-miRNAs) are then exported from the nucleus to the cytoplasm, where a second RNase enzyme (Dicer) removes the hairpin loop. This produces a segment of double stranded RNA that is separated into two single strands by helicase enzymes. After separation, one of the single stranded RNAs is combined with an Argonaute (Ago) protein to form an RNA-induced silencing complex (RISC). (Although other proteins may be incorporated into the structure, an Ago protein and miRNA compose a minimal functional RISC [18],[19].) Once assembled, RISCs composed of a given miRNA interfere with the translation of select mRNAs by hybridizing to them at target sites complementary to the miRNA sequence and either cleaving the mRNA or blocking its translation while leaving the molecule intact. Any mRNA translationally regulated by a particular miRNA can be anticipated to have a limited number of target sites usable by that miRNA. Each miRNA can target multiple mRNAs, and an mRNA may contain target sites for multiple miRNAs.
While both mRNA cleavage and blocking ribosomal activity disrupt translation, the latter does not directly alter mRNA abundance. Whether a particular RISC cleaves or blocks translation is determined by both the qualities of the hybridization and properties of the Ago protein contained in the RISC. The number and function of distinct Ago proteins shows substantial variability across organisms. For example, in Arabidopsis there are 10 different variants of Ago and miRNAs preferentially associate with only one in forming RISCs [20],[21], while in humans there are 4 commonly coexpressed Ago proteins, and miRNAs can be effectively regarded to have equal propensity to combine with each [22],[23]. The variant of Ago primarily utilized by miRNAs in Arabidopsis is competent for target cleavage [21], which is consistent with previous observations that the dominant means of miRNA-based regulation of mRNA translation is cleavage rather than translational repression. In humans, only RISCs composed of Ago 2 have been demonstrated to have the ability to cleave and degrade targeted mRNAs [22],[24]. Since miRNAs have equal propensity to combine with each of these, it is reasonable to conclude that targeted mRNAs are repressed through a combination of both cleavage and ribosomal blockage. This is consistent with results described by Nakamoto et al [25] which demonstrate simultaneous increases in both target mRNA and polyribosomal fraction in human miRNA knockdown studies, and recent experiments reported by Bartel et al [26] that suggest in mice (which share many of the complexities found in human Ago properties and RISC formation), most mRNA targets of miRNA-mediated repression are cleaved.
To determine whether a miRNA targets a particular mRNA, sequence-based computational target prediction methods may be used to identify potential miRNA hybridization sites within that mRNA [27]–[34]. Algorithms such as miRanda [35] use m/miRNA alignments and hybridization energies as metrics to score mRNA subsequences, and report high-scoring subsequences as putative target sites. More recently proposed methods additionally utilize evolutionary conservation of a predicted site across multiple organisms (PicTar [36]), information regarding target site position and base content (TargetScan [37]–[39]), or mRNA secondary structure [40] to improve prediction performance. Although existing computational target prediction algorithms provide important information regarding potential m/miRNA target pairings, they are acknowledged to have issues with specificity and sensitivity [29],[30] as well as inter-algorithm consistency [30],[34]. (These issues are discussed in relation to this study in the Methods and Discussion sections.) The problem of how to reliably predict target pair relationships from sequence data alone is currently unresolved.
With the limitations of purely sequence-based methods of miRNA target prediction, it has been suggested that the statistical analysis of expression data may play an important role not only in verifying computationally predicted m/miRNA targeting relationships, but also for generating de novo target pair predictions [27],[29]. Such analysis would require that both mRNA and miRNA abundance be measured on the same tissue samples, and naturally would consider the marginal correlation between a miRNA and its putative target. Marginal approaches are attractive because they are simple and they aim to capture the fundamental negative relationship between miRNAs and their targets. However, determining reliable and replicable targeting relationships through marginal expression comparisons either on their own or in combination with computational prediction has proven to be difficult both previously [41] and in our own analysis (see following results).
We hypothesize that statistical models guided by knowledge of the miRNA pathway can be used to reduce error in both validating and predicting targeting relationships. The premise of our approach is that although a negative abundance relationship may exist in an m/miRNA pair, this relationship may only be detectable within the context of the abundance of other molecules that participate in mRNA silencing. In a marginal comparison of m/miRNA expression levels for the purpose of verifying a predicted targeting relationship, miRNA expressions are compared directly to those of a putatively targeted mRNA. When expression data from homo sapiens are under study, such a comparison uses miRNA expressions as a direct substitute for those of RISC composed of Ago 2 protein and a targeting miRNA. Additionally, marginal comparisons do not compensate for indirect effects on mRNA abundance caused by the blocking RISCs composed of Ago 1, 3 or 4 proteins and the targeting miRNA. Although ceteris parabis increases of the levels of these RISCs cannot observably reduce the concentration of the targeted mRNA, because they utilize the same target sites as RISCs containing Ago 2 such increases can be anticipated to affect the ability of Ago 2 RISCs to cleave targeted mRNAs. Finally, marginal comparisons do not compensate for either the targeting of the mRNA in a putative target pair by RISCs constructed from miRNAs other than that under consideration, or targeting of mRNAs other than the one under analysis by the miRNA.
In this paper, we develop a linear regression model that accounts for a variety of elements and interactions in the human miRNA pathway and that compensates for idiosyncratic aspects of two data collections on which it is applied. Central to this model is the comparison of the expression levels of a putatively targeted mRNA to a proxy for RISC expression composed of an interaction between Ago 2 and a targeting miRNA, rather than to miRNA expression alone. To demonstrate that our approach offers superior performance to marginal m/miRNA comparisons, we compare the two methods on sets of m/miRNA pairs both previously shown and predicted to have targeting relationships using expression data from two different studies as well as a combination of the data. We find that: 1) the system biological regression approach explains a higher proportion of the observed variation in known mRNA target levels, even after compensating for increases in model complexity. 2) The estimated effects of proxies to targeting Ago 2 RISC expressions on the expressions of known mRNA targets are more consistently and appropriately negative than those of marginal miRNA expressions. 3) A larger number of known m/miRNA target pairs are identified as such using the regression approach compared to marginal m/miRNA methods. 4) The system biological regression approach provides evidence supporting substantially more computationally predicted m/miRNA pairs as bona fide than do marginal m/miRNA comparisons. Because we obtain these improvements in performance without directly utilizing exogenous information from sequence-based computational target prediction methods, our approach provides a basis for statistical methods to putative m/miRNA target pair analysis that can play useful roles in both verifying computational target predictions as well as generating de novo information regarding m/miRNA target relationships.
There are two categories of covariates that ought to be compensated for when comparing the expression levels from a putative m/miRNA target pair in homo sapiens for the purpose of inferring a targeting relationship: those corresponding to elements of the miRNA system biology, and those corresponding to idiosyncratic data effects (if any). Of these two categories, covariates related to the miRNA system biology can be further subdivided into those pertaining to the effect of the particular miRNA under analysis on the putatively targeted mRNA rather than that of other miRNAs potentially targeting the mRNA, those related to observable target cleavage rather than those resulting in translational repression without cleavage through a maintained hybridization at a target site, and those related to the affinity of both the miRNA under analysis as well as other miRNAs to mRNAs not under direct consideration.
It can be presumed that the covariates in these categories are related to one another and to target mRNA expression in a complicated and nonlinear manner, and any statistical or computational procedure for inferring m/miRNA targeting relationships ought to have some degree of fidelity to the system biology represented by the model it is explicitly or implicitly based upon. However, the fidelity of the model also should be balanced against the need for a computationally efficient procedure that works well given the limitations of sample size and the levels of variation in the system. A well-formulated regression model is computationally tractable (especially if large numbers of putative m/miRNA pairs are to be evaluated) and is a standard approach to decomposing variation in a response. Further, although a linear formulation may not emerge from first principles, it may capture the dominant relationships sufficiently well to identify bona fide targeting relationships. Thus we relate the categories of system biologic covariates to the expression of a putatively targeted mRNA as in (1):(1)where [message] refers to expression of the putative targeted mRNA; [putative cleaving RISC] represents the effect of RISCs composed of the putative targeting miRNA and Ago 2 on the targeted mRNA; [putative blocking RISC] is the effect of RISCs composed of the putative targeting miRNA and Ago 1, 3 or 4; [non-specific cleaving RISC] is the effect of RISCs composed of Ago 2 and miRNAs not under particular consideration; [non-specific blocking RISC] is the effect of RISC composed of Ago 1, 3 or 4 and the unconsidered miRNAs; [other targets] refers to the effect that the expression of other mRNAs have on the putative targeted mRNA, especially through their affinity for interactions with the putative targeting miRNA; [idiosyncratic effects] are dataset-specific effects; noise represents natural variation in [message] as well as that due to systemic effects not adequately captured in our model.
Although the levels of RISCs of various types used in (1) are unobserved in RNA microarray expression level measurements, proxies to them can be obtained using available microarray expression data by constructing interaction terms from observable targeting miRNA Ago RNA levels. This preserves a representation of the relevant miRNA biology leading to target cleavage while avoiding complications leading to model nonlinearities, such as seen in equilibrium points of typical chemical kinetics systems. We note that Ago RNA levels are proxies to (unobserved) protein levels. As discussed in Protocol S1, the microarray data was processed to approximate mRNA concentration levels. We assume that these levels are positively related to protein concentration, and so the interaction between Ago mRNA and targeting miRNA levels ought to be positively related to RISC concentration.
Model (2) refines the system biological elements in (1) and provides the beginnings of a formal statistical model. Let i index tissue sample, j index an m/miRNA pair, and consider that expression levels are measured on the logarithmic scale. Further, let mRNAij represent the level of the putative target mRNA in the ith tissue sample of the jth pair; Ago2i and Ago134i be levels of Ago 2 and Ago 1, 3 and 4 (combined); miRNAij and miRNAi−j be levels of the targeting miRNA in the jth pair and the combined levels from other miRNAs; and εij be a random error term assumed to be normally distributed. As suggested, proxies for the concentration of targeting RISCs composed of Ago 2 and Ago 1, 3 or 4 are obtained as products of miRNAij and Ago2i or Ago134i respectively, and analogously for such RISCs composed of miRNAs not under explicit study.(2)
Under model (2), if the jth m/miRNA pair have a targeting relationship then β1j<0 (indicating a negative relationship between expression levels of the mRNA and putatively targeting Ago 2 RISC proxy) would be anticipated. Therefore, a targeting relationship between the jth m/miRNA pair under consideration can be inferred by evaluating the no-targeting relationship hypothesis H0: β1j = 0 vs. HA: β1j<0. To contrast this approach with marginal expression level comparisons of mRNAs to miRNAs, note that an alternative to correlating m- and miRNA levels and evaluating the analogous no-targeting hypothesis H0: ρj = 0 vs. HA: ρj<0 (where ρ represents the true correlation level between m- and miRNA expression levels) would be to estimate the simple linear regression:(3)and evaluate the hypothesis H0: β1j = 0 vs. HA: β1j<0. Of the other effect terms in (2), β5j has arguably the most compelling physical interpretation - if the m/miRNA possess a targeting relationship (as evidenced by rejection of the no-targeting hypothesis), β5j is anticipated to be positive and scaling in magnitude with β1j due to the aforementioned competition for targeting sites between RISCs composed of Ago 2 and Ago 1, 3 or 4. The remainder of covariates and effects used in (2) are included to conform to statistical modeling standards that require inclusion of individual covariates in models that analyze interaction terms (e.g. miRNAij and Ago2i terms), and to have a full representation of the variety of possible effects justified by the system biology (e.g. Ago2imiRNAi−j).
Regression models (2) and (3) were developed on and fit to data from two studies in which both human m- and miRNA expression levels were measured on a reasonably large set of tissue samples. A study of nasopharyngeal cancer (NPC) by researchers in Madison, WI and elsewhere [10],[42] derived whole genome Affymetrix hgu133plus2 microarrays for mRNA profiling, a custom cDNA array for miRNA profiling and RT-PCR for the expression of Epstein-Barr (EBV) genes. Data are available on 31 NPC and 10 normal tissue samples. The second data source was derived from that produced from a study of miRNA expression patterns over a wide variety of tumor and normal tissue types conducted by the Broad Institute [43]. This data collection measures m- and miRNA expression across 67 tissue samples from 10 different normal and tumor tissue types, each tissue type is represented by at least 5 sample observations. Additionally, we merged the Madison and Broad data to create a third dataset in order to fit (2) and (3) to data from the largest number of tissue samples possible. The merged dataset measured m- and miRNA expression across 108 tissue samples from 12 different normal and tissue types (the tissue states from the Madison dataset were not represented in the Broad study). Details of the Madison, Broad and combined data collections is provided in Protocol S1.
In order to validate the system biological regression model, the TarBase miRNA target database [44] was used to derive a set of m/miRNA target pairs that both had been previously validated through the use of gene mRNA and protein-specific techniques (such as PCR, luciferase reporters and immunoblotting) and were represented in the Madison and Broad datasets. (We did not include relationships that were supported by microarray data alone.) In total, there were 76 such m/miRNA target pairs that were commonly measured in both the Madison and Broad datasets and that fit the above criteria (these target pairs were used in the combined data analysis), and 23 additional pairs measured in the Madison data alone. See Table S1 for information pertaining to each of these m/miRNA target pairs.
We note that TarBase classifies target pairs into those reported to result in cleavage or translational repression. To assure that the known target pairs used in this study are competent for observable cleavage, we examined the original studies supporting their inclusion in TarBase. We found no reason to reject any of the pairs labeled in TarBase as resulting in mRNA cleavage as being so competent. However, simultaneous translational repression and cleavage of was demonstrated by a number of target pairs classified in TarBase as translationally repressive [25], and in other studies the use of only protein to miRNA comparisons could not justify such a distinction. Based on our examination of the supporting studies and underlying system biology (as previously described), we did not reject any of the known target pairs based on their TarBase cleavage/translational repression classification and instead regarded all target pairs as competent for Ago 2 RISC-mediated cleavage.
To evaluate the performance of the system biological regression model on computationally predicted but unverified m/miRNA target pairs we used the results of sequence-based comparisons summarized in the miRBase [45]–[47] and TargetScan databases and expression data from the Madison dataset to derive a set of putative target pairs that met three criteria: 1) They were predicted by both miRBase and TargetScan simultaneously, rather than either database singularly; 2) The putative targeting miRNAs in the pairs under consideration were previously identified as differentially expressed between NPC and normal tissue samples [10]; 3) The putative targeted mRNAs in the pairs were those that had above median expression variability. These criteria were used to assure confidence in both the computational target predictions as well as the data used to verify them. The use of putative target pairs simultaneously predicted by both miRBase and TargetScan was motivated by the relatively low overlap between predicted target pairs from these databases – conditional on the miRNA under consideration, TargetScan averaged 301 predicted targets meeting criteria (2) and (3) and miRBase averaged 379, with 48 in common. Constraining the analysis to those pairs with differentially expressed miRNAs and targeted mRNAs with above average expression variability assured that there was sufficient variability in expression levels to permit a statistical analysis to be conducted. (Using mRNAs with above median mean expression rather than variability yielded no substantial differences in the results of our study.) In total, there were 874 putative m/miRNA target pairs that were evaluated using the Madison dataset. See Table S2 for the specific predicted target pairs studied.
The Madison, Broad and combined data collections each exhibit a number of idiosyncratic data effects that might affect the ability to detect m/miRNA target pair relationships. Both Madison and Broad datasets consist of expression measurements from multiple tissue types with highly differentiated expression profiles not directly related to m/miRNA targeting. In the Madison data the tumor samples exhibit varying levels of EBV activity, which has been related to the up- and downregulation of a wide variety of genes both previously [48] and in the Madison data set [42]. In the Broad data, no measurements for Ago 3 expression are available. Finally, in addition to the idiosyncratic effects from the Madison and Broad datasets individually, the composition of the merged dataset from two data studies can be anticipated to introduce complications to even a marginal analysis of m/miRNA expressions.
To compensate for these issues, when analyzing target pair expressions isolated from the Madison data we added two covariates to model (2): a dichotomous variable representing tumor/normal tissue sample state and the expression of the EBV gene EBNA 1. When analyzing target pairs from the Broad data, we added a vector of dichotomous covariates representing tissue type to compensate for tissue type effects and substituted terms aggregating only Ago 1 and 4 for those using Ago 1, 3 and 4. In marginal analyses of the Madison and Broad datasets, no compensation for tissue state was made – as described below, introduction of similar dichotomous variables to model (3) had no effect on the substantive results of the marginal analyses. When analyzing target pairs from the combined dataset, we added to model (3) a dichotomous covariate that represented the dataset origin (Madison/Broad) of the observation under analysis, and added to model (2) the expression of the EBV gene EBNA 1, a vector of dichotomous covariates representing tissue state, and a dichotomous covariate that represented dataset origin.
As for (2) and (3), models that include idiosyncratic data covariates can be used to infer a targeting relationship for the jth m/miRNA pair by evaluating the suggested no targeting relationship hypothesis. Evaluation of such a hypothesis is typically performed via a t-test, and for marginal m/miRNA comparisons using any of the Madison, Broad or combined datasets this procedure is appropriate as the number of parameters are relatively low compared to the number of tissue samples available. However, the high parameterization of the system biological models motivated an alternative analysis based on AIC score minimization [49]. From a fully specified model containing both system biology and idiosyncratic data effects, minimum AIC submodels were computed and examined to determine whether the proxy variable to RISCs composed of Ago 2 and the putatively targeting miRNA was retained as a covariate, and if so, whether the effect of that variable was negative. For observational studies, estimated parameters in models selected for parsimony are sufficient to infer an effect of the associated covariate and so such cases were taken as rejections of the no targeting hypothesis. Additionally, we note that models selected by the AIC criterion can be regarded as implicitly passing a cross-validation test [50]. Therefore, there is a strong relationship between our technique and those that would be predicated upon dividing the data into training and validation sets (e.g. for developing a predictive model for mRNA expression, in which putatively targeting miRNAs might be evaluated as a potential predictor).
To evaluate the significance of the numbers of positive identifications we repeatedly applied the marginal and system biology-based regression model approaches to randomized control data, recording the number of m/miRNA pairs identified as targeting for each repetition. Two complementary randomization schemes were used. In the “no-targeting null” scheme miRNA expressions from each pair under study were permuted across tissue samples, holding the m/miRNA pairing constant – i.e. we condition on the set of m/miRNA expression levels in a given pair, but we randomize their association by resampling the observed miRNA levels. (This randomization was done separately in the two sources to preserve dataset-specific effects in the combined dataset analysis.) By contrast, in our “random pairs” scheme sets of non-targeting m/miRNA pairs were constructed by independently sampling unrelated m- and miRNAs from those under study (i.e. from the set of known target pairs), thus randomizing the pairings while holding the expression levels unchanged across tissue samples.
For the analysis of the known target pair data, multiple (1000) iterations of both randomization procedures were used to construct no-targeting null and random pairs distributions of numbers of positive identification. These distributions provided the basis for calculating p-values for the numbers of positively identified target pairs actually obtained by the marginal m/miRNA comparisons and system biology-based regression models. Under the no-targeting null randomization, dependency between m- and miRNA expressions is explicitly removed. Therefore the distributions of numbers of identifications across repetitions obtained from this procedure can be regarded to be what might be expected if none of the m/miRNA pairs under consideration had true targeting relationships, and the p-values correspond to tests of the hypothesis that the statistical procedure detects more targeting relationships that what would be anticipated if none of the m/miRNA pairs under consideration were bona fide target pairs. Further, the median numbers of identifications from the distributions can be used to infer measures of test specificity. We note that the random pairs procedure does not guarantee that the pairs under analysis do not have a targeting relationship (although known target pairs are rejected from those used in the method, it is possible that the m/miRNA pair is targeting but not yet verified as such), and so inflated numbers of identifications relative to what might be observed under the no-targeting null are expected. In other respects, the distribution and p-values of observed numbers of identifications against the random pairs distribution can be used in the same manner as those from the no-targeting null.
We performed two different analyses that verified the intuitions and results from our randomization tests on the known target pair data. To assure that our no-targeting null distributions were composed of a sufficient number of samples, we reconstructed no-targeting null distributions for the Madison data using 10000 iterations of the procedure described above (rather than the 1000 originally used), and recomputed p-values for the numbers of positive identifications obtained by the marginal and model-based procedures. These p-values were substantially identical to those obtained using 1000 iterations, and considering the heavy computational resources these procedures require we therefore constrained our analysis to the 1000 iteration case. Next, to verify our expectations regarding inflated numbers of identifications in the random pairs distribution we constructed a version of this distribution for the Madison data that was composed of randomly paired m/miRNA expressions taken from the full set of measurements (rather than the subset of m- and miRNAs involved in known target pairs), recomputed p-values as previously and compared these p-values to those originally obtained. In this analysis, no effort was made to remove known target pairs from those randomly sampled. The results of this version of the random pairs scheme are described below, however the test strongly verified our original intuitions.
In the analysis of the computationally predicted target pairs, we conditioned on miRNA and used no-targeting null distributions to obtain 95% upper bounds for the numbers of verifications that might observed from either the marginal comparisons or system biology-based models if none of the predicted m/miRNA target pairs were bona fide. The numbers obtained from the marginal and system biology-based methods on the actual data were then compared to these bounds to provide an indicator of the relative commonality of the results and an informal assessment of the specificity of the methods, analogous to those obtained on the known target pairs. The 95% upper bounds of the no-targeting null distributions were generated using 100 iterations – again, considering the computational resources required we regarded this number as sufficient to obtain a reasonable estimate of the 95% level.
We began by analyzing the expression data from known target pairs. Marginal m/miRNA comparisons were made initially in order to provide a performance baseline for the regression models that incorporated system biological covariates. For the Madison and Broad data analyses, we calculated Pearson correlation statistics on m/miRNA expression levels to measure negative marginal associations, and R2 statistics from the simple linear regression described in model (3) to determine the amount of variation in targeted mRNA expressions attributable to that of targeting miRNA. For the combined data analysis partial correlations of m/miRNA expression controlling for data source were calculated, and adjusted R2 statistics were computed to compensate for the increase in model complexity due to the introduction of the dichotomous data origin covariate. (A partial correlation is a measure of the amount of common variation between two variables after accounting for the effects of a set of related covariates on both. It is analogous to a standard marginal correlation between two variables, which does not account for covariate effects. An adjusted R2 statistic is a measure of model fit analogous to the standard R2 statistic that compensates for the number of covariates in the model. See [51], Chapter 7.10 and 7.7 for technical descriptions of the partial correlation and adjusted R2 statistic respectively.) In each data analysis using marginal methods, the total numbers of positive identifications of m/miRNA target pairing obtained from evaluation of the no-targeting hypothesis through a t-test at the 5% level were obtained. Next, the performance of the system biological regression model on the data from known target pairs was evaluated. Partial correlation statistics for pairs of targeted mRNAs and proxies to RISCs constructed of targeting miRNA and Ago 2, adjusted R2 statistics, and numbers of positive identifications of m/miRNA target pairing obtained from use of the minimum AIC submodel procedures on the versions of model (2) that included data idiosyncratic covariates were computed and compared to the analogous baselines from the marginal m/miRNA comparisons. The number of positive identifications were additionally evaluated using the randomization controls to assure that we obtained greater numbers of identifications than what would be expected under the null hypothesis of none of the m/miRNA pairs under analysis being a legitimate target pair.
We continued by analyzing the computationally predicted target pairs using the Madison dataset. For each putative target pair, the simple linear regression described in model (3) and the version of the system biology-based regression model (2) that incorporated idiosyncratic data effects was used to evaluate the no-targeting hypothesis through a t-test at the 5% level and the minimum AIC submodel procedure respectively. The results from the marginal procedure provided a baseline for evaluating the performance of the system biology-based regression model. The total numbers of verifications both the marginal and system biology-based procedures were conditioned on miRNA and compared directly to one another.
Our analyses were implemented as scripts in the R programming language [52], which were executed on Macintosh OS X computers with installations of R 2.8.0 (earlier versions of R were used at earlier stages in our analysis). Dataset S1 contains the scripts and associated data used to study the known target pairs in the Madison, Broad and combined datasets. Alternatively, the first author may be contacted to provide the archive directly. The archive is commented and can be used to provide further information regarding our procedures, or to rerun our analyses on any system with an R installation (available through the Comprehensive R Archive Network, http://cran.r-project.org). Please direct any questions regarding the archive to the first author.
Because the known target pairs under examination were previously observed to have targeting relationships, it was anticipated that the marginal correlations between m/miRNA expression levels using any of the Madison, Broad and combined datasets would typically be significantly negative. Contrary to these expectations, the sensitivity of marginal m/miRNA expression level comparisons was demonstrated to be quite low. Only 5 of the 99 target pairs in the Madison dataset, 6 of the 76 pairs in the Broad dataset and 7 of the 76 pairs in the combined dataset have significantly negative marginal relationships between m/miRNA expressions (Table 1), and the majority of observed correlations are positive (Figure 2). An example of the relationship between marginal m- and miRNA expression levels in the Madison data is provided in Figure 3 (top row, left column). The example provided compares miR-17-5p to E2F1, a known oncogene. Although miR-17-5p is known to target E2F1, the relationship between m- and miRNA levels is positive.
As suggested in Methods, adding idiosyncratic data effects to our marginal m/miRNA comparison in (3) resulted in nearly no differences in the number of known m/miRNA target pairs successfully identified as such. Using t-testing procedures to evaluate the no-targeting hypothesis after doing so yields 3 of 99, 7 of 76 and 6 of 76 known m/miRNA target pairs identified as such in the Madison, Broad and combined datasets respectively. Similarly, a variety of data transformations were used to attempt to generate an improvement in the overall results without success, and the model fits were checked to assure that the results were not due to systemic outlier effects, model misspecifications or non-normal error terms.
Finally, it was notable that the number of detections obtained by marginal comparisons was well within what might be observed under either the no-targeting null or random pairs distributions (Figure 5, second row). For the analysis of the Madison data, the p-values of the number of positive identifications under the no-targeting null and random pairs distributions were 0.491 and 0.279 respectively, for the Broad data p = 0.193 and 0.800 respectively, and for the combined analysis p = 0.947 and 0.419 (Table 2). In the context of the previously discussed identification performance, these values suggest that the specificity of the marginal procedure approximates the false positive rate under the null hypothesis of no targeting, and therefore that marginal m/miRNA expression level comparisons are as likely to detect evidence of a targeting relationship for unrelated m- and miRNAs as they are for bona fide target pairs.
The observed R2 values from marginal m/miRNA expression level comparisons using data from known target pairs range from less than 0.001 to 0.365 with an average score of 0.061 for the pairs in the Madison data, less than 0.001 to 0.196 with an average of 0.035 for the pairs in the Broad data, and 0.008 to 0.880 with an average score of 0.411 for the pairs in the combined data (Figure 2). In the case of the Madison and Broad analyses, these values indicate that variation in the expression levels of targeting miRNAs explains only a small proportion of that in targeted mRNA levels. As a consequence of this, it would be anticipated that marginal comparisons of m/miRNA expression levels would not be useful for determining whether or not a targeting relationship exists, and therefore the low R2 values rationalize the previously observed performance of the marginal m/miRNA comparisons in identifying the known target pairs as such. In the case of the combined data analysis the observed R2 scores are substantially larger. However, the true relationships of the known m- and miRNA target pairs were not captured when analyzing the combined dataset with the marginal model. Therefore, these high R2 scores simply suggest that the majority of the observed variance in targeted mRNA expression is explained by the origin of the data observation, rather than the appropriateness of the model.
The mean and range of adjusted R2 values for fits of the system biological regression model were (0.524, −0.102–0.922) for the Madison data, (0.310, −0.077–0.602) for the Broad data and (0.712, 0.055–0.974) for the combined data (Figure 4). The increases in observed R2 scores from the baselines obtained from fits of the marginal model indicate that the regression model captures a greater percentage of the variation in targeted mRNA levels, even after compensating for its increased complexity.
Because inclusion of the system biological covariates yielded a model that explained greater amounts of variation in target mRNA levels explained than the marginal model, it was anticipated that it would also better represent the true negative relationship between m/miRNA expression levels from the known target pairs. In fact, partial correlations between targeted mRNAs and proxies to targeting Ago 2 RISCs under model (2) were appropriately negative at substantially higher rates than marginal m/miRNA correlations (53% vs. 38%, 59% vs. 42% and 61% vs. 32% for the target pairs in the Madison, Broad and combined datasets respectively). Additionally, there was a reduction in observed correlation scores taken across the sample of m/miRNA target pairs (mean marginal and partial scores were (0.0934, −0.0247), (0.057, −0.012) and (0.232, −0.083) for the Madison, Broad and combined data). To formalize this comparison, a null hypothesis of equality of marginal and partial correlation scores was tested and rejected for all three datasets using a paired Wilcoxon rank-sum test (p = 0.014, 0.018 and <0.001 for the Madison, Broad and combined data).
Validation of these results consisted of checking the model fits for evidence of systemic outlier effects, model misspecifications or non-normal error terms, as was done for the marginal model fits. A comparative example of the model fits achieved in the Madison data is provided in the top row, right column of Figure 3. After controlling for system biological and idiosyncratic covariates, the relationship between miR-17-5p (which was positive under the marginal model) is appropriately negative.
In a further examination, the effects of the covariates used in the AIC-optimal submodels of the fits of (2) on the Madison data were studied to assure that the model was not overspecified. Of the variety of covariates used in the version of (2) compensating for the idiosyncratic data effects, only the dichotomous variable indicating tissue type found low levels of use in the AIC-optimal submodel – in fact, it was never included in the AIC-optimal submodels, indicating that tissue type never had a substantive effect on a targeted mRNA level after compensating for other effects. Because few if any of the known m/miRNA target pairs under consideration have been previously observed to be differentially expressed in NPC, this might be reasonable. Alternatively, this result can be explained by noting that EBV expression in the Madison data is highly associated with NPC, and therefore statistical control of EBV expressions rather than tissue type may be sufficient for both.
Related to this analysis, the estimated effects of proxies for targeting RISCs composed of Ago 1, 3 and 4 from the AIC-optimal submodels were compared to those composed of Ago 2 in order to assure that the model was performing in a reasonable manner. Figure 6 displays the relationships of estimated effects of targeting Ago 2 RISC proxies to targeting Ago 1, 3 and 4 RISC proxies, for AIC-optimal submodels estimated on the Madison data in which both covariates were included and the estimated effect of the targeting Ago 2 RISC covariate was appropriately negative (there were 13 such cases out of the 33 in which the effect of the targeting covariate was so). It can be observed that, as anticipated, the estimated effects of targeting Ago 1, 3 and 4 RISCs on targeted mRNA levels are indeed generally positive with effect sizes scaling with those of targeting Ago 2 RISCs.
Overall, these results demonstrate that the relationships between targeted mRNAs and proxies to targeting Ago 2 RISC, compensating for other relevant biological covariates, better represent the actual relationship of the known target pairs than marginal m/miRNA expression level correlations.
Based on the improvements in model fit, it was further anticipated that evaluating the no-targeting hypothesis using the system biological model and the minimum AIC submodel procedure would indicate a greater number of positive identifications of targeting relationships than obtained by marginal m/miRNA comparisons. In fact, model (2) identified 33 of 99, 20 of 76 and 36 of 76 known m/miRNA target pairs as having expression profiles consistent with targeting relationships in the Madison, Broad and combined datasets respectively. This represents up to a sevenfold increase from the baseline obtained by marginal m/miRNA expression level comparisons (Table 1), and demonstrates the improved sensitivity of model (2) in detecting m/miRNA target pair relationships. We note that although under 50% of known target pairs were recovered by model (2), this level of identification performance is similar to the individual performances obtained by a number of sequence-based computational methods [30]. In particular, using the Madison and combined datasets we were able to successfully identify 33 and 47% of the known targets pairs we evaluated, whereas TargetScan and miRBase are reported to have 21 and 48% consistency with experimentally supported target pairs.
The numbers of detections obtained by model (2) relative to what might be expected under either of the randomization techniques show similar improvements from the baseline obtained by marginal m/miRNA expression level comparisons (Figure 5, first row). For the analysis of the Madison, Broad and combined datasets, the p-values of the number of positive identifications under the no-targeting null were 0.008, 0.096 and 0.001 respectively. Likewise, under the random pairs distributions the p-values were 0.053, 0.241 and 0.072. In the case of the Madison, Broad and combined data analyses, the numbers of identifications obtained are at least marginally significantly greater than what is typically observed under the no-targeting null. Under the random pairs distribution the Madison and combined data analysis show similar results, while the number of identifications made using the Broad data is not significantly greater than what might be expected under the null. As suggested above we anticipated an overall inflation in p-values under the random pairs technique due to inadvertent sampling of as-of-yet unverified target pairs from the sets of known target pairs used as a basis for the technique. As discussed in Methods, to verify this intuition we performed a secondary analysis of the number of detections obtained for the Madison data against a random pairs distribution constructed from the full set of m- and miRNAs for which expression measurements were available. The p-value from this study was 0.007; based on this result we regarded the inflated p-values under the random pairs distributions as a statistical artifact and focused our attention on the results from the no-targeting null.
In total, the results from our analysis imply that the improvements in sensitivity for detecting target pairs obtained through model (2) are greater than any loss in specificity that might be incurred relative to that of the marginal procedure. The overall specificity of (2) for rejecting non-targeting pairs in the Madison and combined datasets are approximately 80%, as can be observed from median numbers of acceptances under the non-targeting null distribution. The analysis of the Broad data did not yield significantly larger numbers of correct identifications under either the no-targeting null or random pairs distributions, however it is useful to note that the high number of tissue types and missing Ago 3 measurements in the Broad dataset can be anticipated to negatively affect our ability to detect m/miRNA target pair relationships from expression levels. As well, the Madison dataset was processed to provide measurements in terms of concentration estimates that can more naturally be aggregated than the RMA measurements provided in the Broad data.
The overall results of evaluating the computationally predicted m/miRNA target pairs on the Madison data with the system biological regression model and marginal m/miRNA comparison are described in Figure 7. (Table S2 provides further detail on results obtained for particular m/miRNA pairs analyzed by the system biologic regression model.) For each miRNA under consideration, the first, second and third columns of Fig. 7 provide the numbers of putative target pairs evaluated and positive validations obtained by the system biological model and marginal m/miRNA comparisons respectively. In the second and third columns, 95% upper bounds on number of positive validations expected under the no-targeting null are provided.
Visual inspection of Figure 7 suggests that model (2) yields substantially more verifications than the marginal method in nearly every case. In fact, the marginal method most often yields no verifications of computationally predicted targets of any of the miRNAs considered. The average percentage of predicted targets validated by the system biologic regression model is 25.68%, taken across all miRNAs. For 6 of the 18 miRNAs conditioned upon, the number of verifications obtained was significantly (p<0.05) greater than what might be expected under the no-targeting null (miR-130b, -15a, -16, -181a, -181c, -30d), and analyses conducted on targets predicted for an additional three miRNAs (miR-192, -224 and -212) yielded numbers of identifications that were substantially greater (p = 0.08, 0.13 and 0.28 respectively). In comparison, marginal comparisons validate an average of 7.83% of predicted targets, and yielded three miRNAs (miR-212, -29a and –29c) associated with significantly greater numbers of verifications than what might be expected under the no-targeting null with one additional miRNA (miR-133a) having a substantially greater number (p = 0.21). Although some inflation in the number of verifications that might be observed under the no-targeting null was incurred through when using the system biologic regression model rather than marginal m/miRNA comparisons, the results obtained here are roughly consistent with the performance of the marginal and system biologic regression methods on the set of known target pairs.
Based on these results, a further comparative inspection was made of the distributions of the estimated marginal and Ago 2 mediated effects from fits of the miRNAs under analysis against all mRNAs in the Madison dataset. Sample distributions for estimated and normalized marginal effects of miR-29c and estimated miR-30d Ago 2 RISC effects are provided in Figure 8. (These were selected due to their high numbers of predicted target pair verifications, as seen in Figure 7.) The estimated marginal effects of miR-29c are clearly negatively biased, explaining the high numbers of validations. The estimated miR-30d Ago 2 RISC effects do not have such a bias. Instead, they demonstrate a bimodality with a main mass centered at 0 effect and a smaller mass centered at −1.5. Such a distribution is consistent with a categorization of genes into two classes: those regulated by miR-30d, and those not. Although analyses of such large-scale screen results are ongoing, the results in Figures 7 and 8 provide further evidence that that use of statistical models which compensate for the system biology related to miRNA-based gene silencing are more appropriate for validating and predicting m/miRNA targeting relationships than marginal expression level comparisons.
The effects of miRNAs on mRNA stability and translation are presently understood to have effects on organism development and physiological function, and have been linked to diseases such as cancer. It is of acknowledged importance to develop greater insight into the targeting relationships between m- and miRNAs. In this paper, we considered the role that biology-based statistical modeling and methods might play in the m/miRNA target prediction problem. Currently, the statistical techniques used for these purposes are typically based on marginal comparisons of individual m- and miRNA expressions across tissue samples. In some respects this is a natural comparison to consider – many early studies verifying predicted targeting relationships were based on transfection experiments with small numbers of samples, for which marginal m/miRNA comparisons might be the only procedure available. However, it has been observed previously (and was demonstrated here) that in practice these methods typically yield relatively disappointing results.
We hypothesized that improvements in the performance of statistical methods for detecting m/miRNA target pair relationships might be achieved through development of a statistical model and associated hypothesis testing procedure better tied to the underlying system biology. In an investigation of this biology in homo sapiens we identified a number of factors that we expected to affect the ability of marginal m/miRNA expression level comparisons to detect targeting relationships, many related to the dependence of the gene silencing mechanism on the construction and varied actions of RISCs. Based on this as well as additional information pertaining to the data under analysis, we developed regression methodology for testing hypotheses of no targeting relationship between m- and miRNA. Our rationale for choosing regression methods (as opposed to other possible statistical or computational methods) was motivated by the balance it offered between the competing goals of fidelity to the system biology, having a methodology with understood theoretical underpinnings and computational tractability for analyzing large number of putative m/miRNA target pairs, while being appropriate to the data quality and sample size.
In comparison to procedures based on marginal m/miRNA expressions, our models and procedures were shown to provide substantial improvements in overall model fit and detection performance for sets of known m/miRNA target pairs, although the degree of such improvement was somewhat dependent on the study design. As would be hoped, we further demonstrated that such improvements were carried over into the problem of validating predicted m/miRNA target pairs. Our study suggests that use of the regression models and associated hypothesis testing procedures developed here (or equivalent techniques based on the system biology) represent a reasonable alternative to methods based on marginal m/miRNA comparisons for analyzing expression data in m/miRNA targeting studies, and in conjunction with high throughput data can be used to either verify computationally predicted relationships or generate de novo information regarding m/miRNA target pairs. In fact, our model demonstrates consistency with known target pairs on par with many computational target prediction algorithms [30].
Because there have been few systematic studies of statistical methods for detecting m/miRNA targeting, there is little context that can be used to help evaluate our results. The most relevant external work is that recently conducted by Huang et al [53]–[55], however there are a number of differences between our studies. Huang et al focus on Bayesian methods to update a set of prior probabilities of targeting relationships between m- and miRNAs using marginal expression comparisons. These prior probabilities are, in their reported work, highly tied to the results of computational target prediction algorithms (in particular, TargetScan). The posterior probabilities obtained through their technique are compared to a threshold based on those obtained from a high-confidence set of m/miRNA target pair expression values; m/miRNA pairs with posterior targeting probabilities meeting the threshold are accepted as valid target pairs. In contrast, our study is framed in terms of evaluating a single m/miRNA pair for evidence of a targeting relationship, compensating for the underlying system biology (which includes the effects of other targeted and targeting m- and miRNAs on the m/miRNA pair under consideration). Our use of a hypothesis testing framework allows us to avoid the need to set a thresholding value based on a separate set of m/miRNA expression data for evaluating whether potential m/miRNA pairs evidence a targeting relationship. We do not tie our work to any particular computational target prediction algorithm, a position we view as appropriate given the issues with their specificity, sensitivity and inter-algorithm consistency.
Further, the emphasis of our presentation of algorithm development and results is substantially different from Huang et al. We choose to focus development of a statistical method on known m/miRNA pairs and then use the resulting procedure to validate a set of computational target predictions. Huang et al are primarily concerned with using their algorithm to validate computational predictions, with verification of their method on known target pairs taking place only on those that are represented in their set of computational target predictions [53]. It is unclear whether these differences in presentation have a substantial difference in performance. The methods proposed here and by Huang et al verify approximately the same proportion of computational target predictions evaluated, and Huang et al [53] demonstrate that of 19 known target pairs contained in the set of computationally predicted targets that they attempt to evaluate, 9 are identified as such. Overall, comparing the two methods and constructing new statistical procedures that incorporate elements of each may be one direction for achieving further improvements in the ability to detect m/miRNA target relationships from high-throughput expression data.
A similar issue that this study only indirectly addresses is the topic of how to best combine results across multiple sequence-based computational or expression-based methods, in order to obtain an aggregate estimate of the full set of m/miRNA target pairs occurring in humans. Such techniques can be classified into two categories: Those that would use sequence-based and expression-based methods sequentially (e.g. using expression-based methods to validate sequence-based predictions or using sequence-based methods to rationalize de novo expression-based predictions with a target site), and those that would use them simultaneously (i.e. without using one type of method conditional on the results of the other). Here, after establishing the utility of our data on known target pairs, we demonstrate how it might be used in a sequential study conditional on the results of sequence-based methods. To perform either a sequential study in which sequence-based methods are used conditional on de novo expression-based predictions or a simultaneous study using both sequence-based and expression-based predictions, the development of statistical methods which can distinguish between a bona fide m/miRNA target pair and m/miRNA pairs related through an intermediate, targeted, translationally activating mRNA must be developed. We are currently working on the development of such a technique. Additional complications that ought to be addressed in such studies is how best to handle the multiple comparisons problems that occur due to the large number of m/miRNA pairs that might be evaluated (which are orders of magnitude larger than those encountered in typical differential expression studies, for example), and how to best align results from multiple algorithms and datasets. We feel that, much as this study utilized known m/miRNA target pairs to validate our regression model, it is reasonable for future proposed methods for handling these technical problems to use them as a basis for evaluation and validation.
Aside from our current work towards the development of a statistical technique capable of de novo m/miRNA target pair prediction, we are extending our work in large-scale screening of putative m/miRNA target pairs (such as described in Figure 8). Our work consists of both investigating and improving our statistical procedures for inferring such relationships as well as aligning predictions from sequence- and expression-based methods, and by further supplementing the data used in this study with new samples as they become available. In a study of a recent dataset originally analyzed by Ambs et al [56], many of the results obtained here are reiterated. Figure 3 provides an example. Consistent with our result using the Madison data, miR-17-5p shows no substantial relationship with E2F1 in a marginal analysis (bottom left panel), but after controlling for the biological and idiosyncratic covariates the true negative relationship between them can be observed (bottom right). Based on this study, those of Huang et al, and the continued release of high-throughput data studies comparing m- and miRNA expression, we look forward to the further development of statistical methods for detecting m- and miRNA targeting relationships from expression data.
|
10.1371/journal.pntd.0001620 | Long-Term Impact of the World Bank Loan Project for Schistosomiasis Control: A Comparison of the Spatial Distribution of Schistosomiasis Risk in China | The World Bank Loan Project (WBLP) for controlling schistosomiasis in China was implemented during 1992–2001. Its short-term impact has been assessed from non-spatial perspective, but its long-term impact remains unclear and a spatial evaluation has not previously been conducted. Here we compared the spatial distribution of schistosomiasis risk using national datasets in the lake and marshland regions from 1999–2001 and 2007–2008 to evaluate the long-term impact of WBLP strategy on China's schistosomiasis burden.
A hierarchical Poisson regression model was developed in a Bayesian framework with spatially correlated and uncorrelated heterogeneities at the county-level, modeled using a conditional autoregressive prior structure and a spatially unstructured Gaussian distribution, respectively. There were two important findings from this study. The WBLP strategy was found to have a good short-term impact on schistosomiasis control, but its long-term impact was not ideal. It has successfully reduced the morbidity of schistosomiasis to a low level, but can not contribute further to China's schistosomiasis control because of the current low endemic level. A second finding is that the WBLP strategy could not effectively compress the spatial distribution of schistosomiasis risk. To achieve further reductions in schistosomiasis-affected areas, and for sustainable control, focusing on the intermediate host snail should become the next step to interrupt schistosomiasis transmission within the two most affected regions surrounding the Dongting and Poyang Lakes. Furthermore, in the lower reaches of the Yangtze River, the WBLP's morbidity control strategy may need to continue for some time until snails in the upriver provinces have been well controlled.
It is difficult to further reduce morbidity due to schistosomiasis using a chemotherapy-based control strategy in the lake and marshland regions of China because of the current low endemic levels of infection. The future control strategy for schistosomiasis should instead focus on a snail-based integrated control strategy to maintain the program achievements and sustainably reduce the burden of schistosomiasis in China.
| Schistosomiasis japonica is an important disease in China with a documented history of more than 2,100 years. The World Bank Loan Project (WBLP) implemented during 1992–2001 contributed greatly to China's schistosomiasis control. This study shows that the long-term impact of WBLP strategy on schistosomiasis control was not ideal. It can only maintain the morbidity of schistosomiasis at a low level, but can not reduce it further. Also, the WBLP strategy could not effectively compress the spatial distribution of schistosomiasis risk. To achieve further reductions in schistosomiasis-affected areas, and for sustainable control, focusing on controlling the intermediate host snail in the lake and marshland regions was suggested to be the next step to interrupt schistosomiasis transmission within the two most affected regions surrounding the Dongting and Poyang Lakes. While in the lower reaches of the Yangtze River, the WBLP's morbidity control strategy may need to continue for some time until snails in the upriver provinces have been well controlled.
| Schistosomiasis japonica, a disease caused by the trematode Schistosoma japonicum, has a documented history of more than 2100 years in China [1]. It severely impacts the health of residents within endemic areas, causing substantial morbidity such as wasting, weakness, ascites and growth retardation [1], [2], [3]. Recognizing the large public health and socio-economic impact of this disease, the government of China initiated a large-scale schistosomiasis control program in the mid-1950s and achievements have been monitored during nearly 60 years of continuous endeavor [2], [4], [5]. At present, the schistosomiasis endemic regions have been largely reduced and confined to seven provinces along the Yangtze River: five provinces of Hunan, Hubei, Anhui, Jiangxi and Jiangsu in the lake and marshland regions and two in the mountainous regions, Yunnan and Sichuan provinces [6], [7], [8]. Many projects or programs have contributed to this success. Among others, the World Bank Loan Project (WBLP) targeted at schistosomiasis control in China has played an important role during the period of 1992–2001. Zhang&Wong [9] and Chen et al. [10] evaluated the impact of the WBLP strategy before and shortly after the end of the project. These authors concluded that the original objectives of the WBLP strategy– to control schistosomiasis morbidity – had been met, but that snail infested areas had increased to a certain degree and that snail infection fluctuated at low levels. After the termination of the WBLP strategy, there was a consistent gap between available funding and the financial resources required to maintain program achievements and to make further progress [10]. Many researchers have reported that schistosomiasis prevalence has rebounded in some regions, even where the criteria of transmission interruption or control had been previously met [11], [12]. A national survey carried out in 2004 confirmed the re-emergence of schistosomiasis in China [13]. This led us to question the long-term impact of the WBLP strategy and investigate a more sustainable strategy for schistosomiasis control in China [2], [14].
There are two potential limitations regarding the previous assessments of the success of the WBLP strategy. One is that previous studies only evaluated the short-term impact of the WBLP strategy [9], [10]. The WBLP's control strategy focused on the large-scale use of chemotherapy to control morbidity in humans and livestock, an approach which has been frequently questioned regarding its sustainability. For example, the compliance rate of chemotherapy can decrease to a great extent because of fatigue with repeated treatments [6], [15]. The long-term impact of the WBLP strategy could be different from the short-term situation. The second potential limitation is that earlier assessments were only based on a non-spatial perspective. That is, they used the magnitude of the absolute number of cases to evaluate the overall control effect, but neglected to consider the spatial aspects which can provide some new and even different results. It is well known that a certain number of cases in a region could be due to two completely different scenarios, a clustered risk profile with cases concentrated in only a few areas and a random risk profile with cases occurring nearly randomly throughout the region. Different risk patterns require a distinct control strategy and decision-making process. So evaluating the long-term impact of the WBLP strategy from a spatial perspective would be valuable for future planning of schistosomiasis control.
In this study we used the national datasets from the five provinces in the lake and marshland regions from two periods, 1999–2001 and 2007–2008, with the aim of assessing the long-term impact of the WBLP strategy on schistsomiasis. We compared the changes of the spatial risk distribution of schistosomiasis between the two study periods and contrasted the compositions of the two random effects of spatially correlated and spatially independent heterogeneities.
Our study was carried out in the lake and marshland regions of schistosomiasis in the middle and lower reaches of the Yangtze River, which included the five provinces of Hunan, Hubei, Anhui, Jiangxi and Jiangsu. According to the latest report (2009) on the schistosomiasis situation, it was estimated that 98.7% of the snail-infested areas and 97.8% of the S. japonicum infected people in China were concentrated in these five provinces [16]. These provinces included 261 schistosomiasis endemic counties. Of these, 115 reached the criterion of transmission interruption, 57 achieved transmission control and 89 had ongoing transmission [16], [17]. It is obvious that the lake and marshland regions should be the focus for China's schistosomiasis control program.
The WBLP project started in 1992. It was completed at the end of 1998 in Anhui, Jiangxi, and Jiangsu provinces and continued until the end of 2001 in Hubei and Hunan. The three provinces that completed the project in 1998 continued to carry out schistosomiasis control activities using their own funds and according to the operational plan set out by the WBLP [10]. Thus, these provinces underwent similar stages of the schistosomiasis control strategy.
The county-level prevalence data on S. japonicum infection in the lake and marshland regions were obtained from the national annual report on schistosomiasis. These data were first collected through village-based field surveys using a two-pronged diagnostic approach (screening by a serological test on all residents of 5 to 65 years old and then confirmation by a parasitological test), then reported to the towns and finally summed at the county level, with only the county-level totalized databases made available to us. This system of recording and annual reporting had been in place since 1999 [18]. The diagnostic criteria and diagnostic approaches for schistosomiasis cases were as per the national guidelines [19].
Prior to and including during 2004, these data were maintained and managed by Fudan University (formerly Shanghai Medical University). Beginning 2005, this task was taken over by the National Institute of Parasitic Diseases in Shanghai (Formerly the Institute of Parasitic Diseases), Chinese Center for Disease Control and Prevention. The national databases for the lake and marshland regions during the two periods of 1999–2001 and 2007–2008 were obtained from the corresponding institutes, respectively. The total number of schistosomiasis cases and population at risk in each county were used to estimate the prevalence of schistosomiasis and to analyze the dynamics of spatial risk distribution and spatial heterogeneities of schistosomiasis-related risk between the two study periods.
County-based digitized polygon maps in the lake and marshland regions were obtained for the five study provinces [5], [6]. Digitized maps of the Yangtze River and Dongting and Poyang Lakes were also obtained. Attribute data on schistosomiasis were linked to the county maps to establish the spatial database and facilitate the visualization of the results. During the 10-year study period, the administrative boundaries of counties changed slightly. To simplify the analysis and for comparability of the results, the administrative divisions in 2008 were used as the standard and data from the other study years were modified accordingly. Two types of topological manipulations were involved, the merging and splitting of polygons, which have only ignorable impacts on this study because the county-level data is much too unspecific to really capture the dynamics of schistosomiasis. The former operation combined two or more polygons into a single, new polygon; and the number of schistosomiasis cases and the population at-risk were then summed to produce the new polygon's disease data. The latter operation divided one polygon into two or more new polygons, in which the number of schistosomiasis cases in each new county was estimated using the proportion of the population of the original county that was present in each of the new counties. All data manipulation was undertaken within ArcGIS9.2 software (Environmental Systems Research Institute, Inc., Redlands, CA, USA).
The analysis consisted of four procedures. Firstly, crude prevalence of schistosomiasis was calculated and summarized for those counties with reported cases using conventional descriptive statistics (e.g., median and quartiles).
Secondly, the counties in the lake and marshland regions were classified into five classes according to the dynamics of their prevalence status: 1.) non-endemic counties where no cases were reported during the two study periods; 2.) unchanged endemic counties where schistosomiasis cases were continuously reported across years; 3.) disappeared endemic counties where cases were continuously reported in 1999–2001, but no cases were reported in 2007–2008; 4.) newly appeared endemic counties where cases were continuously reported in 2007–2008, but no cases were reported in 1999–2001; and 5.) fluctuating endemic counties where cases were reported in one or two years of 1999–2001 and one year of 2007–2008. Maps of these categories of endemicity were created using ArcGIS9.2 software.
Thirdly, a Bayesian random-effect model-which was first introduced by Clayton and Kaldor in 1987 [20] and developed further by Besag et al. in 1991 [21]- was built to analyze the spatial distribution of schistosomiasis for different time points. In this model for estimating relative risk (RR), area-specific random effects are decomposed into two latent components: one component represents the effects of schistosomiasis-related risk factors that vary in a structured manner in space (referred to as correlated heterogeneity, CH) and another component indicates the effects from schistosomiasis-related risk factors that vary in an unstructured way among areas (referred to as uncorrelated heterogeneity, UH). The model is formulated as [22],(1)(2)where is the number of reported schistosomiasis cases in county i; is the expected number of schistosomiasis cases in county i; is the expected or predicted RR in county i; is the overall level of risk assuming the effects of UH and CH are zero; is the correlated heterogeneity (CH) which was modeled using the conditional autoregressive (CAR) structure and is the uncorrelated heterogeneity (UH) that was modeled using a Gaussian distribution, for which the following formulas are used, respectively:(3)(4)where is the weight of the neighbor j for area i, if i, j are neighbors, otherwise ; and are precisions and their inverses are the variance of and , respectively.
Bayesian methods were used to fit the spatial model, implemented using WinBUGS1.4.1 software (Imperial College and MRC, London, UK). Parameters and control the variability of CH and UH effects, for which prior distributions were specified using the same vague gamma prior distributions: Gamma(0.5,0.0005). For the baseline risk , a vague normal prior distribution: N(0,0.0001) was used. Model fitting was carried out using two separate chains starting from different initial values, and 30,000 iterations were run: the first 10,000 samples were discarded as burn-in and the remaining 20,000 iterations from each chain were used for parameter estimation. Convergence was checked by visual examination of the time series plots of samples from each chain and by computing the Gelman and Rubin diagnostic statistic.
Finally, the estimated parameters of the model (overall risk , the variation of the UH and CH components) were summarized in a table. The predicted RR, the posterior probability of RR>1, the ratio of UH to CH and the model residuals were exported and linked with the digitized polygon maps using ArcGIS9.2 software to display their spatial distributions.
The number of counties reporting schistosomiasis cases and the overall crude prevalence increased slightly from 1999 to 2001 during the later period of the WBLP strategy and then decreased during 2007–2008, but the prevalence was still higher than that in 1999. The overall variation (95% CI) in the prevalence showed a tendency of continuous decline except for a slight rebound in 2001 (Table 1).
Figure 1 shows the distribution of the unchanged, disappeared, new appeared and fluctuating endemic counties. The number (proportion) of endemic counties for each type were 110 (64.71%), 25 (14.71%), 4 (2.35%), and 31 (18.24%), respectively. The four newly appeared endemic counties in 2007–2008 were all located in Hubei province. Different types of endemic counties were intermingled along the Yangtze River and the Poyang and Dongting Lakes. Exceptions are one fluctuating and one disappeared endemic county, which were isolated and located in Jiangxi province. Besides, the fluctuating and disappearing counties are mainly on the geographical margins of the endemic areas.
From Table 2, we see that the overall risk of schistosomiasis in the counties tended to decrease gradually and that the risk in 2008 was reduced to less than half of the risk in 1999. The changes in variation of the CH and UH components were different. The former fluctuated across years and reached a peak in 2008 whereas the latter first decreased from 1999 to 2001, then increased again from 2007 and finally rebounded to a level that was similar to 1999. Except for the year of 1999, the variation in the CH component was greater than the variation in the UH component.
The magnitude of the relative risk generally decreased during the study period and the counties with posterior expected RR>1 were mainly located in the areas surrounding Poyang and Dongting Lakes. In the southern part of Yangtze River, the areas with RR>1 were relatively stable in space; while in the northern part of Yangtze River, the spatial distribution of predicted risk counties spread northwards around the Dongting Lake, but was compressed in areas near Poyang Lake (Figure 2).
Figure 3 shows that the counties with the highest probability of greater than average risk (i.e., RR>1) were mostly confined to Poyang and Dongting Lakes and their neighboring counties. For the counties where the posterior probability was high in the lower reaches of the Yangtze River in 1999–2001, the posterior probability was reduced in 2007–2008.
In the endemic regions around the Poyang and Dongting Lakes, schistosomiasis risk was dominated by the UH effects of schistosomiasis-related risk factors in 1999, but the CH effects were dominant in the other years. In contrast, in the lower reaches of Yangtze River the primary component of the heterogeneity effects for schistosomiasis risk was relatively unstable (Figure 4).
The model residuals ranged between −1.40 and 1.65. No obvious outliers were identified, so the distribution of the residuals after adjusting for the effects of the UH and CH components were not displayed here (data are available upon request).
This study presents an application of Bayesian methods to evaluate the long-term impact of the chemotherapy-based WBLP strategy on the spatial distribution of schistosomiasis japonicum. It is widely reported that the transmission dynamics of schistosomiasis are closely related to socio-economic, climatic, demographic, biological and environmental factors [23], [24], [25]. When studying the epidemiology of schistosomiasis and evaluating the impact of control strategies, it is impossible to consider all the potential risk factors related with interested diseases because either the information is unavailable or disease mechanisms are unclear. Hence, previous reports have only included the most important factors or those of specific interest [4], [5], [26], [27], [28], [29], [30], [31], which will contribute to bias in effect estimates because of unadjusted effects from risk factors that have been ignored. From the spatial perspective, all the potential risk factors related with studied diseases can be divided into two latent components of random effects, spatially correlated heterogeneity (CH) and uncorrelated heterogeneity (UH). This idea has been mathematically demonstrated within the well known Besag, York and Mollié model [21], which permits us not only to analyze and predict the disease risk more accurately, but also to study and compare the dynamics of the two latent components. For schistosomiasis, CH could encompass the combined effects of all the unmeasured environmental factors (e.g., normalized difference vegetation index (NDVI), land surface temperature (LST), rainfall), socio-economic factors (e.g., income), and the biological factors of the intermediate host Oncomelania hupensis (e.g., density of snails), which are affected by the other two factors. Among others, the snails are the focus for schistosomiasis control, so the transmission control strategies are closely related with CH such as mollusciciding, environmental modification and liming the banks along canals and drainage ditches. Unfortunately, this was not the focus for WBLP's schistosomiasis control strategy. By contrast, UH could encompass spatially unstructured risk factors, which might include demographic factors (e.g., age, gender, education level, and occupation), human behavioral factors (e.g., frequency of contact with infected water, personal protections) and livestock-related factors. Here, the reservoir hosts (human and cattle) are the intervention targets and the morbidity control strategy with chemotherapy-based population treatment emphasizes the observed UH. Hence, the dynamic changes in CH and UH can help signify the control effect of schistosomiasis to a certain degree and direct future control emphasis.
From our study, we see that the prevalence of schistosomiasis has been greatly reduced and maintained at a low level. The prevalence during 2007–2008 was reduced further, but still in the same order of magnitude (10−3). This suggests that the chemotherapy-based WBLP strategy has had little effect under the low endemic levels of schistosomiasis in China and some new control strategies are needed [32], [33]. Also, we found that the spatial distribution of schistosomiasis risk during 2007–2008 was only slightly reduced and <10% endemic counties in 1999–2001 were declared to be free of cases in 2007–2008. The most severely affected areas were located along the Yangtze River, in the areas of the great lakes (Dongting and Poyang Lakes), and their surroundings [1], [34], [35], where the predicted RR>1 and its posterior probability were high. The spatial distribution of current schistosomiasis risk seemed to be stable and was not obviously compressed in space. This was confirmed by the intersecting distributions of fluctuating, newly appeared, disappeared and unchanged endemic counties. Combined with the above results, we may conclude that continuing a chemotherapy-based control strategy following the decade-long WBLP is likely to contribute little to further schistosomiais control under the current situation of low morbidity levels. We also conclude that it is difficult to further restrict the spatial distribution of schistosomiasis endemic regions using current control methods. The rebounding of prevalence during 1999–2001 and the increasing RR in some counties during 2007–2008, which were partly due to reduced compliance with drug use and the persistence of extensive snail habitats, suggests that the impact of the WBLP strategy was inconsistent. This has been frequently highlighted by many researchers who suggest that snail control should be given more attention for sustainable control of schistosomiasis [2], [5], [6], [14], [36].
The variation in UH was decreased from 1999 to 2001 and then increased again from 2007 to 2008, reflecting a good short-term control effect of chemotherapy-based WBLP strategy, but the long-term effect may not be optimal. The variation in CH showed a tendency of increasing and in 2008 it was over 3 times as that of 1999. This may suggest the risk from the intermediate host snails increased continuously, possibly because snail control was not emphasized within the WBLP strategy. For the predicted high risk regions around the Dongting and Poyang Lakes, the ratios of UH to CH were over 1 only occurred in 1999, suggesting that CH has become the main component of current schistosomiasis risk in those regions where transmission control strategy focusing on snail control should be implemented. While for the lower reaches of Yangtze River, it is more complicated for the changes of UH to CH. The possible reason is the control effect was not only affected by itself, but also was influenced by the schistosomiasis epidemics in the upriver provinces where the cercaria could be brought to the downriver provinces following the water flows. So morbidity control strategy in the lower reaches of Yangtze River should be maintained until the snails in the upriver provinces have been well controlled.
Besides, the frequent flooding of the Yangtze River, water resource development projects (e.g., Three Gorges Dam Project and South-to-North water transfer project) [37], [38], climate change/global warming, anti-flood policies (returning reclaimed land to lake, leveling dykes between main levees and building new towns for resettlement), mobility of populations, the frequent trade of livestock and increased tourism and travel to endemic regions are all important drivers for the fluctuation, (re)-emergence and spread of schistosomiasis and contribute to the continuing challenge of schistosomiasis control, especially sustainable control. Experiences in China and Japan indicate that controlling the intermediate host snail should result in a more sustainable impact compared to other control approaches [17], [39]. The primary CH component of current schistosomiasis risk identified in this study also implies that a future control strategy should shift to transmission control strategy with snail control as an emphasis, which would have a sustainable impact on controlling schistosomiasis and be helpful for facilitating the ultimate elimination of schistosomiasis in China.
There are two potential shortcomings in our study that warrant discussion. One limitation is that the quality of reported schistosomiasis data from different regions may be inconsistent. Case data were collected through a bottom-up disease reporting system aimed at monitoring the national disease status, so the data reliability was largely based on the quality of data reported by local institutions. Another limitation is that the diagnostic approach for schistosomiasis is not 100% sensitive and specific. The parasitological test (e.g., Kato-Katz technique) has a low sensitivity, while serological tests (e.g., indirect hemagglutination assay, IHA) have low specificity [40]. According to expert opinion, the estimated sensitivity and specificity of the Kato-Katz technique are about 20–70% and 95–100%, respectively and for the IHA technique about 90–95% and 85–90%, respectively [41]. Therefore, some correction methods for the prevalence estimates are needed for a precise analysis. Wang et al. (2008) reported a simple Bayesian approach to correct the reported prevalence of schistosomiasis by considering the uncertainties of the diagnostic approaches [42], but they assumed that the diagnostic methods had the same uncertainties for all levels of endemicity. In fact this is not the case and it could introduce other biases to the results. The best possible solution may be to develop a correction method weighted according to prevalence, where the determination of appropriate weights is the most important issue. This is an area of future work.
In conclusion, we conducted a comparative study of the spatial epidemiology of schistosomiasis using two datasets from 1999–2001 and 2007–2008 in the lake and marshland regions to evaluate the long-term impact of WBLP strategy on China's schistosomiasis status. The WBLP strategy appears to have had a good short-term impact on schistosomiasis control, but its long-term effect has not been ideal. It has successfully reduced the morbidity of schistosomiasis to a low level, but can not contribute further to current schistosomiasis control considering the low endemic level in China. The WBLP strategy has also failed to reduce the geographical range of affected counties. To achieve this and for sustainable control, transmission control strategies focusing on snails should become the next priority in the two most seriously affected regions surrounding the Dongting and Poyang Lakes. Whereas in the lower reaches of the Yangtze River, WBLP's morbidity control strategy should be continued for some time until the snails in the upriver provinces are well controlled.
|
10.1371/journal.pcbi.1005380 | A model for brain life history evolution | Complex cognition and relatively large brains are distributed across various taxa, and many primarily verbal hypotheses exist to explain such diversity. Yet, mathematical approaches formalizing verbal hypotheses would help deepen the understanding of brain and cognition evolution. With this aim, we combine elements of life history and metabolic theories to formulate a metabolically explicit mathematical model for brain life history evolution. We assume that some of the brain’s energetic expense is due to production (learning) and maintenance (memory) of energy-extraction skills (or cognitive abilities, knowledge, information, etc.). We also assume that individuals use such skills to extract energy from the environment, and can allocate this energy to grow and maintain the body, including brain and reproductive tissues. The model can be used to ask what fraction of growth energy should be allocated at each age, given natural selection, to growing brain and other tissues under various biological settings. We apply the model to find uninvadable allocation strategies under a baseline setting (“me vs nature”), namely when energy-extraction challenges are environmentally determined and are overcome individually but possibly with maternal help, and use modern-human data to estimate model’s parameter values. The resulting uninvadable strategies yield predictions for brain and body mass throughout ontogeny and for the ages at maturity, adulthood, and brain growth arrest. We find that: (1) a me-vs-nature setting is enough to generate adult brain and body mass of ancient human scale and a sequence of childhood, adolescence, and adulthood stages; (2) large brains are favored by intermediately challenging environments, moderately effective skills, and metabolically expensive memory; and (3) adult skill is proportional to brain mass when metabolic costs of memory saturate the brain metabolic rate allocated to skills.
| Complex cognition and relatively large brains occur in a diversity of mammal, bird, and fish species among others, and there is a large number of mostly verbal hypotheses to explain what causes their evolution in certain species but not others. However, these hypotheses have scarcely exploited the power of formulating them in mathematical terms, which has been very useful to understand the evolution of other traits. To address this issue, we formulate a mathematical model that allows incorporating many of those hypotheses and that can be used to obtain predictions for how much and how fast the brain should grow depending on the hypothesis assumed. We apply the model to humans in a setting where each individual must extract energy from the environment alone (e.g., by hunting or cooking) but possibly with its mother’s help when young (“me vs nature”). We find that a me-vs-nature setting can be enough to produce a variety of human features, including large brain sizes and an adolescence life stage. Our model can be used to compare hypotheses promoting brain evolution, such as harsh environments or plentiful social interactions.
| Complex cognitive abilities and relatively large brains are distributed across a variety of taxa [1], and there is a large number of primarily verbal hypotheses proposed to explain such diversity. Leading hypotheses suggest that various ecological and social challenges favor enhanced cognition or relatively large brains [1–9]. Empirical tests for these hypotheses regularly involve assessment of correlations, for instance, between diet quality or group size with cognitive ability or proxies thereof [10–17]. A complementary approach has been via functional studies; for example, behavioral experiments in humans find refined cognitive skills for social rather than general function [6, 18], and brain imaging also in humans has identified various brain regions specialized for social interaction [19, 20]. More recently, studies have addressed more directly the causes of large-brain evolution via phylogenetic analyses, artificial selection experiments, and genomic patterns of selection [21–25]. However, as hypotheses have remained mostly verbal, understanding of brain evolution may benefit from mathematical formalization to be used in synergy with empirical research, which can reveal features that are otherwise difficult to identify given biological complexity [26, 27]. In addition, mathematical models that can make quantitative rather than essentially qualitative predictions could facilitate obtaining contrasting predictions from the diversity of hypotheses, thereby sharpening their tests [28, 29]. Thus, here we formulate an evolutionary model that can be used to ask what quantitative and qualitative predictions arise from the various hypotheses for complex-cognition or large-brain evolution once these hypotheses are expressed in mathematical form.
A successful modeling approach in evolutionary research is life history theory [30–35]. Life history theory considers decisions regarding resouce allocation that individuals make over their lifespan, taking into account tradeoffs between competing ends (e.g., current vs future reproduction, number vs size of offspring, and growth vs reproduction) [30, 33, 34]. Such tradeoffs are of particular importance in complex-cognition and large-brain evolution because the brain uses copious amounts of energy that could be used for other functions [36–39]. While life history theory has been thoroughly used to explain the distribution of a variety of traits (e.g., [40–44]), it has remained relatively underdeveloped for cognition and brain (but see [29, 45, 46]). A first barrier is that mathematical modeling of brain evolution must describe how the brain impacts fitness without being overwhelmed by brain mechanistic details and at the same time it should consider enough mechanistic details to be able to make testable predictions. Existing models have described brain’s impact on fitness as facilitating energy acquisition from the environment allowing this energy to be used to increase survival [46], as facilitating energy production and/or decreasing the probability of being scrounged by others [47], as increasing offspring survival via parental care despite increasing mortality at birth [48], as increasing collaborative efficiency [49], as increasing mating ability [50], and as increasing the complexity of decision making regarding cooperation [51]. While these models have contributed to the understanding of brain evolution, a life history model is still lacking that incorporates real estimates of the metabolic costs of the brain while causally yielding quantitative predictions for brain and body size throughout ontogeny from a given brain-evolution hypothesis.
However, producing quantitative predictions that match empirical data from causal mathematical models is challenging given biological complexity. Despite this difficulty, metabolic theory has been successful at making quantitative predictions about ontogenetic body mass with the added bonus that its focus on metabolism allows using a top-down perspective without the need to describe the inner functioning of the system [52–56]. Thus, with the aim of producing quantitative predictions from given brain-evolution hypotheses, we combine elements of life history and metabolic theories to derive a mathematical model to study brain life history evolution. The model can be used to determine an individual’s optimal strategy regarding its energy allocation to the growth of its different tissues at each point in its life, which allows obtaining quantitative predictions for body and brain size under different biological settings.
Our model builds on previous approaches considering brain as embodied capital invested in fitness [46]. We consider separately the physical and functional embodied capital, the former being brain itself and the latter being skills (or cognitive abilities, knowledge, information, etc.) generated by the brain during ontogeny. Thus, our life-history approach considers skills throughout ontogeny. This implements the notion that information gained and maintained by the brain during growth should be considered when modeling brain evolution because selection for skill learning may delay maturity and thus impose tradeoffs among brain, body, and reproduction [1, 8, 45, 46, 57]. A defining feature of our model is that it assumes that some of the brain’s energetic consumption at each time is due to acquisition and maintenance of skills (or cognitive abilities, knowledge, information, etc.). In turn, we consider skills that allow overcoming energy-extraction challenges that the individual faces at each age, and by doing so the individual obtains some energetic reward. The model allows to study various biological settings depending on what poses the challenge faced at a given age and on who engages in overcoming the challenge: that is, the challenge can be posed by the non-social environment (nature), or it can be posed by social partners (them); also, the individual can engage in overcoming the challenge alone (me) or in concert with social partners (us): me (or us) vs nature (or them) [4, 5, 8, 45, 46, 49]. By applying different settings, our framework allows to investigate different brain-evolution hypotheses.
We apply our model to a baseline setting where individuals face exclusively ecological (non-social) challenges which are overcome by the individual alone (“me vs nature” [49]). This application captures basic aspects of hypotheses emphasizing ecological challenges as drivers of large-brain and complex-cognition evolution (e.g., that the non-social environment is a primary driver). Then, given that the brain consumes some of its energy to gain and maintain skills and given the various types of challenges that the individual faces at each age, we obtain a model that allows to predict how much an individual should grow its brain to obtain the energetic returns from skills. By feeding the model with parameter values for modern humans, we show how the model can yield predictions for life history stages as well as ontogenetic body and brain mass. We intend our me-vs-nature application to be compared with future applications of the model that include additional aspects of ecological-challenge hypotheses (e.g., variable environments [58]) and aspects of social-challenge hypotheses.
We consider a clonal, well-mixed, female population of large and constant size, where the environment is constant, generations are overlapping, and individuals’ age is measured in continuous time (i.e., standard demographic assumptions of models of life history evolution [32, 35, 40, 42, 59, 60]). We partition the body of each female into three types of tissues (or cells): reproductive tissue, brain tissue, and the remainder tissue, which we refer to as somatic. To have energy at each age for body growth, body maintenance, and reproduction, each female extracts energy from its environment (e.g., by locating food, or by making resources usable through cracking or cooking), possibly with the help of its mother (maternal care) and/or by interacting with other individuals in the population (e.g., through cooperative hunting or social competition for resources). To extract energy, each individual is assumed to use a number of relevant energy-extraction skills, which are produced and maintained by the brain.
We aim to determine the optimal allocation strategy of an individual’s energy budget to the growth of the different tissues throughout its lifespan, which simultaneously addresses the central life history question that asks what the optimal allocation to reproduction is ([61] p. 43; [33] p. 72; [62] p. 109; [32, 35, 40, 42, 59, 60]). An allocation strategy is here a vector of evolving traits that is a function of the individual’s age, and that determines the individual’s energy allocation to the growth of its different tissues throughout the individual’s lifespan. To analyze how selection affects the evolution of the allocation strategy, we carry out an evolutionary invasion analysis (e.g., [35, 63–65]), and thus consider that only two strategies can occur in the population, a mutant u and a resident (wild-type) v allocation strategies. As is standard in evolutionary invasion analysis, we thus seek to establish which strategy is uninvadable, that is, resistant to invasion by any alternative strategy taken from the set U of feasible allocation strategies, and which thus provides a likely final point of evolution [66–68]. From demographic assumptions we make below, it is well established [59, 68–72] that an uninvadable strategy u * satisfies
u * ∈ arg max u ∈ U R 0 ( u , u * ) , (1)
which implies that u * is a best response to itself, where
R 0 ( u , v ) = ∫ 0 T l ( t ) m ( t ) d t (2)
is the basic reproductive number (lifetime number of offspring) of a single mutant in an otherwise monomorphic resident population, and T is an age after which the individual no longer reproduces or is dead. The basic reproductive number depends on the probability l(t) that a mutant individual survives from birth until age t and on its rate m(t) of offspring production at age t with density dependence (“effective fecundity” [73], or the expected number of offspring produced at age t per unit time with density dependence), where these two vital rates may be functions of mutant and resident traits, u and v.
To determine the lifetime offspring production R0 and how it connects to the state variables (tissues and skill) and to the evolving traits, we relate brain and skill growth to vital rates, which in turn is mediated by the connection between energy extraction, metabolism, and tissue growth. We thus formally derive our model by making these connections.
Standard life history models refer to complete components of the energy budget (e.g., assimilated energy [74]). In practice, it is easier to measure heat release (metabolic rates [75]). Hence, to facilitate empirical parameter estimation, we follow the approach of [54] and formulate our life history model in terms of resting metabolic rate allocation, rather than energy budget allocation. Thus, we track how resting metabolic rate is due to growth and maintenance of different tissues, in particular the brain.
We start from the partition of the individual’s energy budget used by [55], which divides the energy budget (assimilation rate) into heat released at rest (resting metabolic rate) and the remainder (Fig 1; see [75] for details justifying this partition). The amount of energy used per unit time by an individual is its assimilation rate. Part of this energy per unit time is stored in the body (S) and the rest is the total metabolic rate, which is the energy released as heat per unit time after use. Part of the total metabolic rate is the resting metabolic rate Brest and the remainder is the energy released as heat per unit time due to activity Bact. In turn, part of the resting metabolic rate is due to maintenance of existing biomass Bmaint, and the remainder is due to production of new biomass Bsyn. We refer to Bsyn as the growth metabolic rate (Fig 1). We formulate our model in terms of allocation of growth metabolic rate Bsyn to the growth of the different tissues.
Denote by Ni(t) the number of cells of type i of a focal mutant female of age t, where i ∈ {b, r, s} corresponds to brain, reproductive, and the remainder cells which we refer to as somatic, respectively. Assume that an average cell of type i in the resting body releases an amount of heat Bci per unit time. Hence, the total amount of heat released per unit time by existing cells in the resting individual is
B maint ( t ) = N b ( t ) B c b + N r ( t ) B c r + N s ( t ) B c s , (3)
which gives the part of resting metabolic rate due to body mass maintenance [55].
Denote by N ˙ i ( t ) the time derivative of Ni(t). Assume that producing a new average cell of type i releases an amount of heat Eci. Hence, the total amount of heat released per unit time by the resting individual due to production of new cells is
B syn ( t ) = N ˙ b ( t ) E c b + N ˙ r ( t ) E c r + N ˙ s ( t ) E c s , (4)
which gives the rate of heat release in biosynthesis [55], and we call it the growth metabolic rate. From Eq (4), we have that
N ˙ i ( t ) E c i = u i ( t ) B syn ( t ) , (5)
for i ∈ {b, r, s}, where ui(t) is the fraction of the growth metabolic rate due to production of new type-i cells at time t (summing over all cell types in Eq 5 returns Eq 4). The resulting time sequence u = { u ( t ) } t = 0 T ∈ U, where u(t) = (ub(t), ur(t), us(t)), of allocations from birth to (reproductive) death is the evolving multidimensional trait in our model and U is the set of all feasible allocations strategies.
From our partitioning in Fig 1, the total amount of heat released by the resting individual per unit time at age t is
B rest ( t ) = B maint ( t ) + B syn ( t ) , (6)
which is the individual’s resting metabolic rate at age t.
Let the mass of an average cell of type i be xci for i ∈ {b, r, s}. Then, changing units from cell number to mass, the mass of tissue i at age t is
x i ( t ) = x c i N i ( t ) , (7)
and hence, using Eq (5), we have that the growth rate in mass of tissue i is
x ˙ i ( t ) = x c i N ˙ i ( t ) = x c i E c i u i ( t ) B syn ( t ) . (8)
To continue the change of cell-number units to mass, we denote the heat released for producing an average mass unit of tissue i as Ei = Eci/xci, which we assume constant with respect to time for simplicity. This gives
x ˙ i ( t ) = u i ( t ) B syn ( t ) E i . (9)
Substituting Eq (6) in Eq (9), we obtain the model’s first key equation specifying the growth rate of tissue i:
x ˙ i ( t ) = u i ( t ) B rest ( t ) - B maint ( t ) E i , (10)
where from Eqs (3) and (7), we have that
B maint ( t ) = x b ( t ) B b + x r ( t ) B r + x s ( t ) B s (11)
and the mass-specific cost of tissue maintenance is Bi = Bci/xci, which we also assume constant with respect to time for simplicity. From Eqs (10) and (11), the mass unit in which xi(t) for i ∈ {b, r, s} is measured (e.g., gram or kilogram) is determined by the mass unit in Bi and Ei.
We assume that the individual at age t has a skill level xk(t) at energy extraction, which is a quantitative variable measuring the individual’s ability to overcome challenges of energy extraction. Skill level xk(t) can be measured on a scale that is best suited for each application, for example as the individual’s number of skills at age t (e.g., as may be useful in anthropology; [81]), or as an index of performance at a series of tasks (e.g., as in comparative psychology studies; [1]). We assume that some of the brain metabolic rate is due to acquiring and maintaining energy-extraction skills [76–80]. Denote by Brest,b(t) the brain metabolic rate of the individual at age t (i.e., the heat released by the brain per unit time with the individual at rest). From energy conservation, the brain metabolic rate must equal the heat released by the brain per unit time due to brain growth and brain maintenance; that is, from Eqs (3) and (4) in mass units, the brain metabolic rate must satisfy
B rest , b ( t ) = x b ( t ) B b + x ˙ b ( t ) E b . (12)
Let sk be the fraction of brain metabolic rate allocated to energy-extraction skills, which we assume constant for simplicity. Suppose that the brain releases on average an amount of heat Ek for increasing skill level by one unit (learning cost; [76–78]). Similarly, assume that the brain releases on average an amount of heat Bk per unit time for maintaining a skill unit (memory cost; [79, 80]). For simplicity, we assume Ek and Bk to be constant with respect to time. From energy conservation, the rate of heat release by the brain due to skill growth and skill maintenance must equal the brain metabolic rate due to energy-extraction skill:
x k ( t ) B k + x ˙ k ( t ) E k = s k B rest , b ( t ) . (13)
Rearranging, we obtain the model’s second key equation specifying skill learning rate:
x ˙ k ( t ) = s k B rest , b ( t ) - x k ( t ) B k E k . (14)
In analogy with Eq (10), the first term in the numerator of Eq (14) gives the heat released due to skill learning and memory whereas the second term gives the heat released for memory. [Note that an equation for skill growth rate can be similarly derived, not in terms of allocation to skill growth and maintenance sk, but in terms of allocation to skill growth uk as for Eq (10).] As with Eq (10), from Eq (14) the unit in which xk(t) is measured is determined by the skill unit in Bk and Ek.
We now derive an expression that specifies how brain affects energy extraction in the model. We consider that energy extraction depends on the focal female’s skills but possibly also on the skills of other females in the population. To make this dependence explicit, we denote by E ( t , u , v ) the amount of energy extracted by the focal female per unit time at time t from the environment. So, using Fig 1 the energy-extraction rate is
E t , u , v = A ( t ) + surplus . (15)
E ( t , u , v ) depends on the mutant’s skill and possibly other features (control or state variables, or body mass) which ultimately depend on the mutant allocation strategy u. Additionally, the energy-extraction rate E ( t , u , v ) also depends on the skill or other features of the resident population which ultimately depend on the resident allocation strategy v. Let Emax(t) be the amount of energy that the individual obtains from the environment per unit time at age t if it is maximally successful at energy extraction (which also possibly depends on body mass). We define the energy-extraction efficiency e ( t , u , v ) at age t as the normalized energy-extraction rate at age t:
e t , u , v = E t , u , v E max ( t ) , (16)
which is thus a dimensionless energy extraction performance measure.
We also define the ratio of resting metabolic rate to energy-extraction rate as
q t , u , v = B rest ( t ) E t , u , v (17)
and, motivated by Eq (17), we define
B rest , max ( t , u , v ) = q t , u , v E max ( t ) (18a)
= B rest ( t ) e t , u , v . (18b)
From Eq (18b), we have that
B rest ( t ) = e t , u , v B rest , max ( t , u , v ) . (19)
Consequently, B rest , max ( t , u , v ) gives the resting metabolic rate when the individual is maximally successful at energy extraction.
Adult resting metabolic rate typically scales with adult body mass as a power law across all living systems [82–85], and also ontogenetically in humans to a good approximation (Fig C in S1 Appendix; but see [86]). Such scaling is empirically obtained from measurements in mostly well-fed individuals, and thus we assume that the empirically measured power-law scaling applies to the resting metabolic rate when the individual is maximally successful at energy extraction, B rest , max ( t , u , v ); that is, we assume
B rest , max ( t , u , v ) = K x B β ( t ) , (20)
where xB(t) = xb(t) + xr(t) + xs(t) is body mass at age t, β is a scaling coefficient, and K is a constant independent of body mass (both β and K may depend on the resident strategy; note that β need not be 3/4). We further assume that energy-extraction efficiency e ( t , u , v ) is independent of body mass, whereby Eqs (19) and (20) yield the model’s third key equation specifying resting metabolic rate as:
B rest ( t ) = K e t , u , v x B β ( t ) . (21) Eq (21) has two noteworthy implications. First, energy-extraction efficiency e ( t , u , v ) regulates the amount of heat that is released from tissue maintenance and synthesis: e.g., a zero energy-extraction efficiency means no heat is released at rest. Second, Eq (21) implies that an increase in body mass is accompanied by an increase in the rate of heat release Brest(t). But from Eq (15) and Fig 1 we have that E(t) = Brest(t) + S(t) + Bact(t) + surplus, which means that an increase in the rate of heat release Brest(t) due to increased body mass must be accompanied by either (1) an increase in the energy-extraction rate E(t), (2) a decrease in the rate of energy stored S(t), (3) a decrease in the rate of heat release due to activity Bact(t), or (4) a decrease in the surplus:
∂ B rest ( t ) ∂ x B = ∂ E ( t ) ∂ x B - ∂ S ( t ) ∂ x B - ∂ B act ( t ) ∂ x B - ∂ surplus ∂ x B . (22)
That is, Eq (21) specifies a physical constraint imposed by body size that requires that an increased resting metabolic rate due to a larger body mass is achieved by balancing the energy budget from the mentioned sources so that Eq (22) is satisfied. Given that a power law between resting metabolic rate and body mass is ubiquitously observed across living systems [82–85], we assume that individuals adjust their energy budget as just described (i.e., satisfy Eq 22) so that the physical constraint implied by Eq (21) is met.
Consequently, Eq (21) implies that increasing a tissue’s mass increases the resting metabolic rate which may increase the growth metabolic rate. To see this, consider the following. From Eqs (6), (21) and (11), the increase in growth metabolic rate with an increase in the mass of tissue i ∈ {b, r, s} is
∂ B rest ∂ x i - ∂ B maint ∂ x i = K e t , u , v β x B β - 1 ( t ) - B i . (23)
When Eq (23) is positive, an increase in the mass of tissue i increases the growth metabolic rate: resting metabolic rate increases more than maintenance metabolic rate, so there is a remainder due to growth. In addition, this remainder is greatest if i is the cheapest tissue to maintain (i.e., if Bi is the smallest for i ∈ {b, r, s}). Therefore, if a tissue (e.g., brain) is favored to grow, other particularly cheap-to-maintain tissues (e.g., soma) may be favored to grow if they increase the growth metabolic rate, which reflects an increase in the energy available for growth given that the individual balances its energy budget as indicated by Eq (22).
The expression for the resting metabolic rate (Eq 21) closes the model from a metabolic point of view, since after substituting Eq (21) in Eq (10) (and using Eqs 11, 12, 14), the ontogenetic dynamics of the brain, reproductive, and somatic tissue mass, xb(t), xr(t), and xs(t), and of skill level, xk(t), are expressed in terms of such state variables, of empirically estimable parameters, and on the evolving traits (mutant u and resident v). That is, the ontogenetic dynamics are given by
x˙i(t)=ui(t)1Ei[Ke(t,u,v) xBβ(t)−xb(t)Bb−xr(t)Br−xs(t)Bs]fori∈{ b,r,s } (24a)
x˙k(t)=1Ek{sk[xb(t)Bb +ub(t)(Ke(t,u,v) xBβ(t)−xb(t)Bb−xr(t)Br−xs(t)Bs)]−xk(t)Bk}. (24b)
The ontogenetic dynamics of the state variables xi(t) are thus non-linear.
To close the model from an evolutionary perspective and compute an optimal allocation strategy, we need expressions for how the state variables relate to the vital rates [l(t) and m(t)] in Eq (2) and expressions for the energy-extraction efficiency [e ( t , u , v )]. A large number of settings can be conceived with the model so far, both for the vital rates and energy-extraction efficiency. We focus on an application aiming at modeling human brain evolution from a baseline “me-vs-nature” setting.
The optimal strategy we obtain divides the individual’s lifespan in three broad stages: (1) a “childhood” stage, defined as the stage lasting from birth to the age at maturity tm, during which allocation to growth of reproductive tissue is zero; (2) an “adolescence” stage, defined as the stage lasting from tm to the age at adulthood ta, during which there is simultaneous allocation to growth of somatic and reproductive tissue; and (3) an “adulthood” stage, defined as the stage lasting from ta to the end of the individual’s reproductive career, during which all growth allocation is to reproductive tissue (Fig 3A). These life stages are obtained with either power or exponential competence (Fig 3A and 3E). Note that the ages at maturity tm and adulthood ta (switching times) are not parameters but an output of the model.
The obtained childhood stage, which is the only stage where there is allocation to brain growth, is further subdivided in three periods: (1a) “ante childhood”, defined here as lasting from birth to the age of brain growth onset tb0, during which there is pure allocation to somatic growth; (1b) “childhood proper”, defined here as lasting from tb0 to the age of brain growth arrest tb, during which there is simultaneous allocation to somatic and brain growth; and (1c) “preadolescence”, defined here as lasting from tb to tm, during which there is pure somatic growth. Hence, brain growth occurs exclusively during “childhood proper”. The occurrence of an “ante childhood” without brain growth disagrees with observation in humans. Two possible and particularly relevant reasons for this discrepancy may be either the absence of social interactions in this setting of the model, or the approximation of resting metabolic rate by a power law (Eq 21) which underestimates resting metabolic rate (and thus growth metabolic rate) during ante childhood (Fig C in S1 Appendix; note that although improving the fit of resting metabolic rate with body mass is straightforward, this introduces additional non-linearities that make the optimal control problem numerically more challenging and this exploration is beyond the scope of the present paper). The switching times tb0 and tb are also an output rather than parameters of the model (Fig 3A).
With the exception of the age of brain growth onset, the predicted timing of childhood, adolescence, and adulthood closely follows that observed in humans with competence being either a power or an exponential function of skill level, given their respective benchmark parameter values (Table 1). Recall that measurement units (i.e., years, kg, and MJ), excepting skill units, are real in that they result from the units of the parameter values estimated from empirical data (Table B in S1 Appendix). Hence, while using realistic metabolic costs of brain and body, the model can correctly predict major stages of human life history with accurate timing, with the exception of no brain growth allocation during ante childhood (Table 1). Later in the paper we address how these results change with changes in parameter values.
The optimal growth strategy generates the following predicted body and brain mass throughout ontogeny. For total body mass, there is fast growth during ante childhood, followed by slow growth during childhood proper, a growth spurt during preadolescence, slow growth during adolescence, and no growth during adulthood, each of which closely follows the observed growth pattern in humans (Fig 3B). Most body growth is due to somatic growth, and this results because if a tissue mass is favored to grow (e.g., brain) the energy required for its production can be made available by increasing the growth metabolic rate by producing the cheapest tissue to maintain which is the soma (Eq 23; Table B in S1 Appendix), under our assumption that individuals balance their energy budget as in Eq (22). The slow body growth during childhood proper results from the simultaneous allocation to somatic and brain growth and from the decreasing growth metabolic rate due to the increasing energetic costs of brain maintenance (Fig 3C). The growth spurt during preadolescence arises because (1) all growth metabolic rate is allocated to inexpensive somatic growth, and (2) growth metabolic rate increases due to increased metabolic rate caused by increasing, inexpensive-to-maintain somatic mass (Fig 3C). The slow growth during adolescence is due to simultaneous somatic and reproductive growth, and to the elevated costs of reproductive tissue maintenance (Fig 3C). These growth patterns result in two major peaks in growth metabolic rate (Fig 3C). While the first peak in growth metabolic rate is made possible by maternal care, the second peak is made possible by the individual’s own skills (Fig G panel D in S1 Appendix). After the onset of adulthood at ta, growth metabolic rate is virtually depleted and allocation to growth has essentially no effect on tissue growth (Fig 3C).
Whereas predicted body growth patterns are qualitatively similar with either power or exponential competence, they differ quantitatively (Fig 3B and 3F). With power competence, the predicted body mass is nearly identical to that observed in human females throughout life (Fig 3B). In contrast, with exponential competence, the predicted body mass is larger throughout life than that of human females (Fig 3F). Our exploration of the parameter space indicates that the larger body mass with exponential competence relative to power competence is robust to parameter change (see section “A large brain is favored…” below and section 8.8 in S1 Appendix).
Regarding brain mass, the model predicts it to have the following growth pattern. During ante childhood, brain mass remains static, in contrast to the observed pattern (Fig 3D). During childhood proper, brain mass initially grows quickly, then it slows down slightly, and finally grows quickly again before brain growth arrest at the onset of preadolescence (Fig 3D). Predicted brain growth is thus delayed by the obtained ante-childhood period relative to the observed brain growth in humans (Fig 3D). As previously stated, such brain growth delay may be a result of the absence of social interactions in this model setting, or an inaccuracy arising from the underestimation of resting metabolic rate during ante childhood by the power law of body mass.
Predicted brain growth patterns are also qualitatively similar but quantitatively different with power and exponential competence (Fig 3D and 3H). Adult brain mass is predicted to be larger with competence as an exponential rather than as a power function (Fig 3D and 3H). As for body mass, our exploration of the parameter space indicates that the larger brain mass with exponential competence is robust to parameter change (see section “A large brain is favored…” below and section 8.8 in S1 Appendix). Moreover, the encephalization quotient (EQ, which is the ratio of observed adult brain mass over expected adult brain mass for a given body mass) is also larger with exponential competence for the benchmark parameter values (Table 1). For illustration, with competence as a power function, the predicted adult body and brain mass approach those observed in late H. erectus (Fig 3B and 3D). In contrast, with competence as an exponential function, the predicted adult body and brain mass approach those of Neanderthals (Fig 3F and 3H). The larger EQ with exponential competence is also robust to parameter change (see section “Factors favoring a large EQ…” below and 8.8 in S1 Appendix).
The obtained optimal growth strategy predicts the following patterns for the skill level at energy-extraction throughout ontogeny. Under the same parameter values as in Fig 3, the individual’s skill level increases most during childhood and adolescence, skill level continues to increase after brain growth arrest, and skill level plateaus in adulthood (Fig 4). That is, skill growth is “determinate”, in agreement with empirical observations for food production skills (Fig 4). Yet, if memory cost Bk is substantially lower, skill level can continue to increase throughout life (i.e., skill growth is then “indeterminate”; Fig K panel E in S1 Appendix) (see Eq 14). Nevertheless, in that case, the agreement between predicted and observed body and brain mass throughout ontogeny is substantially reduced (Fig K panels B,C in S1 Appendix).
The requirement for skill growth to be determinate is that the brain metabolic rate allocated to skills [skBrest,b(t)] becomes saturated with skill maintenance [xk(t)Bk] within the individual’s life (Eq 14). Thus, skill can continue to grow after brain growth arrest if memory costs do not yet saturate the brain metabolic rate allocated to skills. When skill growth is determinate, an immediate prediction of the model is that adult skill level is proportional to adult brain mass. In particular, with determinate skill growth, the skill level that is asymptotically achieved [from Eq 24b setting x ˙ k * ( t ) = 0 and u b * ( t ) = 0] is
x ^ k = s k B b B k x b * ( t a ) , (34)
where x ^ k is the asymptotic skill level, x b * ( t a ) is the adult brain mass, sk is the fraction of brain metabolic rate allocated to energy-extraction skills, and Bb is the brain mass-specific maintenance cost. Hence, adult skill level is proportional to adult brain mass in the model (1) because of saturation with skill maintenance of the brain metabolic rate allocated to skills and (2) because adult brain metabolic rate is found to be proportional to adult brain mass given energy conservation and the assumptions on the parameters [setting x ˙ b ( t a ) = 0 in Eq (12) yields Brest,b(ta) = xb(ta)Bb]. Weak correlations between cognitive ability and brain mass have been identified across taxa including humans [14, 103–106]. Since skill level is here broadly understood to refer to cognitive abilities, this result offers an explanation for these correlations in terms of saturation of brain metabolic rate with skill maintenance (memory).
In section 8.2.3 of S1 Appendix, we relax the assumption that maternal care is independent of maternal skill by allowing for such dependence (i.e., by letting maternal facilitation in Eq 33 depend on maternal skill). Doing so yields the same results with exponential competence and a slightly faster body growth rate with power competence (section 8.2.3 in S1 Appendix). The latter difference is only quantitative and arises because the chosen benchmark maternal facilitation for newborns (φ0 = 0.6) is lower than the resulting newborn maternal facilitation (φ0 = 0.8) when it is allowed to depend on maternal skill. In addition, the predicted adult body and brain mass are virtually the same with this relaxation for either the power or exponential competence (Fig H in S1 Appendix). Hence, our results are robust to modification of our simplifying assumption of exogenous maternal care.
We now vary parameter values to assess what factors favor a large brain at adulthood in a me-vs-nature setting.
We focus on varying the parameter values that were not estimated from empirical data to also assess how they impact predictions. For the switching times in the cases when the predicted adult body and brain mass are not zero, we find that the age of brain growth onset (tb0) remains largely invariant to parameter change (Fig 5 and Figs Q, S, U, and W in S1 Appendix). Later ages at brain growth arrest (tb) and at maturity (tm) are favored by increasing environmental difficulty (increasing α; Fig 5A and Fig S panel A in S1 Appendix), decreasing skill effectiveness (decreasing γ; Fig 5B and Fig S panel B in S1 Appendix), and costlier memory (increasing Bk; Fig 5C and Fig S panel D in S1 Appendix). A later age at adulthood (ta) is also favored by environmental difficulty (Fig 5A and Fig S panel A in S1 Appendix) and skill ineffectiveness (Fig 5B and Fig S panel B in S1 Appendix), but is favored by either small or large memory costs (Fig 5C and Fig S panel D in S1 Appendix). While the switching times (tb0, tb, tm, ta) vary quantitatively as parameter values change, the qualitative occurrence of an adolescence period after childhood persists (shaded regions in Fig 5 and Figs Q, S, U, and W in S1 Appendix).
Regarding brain mass, a larger adult brain mass is favored by an increasingly challenging environment (increasing α; Eq 32), but is disfavored by an exceedingly challenging environment (Fig 6A). Environmental difficulty favors a larger brain because a higher skill level is needed for energy extraction (Eq 32), and from Eq (14) a higher skill level can be gained by increasing brain metabolic rate in turn by increasing brain mass. Thus, a large brain is favored to energetically support skill growth in a challenging environment. However, with exceedingly challenging environments, the individual is favored to reproduce early without substantial body or brain growth because it fails to gain a sufficiently high skill level to maintain its body mass as maternal care decreases with age (Fig O in S1 Appendix).
A larger adult brain is favored by moderately effective skills. When skills are ineffective at energy extraction (γ → 0; Eq 32), the brain entails little fitness benefit and fails to grow in which case the individual also reproduces without substantially growing (Fig 6B). When skill effectiveness (γ) crosses a threshold value, the fitness effect of brain becomes large enough that the brain becomes favored to grow. Yet, as skill effectiveness increases further and thus a lower skill level is sufficient for energy extraction, a smaller brain supports enough skill growth, so the optimal adult brain mass decreases with skill effectiveness (Fig 6B). Hence, adult brain mass is largest with moderately effective skill.
A larger brain is also favored by skills that are increasingly expensive for the brain to maintain (costly memory, increasing Bk), but exceedingly costly memory prevents body and brain growth (Fig 6C). Costly memory favors a large brain because then a larger brain mass is required to energetically support skill growth (Eq 14). If memory is exceedingly costly, skill level fails to grow and energy extraction is unsuccessful, causing the individual to reproduce without substantial growth (Fig 6C).
A large EQ and high adult skill are generally favored by the same factors that favor a large adult brain, but there are some differences regarding memory and learning costs. First, the memory cost has a particularly strong effect favoring a large EQ because it simultaneously favors increased brain and reduced body mass (Fig 6C and 6F). In contrast, increasing learning cost simultaneously reduces adult body and brain mass causing EQ to be invariant with respect to learning costs (Fig R panel D in S1 Appendix). That is, memory costs have a strong effect on EQ, but learning costs have little to no effect on it. Second, in contrast to its effect on EQ, increasing memory cost disfavors a high adult skill level (Fig 6F). That is, a higher EQ attained by increasing memory costs is accompained by a decrease in skill level (Fig 6C and 6F). The factors that favor a large brain, large EQ, and high skill are similar with either power or exponential competence (Fig 6 and Figs R and T in S1 Appendix). Importantly, although with the estimated parameter values the me-vs-nature setting can recover human growth patterns yielding adult body and brain mass of ancient humans, our exploration of the parameters that were not estimated from data suggests that the me-vs-nature setting cannot recover human growth patterns yielding adult body and brain mass of modern humans.
By combining elements of life history and metabolic theories, we formulated a metabolically explicit mathematical model to study brain life history evolution that can yield testable quantitative predictions from predefined settings. We applied our model to a me-vs-nature setting where individuals have no social interactions except possibly with their mothers through maternal care, but the model can be implemented to study brain evolution more generally. Surprisingly, our results for the me-vs-nature case show that this setting can be sufficient to generate major human life history stages as well as adult brain and body mass of ancient human scale, all without social interactions, evolutionary arms races in cognition triggered by social conflict, or explicit cumulative culture. Overall, we find that in the model the brain is favored to grow to energetically support skill growth, and thus a larger brain is favored when (1) competence at energy extraction has a steep dependence on skill, (2) high skill is needed for energy extraction due to environmental difficulty and moderate skill effectiveness, and (3) skills are expensive for the brain to maintain but are still necessary for energy extraction.
We find that somatic tissue grows more than the other tissues because it contributes to body mass and it is the cheapest tissue to maintain. The reason for this stems from the physical constraint imposed by resting metabolic rate being a power law of body mass (Eq 21). This constraint implies that individuals with a larger body mass release more heat and thus must compensate their energy budget either by extracting more energy, storing less energy, burning less energy in activity, or leaving less energy unassimilated (Eq 22). Given the pervasiveness across living systems of the power law between resting metabolic rate and body mass, we assumed that individuals balance their energy budget this way (as is implicitly assumed by [54]). If an individual balances its energy budget like this, then increasing any tissue’s mass can increase the energy available for growth, particularly if the tissue is cheap to maintain (so Eq 23 is positive). Thus, in our model all tissues have the implicit function of being able to increase the energy available for growth, given that individuals balance their energy budget as indicated, and this is why somatic tissue grows in our model.
The model correctly divides the individual’s lifespan into childhood, adolescence, and adulthood. The model also rightly predicts brain growth to occur only during childhood, although there is a delay in the predicted brain growth which may be due to the absence of social interactions or an underestimation of resting metabolic rate early in life by its power law approximation. Additionally, the predicted childhood stage finishes with a body growth spurt that ends at the age of maturity, as observed in human preadolescence and menarche. The model also recovers an adolescence stage with simultaneous allocation to growth and reproduction, which has been previously difficult to replicate with life history models [32]. While the timing of these predicted life stages depends on the magnitude of parameter values, their relative sequence is robust to change in the parameter values that were not estimated from data. However, the childhood-adolescence-adulthood sequence may depend on the relative magnitude of metabolic costs of maintenance and production of the different tissues (i.e., on whether (1) Bi < Bj and (2) Ei < Ej for i, j ∈ {b, r, s}). Empirically guided refinement of both parameter values and the shape of energy-extraction efficiency is expected to allow for increasingly accurate predictions [91]. Similarly, empirical data for non-human taxa should allow determining the model’s ability to predict diverse life histories and brain growth patterns [107–109].
The model also offers a metabolic explanation for correlations between adult cognitive ability and brain mass observed in inter- and intraspecific studies in birds and mammals including humans [14, 103–106]. The explanation is that when skills are costly for the brain to maintain (costly memory), the brain metabolic rate allocated to skills can become saturated by the energetic consumption of skill maintenance which causes adult skill level (or cognitive ability) to be proportional to brain mass (Eq 34). This proportionality between adult skill level and adult brain mass arises because the adult brain metabolic rate is found to be proportional to brain mass. This follows from energy conservation and our assumption that some of the brain (resting) metabolic rate is due to skill learning and memory (Eq 13), which does not make any assumptions about skill function. Consequently, the proportionality between adult skill level and brain mass is linear rather than non-linear: adding exponents to brain mass and skill level in Eqs (12) and (13) would remove the linear relationship between adult brain mass and adult skill, but it would also violate the assumption of energy conservation given the definitions of the accompanying parameters. Eq (34) also predicts that additional variation in correlations between cognitive ability and brain mass can be explained by variation in maintenance costs of brain and skill (see [110]), and by variation in brain metabolic rate allocation to skill. However, the proportionality between adult skill and adult brain mass arising from Eq (34) assumes that the fraction of brain metabolic rate allocated to the skills of interest (sk) is independent of brain mass (and similarly for Bb and Bk). So, the proportionality could break if any of these parameters depends on brain mass. Also, the model indicates that adult skill and brain mass need not be correlated since saturation with skill maintenance of the brain metabolic rate allocated to skills may not occur during the individual’s lifespan, for example if memory is inexpensive, so skill increases throughout life (Fig K panel E in S1 Appendix).
Predicted adult brain mass and skill have non-monotonic relationships with their predictor variables (Fig 6 and Figs R and T in S1 Appendix). Consequently, conflicting inferences can be drawn if predictor variables are evaluated only on their low or high ends. For instance, increasingly challenging environments favor large brains up to a point, so that exceedingly challenging environments disfavor large brains. Thus, on the low end of environmental difficulty, the prediction that increasingly challenging environments favor large brains is consistent with ecological challenge hypotheses [45, 46, 111]; yet, on the high end of environmental difficulty, the prediction that increasingly challenging environments disfavor large brains is consistent with constraint hypotheses according to which facilitation of environmental challenge favors larger brains [46, 112–114]. Counter-intuitively on first encounter, the finding that moderately effective skills are most conducive to a large brain and high skill is a consequence of the need of higher skill when skill effectiveness decreases (Fig 6B). Regarding memory cost, the strong effect of memory cost on favoring a high EQ at first glance suggests that a larger EQ than that observed in humans is possible if memory were costlier (see dashed lines in Fig 6F). However, such larger memory costs cause a reduced adult skill level (Fig 6F), and a substantial delay in body and brain growth that yields growth patterns that are inconsistent with those of humans (Figs L–N in S1 Appendix).
We have made a number of assumptions that, if modified, the predicted relationships may be qualitatively affected, as is known to occur with life history models [33]. Assumptions whose modification may qualitatively affect predictions, as suggested by previous life-history models, include (1) different forms for the mortality rate (e.g., as a function of skills) [69, 115], (2) different modes of population regulation (e.g., through survival rather than through fecundity) [69, 115], and (3) allowing population size to change [62, 115]. Implementing modifications like these would allow to assess the level of generality of our results. Additionally, our model may be used to study allocation to tissue maintenance to address senescence.
Although our model does not include numerous details relevant to human life history, such as social interactions and cummulative social learning, our results are relevant for a set of hypotheses for human-brain evolution. In particular, food processing (e.g., mechanically with stone tools or by cooking) has previously been advanced as a determinant factor in human-brain evolution as it increases energy and nutrient availability from otherwise relatively inaccessible sources [7, 116]. This is supported by archaeological evidence of mechanical food processing in early humans (1.5 mya in Kenya; [116, 117]), as well as archeological evidence of fire control in early humans with increasing certainty in younger strata (1.5 mya–130 kya; [118–120]). Our results, where individuals engage in energy extraction without considering the effects of social interactions, suggest that food processing alone could indeed have been sufficient to allow for a substantial brain expansion. In addition, food processing may help satisfy at least two of the three key conditions identified for large-brain evolution listed in the first paragraph of the Discussion. First, a shift in food-processing technology (e.g., from primarily mechanical to cooking) could create a steeper relationship between energy-extraction skill and competence by substantially facilitating energy extraction (relating to condition 1). Second, food processing (e.g., by building the required tools or lighting a fire) is a challenging feat to learn and may often fail (relating to condition 2). Yet, there are scant data allowing to judge the metabolic expense for the brain to maintain tool-making or fire-control skills (condition 3). Our results are thus consistent with the hypothesis of food processing as being a key factor in human brain expansion.
In sum, the model identifies various conditions favoring large-brain evolution, in particular steep competence with respect to skill, intermediate environmental difficulty, moderate skill effectiveness, and costly memory. As we did not consider social interactions, our application of the model cannot refute or support social brain hypotheses. However, application of our model to the social realm should allow for assessments of social hypotheses. Overall, our model is a step towards a quantitative theory of brain life history evolution yielding testable quantitative predictions as ecological, demographic, and social factors vary.
|
10.1371/journal.pbio.1001853 | Rates of Dinosaur Body Mass Evolution Indicate 170 Million Years of Sustained Ecological Innovation on the Avian Stem Lineage | Large-scale adaptive radiations might explain the runaway success of a minority of extant vertebrate clades. This hypothesis predicts, among other things, rapid rates of morphological evolution during the early history of major groups, as lineages invade disparate ecological niches. However, few studies of adaptive radiation have included deep time data, so the links between extant diversity and major extinct radiations are unclear. The intensively studied Mesozoic dinosaur record provides a model system for such investigation, representing an ecologically diverse group that dominated terrestrial ecosystems for 170 million years. Furthermore, with 10,000 species, extant dinosaurs (birds) are the most speciose living tetrapod clade. We assembled composite trees of 614–622 Mesozoic dinosaurs/birds, and a comprehensive body mass dataset using the scaling relationship of limb bone robustness. Maximum-likelihood modelling and the node height test reveal rapid evolutionary rates and a predominance of rapid shifts among size classes in early (Triassic) dinosaurs. This indicates an early burst niche-filling pattern and contrasts with previous studies that favoured gradualistic rates. Subsequently, rates declined in most lineages, which rarely exploited new ecological niches. However, feathered maniraptoran dinosaurs (including Mesozoic birds) sustained rapid evolution from at least the Middle Jurassic, suggesting that these taxa evaded the effects of niche saturation. This indicates that a long evolutionary history of continuing ecological innovation paved the way for a second great radiation of dinosaurs, in birds. We therefore demonstrate links between the predominantly extinct deep time adaptive radiation of non-avian dinosaurs and the phenomenal diversification of birds, via continuing rapid rates of evolution along the phylogenetic stem lineage. This raises the possibility that the uneven distribution of biodiversity results not just from large-scale extrapolation of the process of adaptive radiation in a few extant clades, but also from the maintenance of evolvability on vast time scales across the history of life, in key lineages.
| Animals display huge morphological and ecological diversity. One possible explanation of how this diversity evolved is the "niche filling" model of adaptive radiation—under which evolutionary rates are highest early in the evolution of a group, as lineages diversify to fill disparate ecological niches. We studied patterns of body size evolution in dinosaurs and birds to test this model, and to explore the links between modern day diversity and major extinct radiations. We found rapid evolutionary rates in early dinosaur evolution, beginning more than 200 million years ago, as dinosaur body sizes diversified rapidly to fill new ecological niches, including herbivory. High rates were maintained only on the evolutionary line leading to birds, which continued to produce new ecological diversity not seen in other dinosaurs. Small body size might have been key to maintaining evolutionary potential (evolvability) in birds, which broke the lower body size limit of about 1 kg seen in other dinosaurs. Our results suggest that the maintenance of evolvability in only some lineages explains the unbalanced distribution of morphological and ecological diversity seen among groups of animals, both extinct and extant. Important living groups such as birds might therefore result from sustained, rapid evolutionary rates over timescales of hundreds of millions of years.
| Much of extant biodiversity may have arisen from a small number of adaptive radiations occurring on large spatiotemporal scales [1]–[3]. Under the niche-filling model of adaptive radiation, ecological opportunities arise from key innovations, the extinction of competitors, or geographic dispersal [1],[4],[5]. These cause rapid evolutionary rates in ecologically relevant traits, as diverging lineages exploit distinct resources. Rates of trait evolution then decelerate as niches become saturated, a pattern that has been formalised as the “early burst” model (e.g., [6],[7]).
Most phylogenetic studies of adaptive radiations focus on small scales such as island radiations and other recently diverging clades, including Anolis lizards, cichlid fishes, and geospizine finches [2],[6],[8]–[10]. Detailed study of these model systems has demonstrated the importance of ecological and functional divergence as drivers of speciation early in adaptive radiations (e.g., [11],[12]). Surprisingly though, early burst patterns of trait evolution receive only limited support from model comparison approaches for these and other adaptive radiations occurring in geographically restricted areas and on short timescales (<50 million years [Ma]; most <10 Ma) [6] (but see [13],[14]).
Studies of morphological evolution on longer timescales, unfolding over 100 Ma or more, are central to establishing whether niche-filling or early burst patterns of trait evolution are important evolutionary phenomena on large phylogenetic scales. A small number of recent studies quantified patterns of trait evolution on large scales using neontological phylogenies. For example, diversification rates and morphological rates are positively correlated in actinopterygians [15] (∼400 Ma); rapid rates of both morphological and molecular evolution occur on deep, Cambrian, nodes of the arthropod tree of life [16] (∼540 Ma); and the early evolution of placental mammals was characterised by rapid rates of diversification [17] (100–65 Ma) and perhaps body size evolution [18] (but see [19]).
However, even the largest neontological studies [15]–[18],[20],[21] are limited to explaining the rise of important extant groups. A more complete characterisation of macroevolutionary processes on long timescales should also explain the ascent and demise of important extinct groups (e.g., [22]), which in fact represent most of life's diversity. Substantial evidence for the dynamics of past adaptive radiations might have been erased from the neontological archive, and macroevolutionary models for extinct or declining/depauperate clades may be tested most effectively using deep time data from the fossil record [23],[24].
Palaeontologists often quantify patterns of morphological radiation using time series of disparity (e.g., [25],[26]). However, few phylogenetic studies including fossil data have attempted to explain patterns of morphological radiation in large clades on timescales >100 Ma, and most have individually targeted either the roots of exceptional modern clades such as birds or mammals (e.g., [19],[27],[28]) or extinct/depauperate clades (e.g., [29]–[31]; studies based on discrete characters). Thus, patterns of morphological evolution in major extinct clades, and their links to successful modern clades, are not well understood.
Non-avian dinosaurs are an iconic group of terrestrial animals. They were abundant and ecologically diverse for most of the Mesozoic, and included extremely large-bodied taxa that challenge our understanding of size limits in terrestrial animals [32]. The first dinosaurs appeared more than 230 Ma ago in the Triassic Period, as small-bodied (10–60 kg), bipedal, generalists. By the Early Jurassic (circa 200 Ma), they dominated terrestrial ecosystems in terms of species richness [33],[34], and Cretaceous dinosaurs (145–66 Ma) had body masses spanning more than seven orders of magnitude (Figure 1A). Non-avian dinosaurs became extinct at the catastrophic Cretaceous/Paleogene (K/Pg) boundary event, at or near the peak of their diversity [35],[36]. In contrast, extant dinosaurs (neornithine birds) comprise around 10,000 species and result from one of the most important large-scale adaptive radiations of the Cenozoic [3],[21].
The proposed drivers of early dinosaur diversification are controversial. Although various causal factors have been suggested to underlie a presumed adaptive radiation, few studies have tested the predictions of niche-filling models, and these have yielded equivocal results. An upright, bipedal gait, rapid growth, and possible endothermy have been proposed as key innovations of Triassic dinosaurs (reviewed by [34]), and mass extinctions during the Triassic/Jurassic boundary interval removed competing clades, perhaps leading to ecological release and rapid rates of body size evolution in Early Jurassic dinosaurs [37] (but see [34]). However, quantitative studies using body size proxies [34] and discrete morphological characters [33] have found only weak support for the niche-filling model during early dinosaur evolution, instead favouring gradualistic evolutionary rates. These studies focussed on the Late Triassic–Early Jurassic, so it is unclear whether Early Jurassic dinosaur evolution differed from later intervals (consistent with radiation following a mass extinction), or how the Middle Jurassic–Cretaceous radiation of birds and their proximate relatives relates to overall patterns of dinosaur diversification.
We used phylogenetic comparative methods [6],[14],[38],[39] to analyse rates of dinosaur body mass evolution (Materials and Methods; Appendix S1). For this study, we compiled a large dataset of dinosaur body masses (441 taxa; Dataset S1) using the accurate scaling relationship of limb robustness (shaft circumference) derived from extant tetrapods [40] (Appendix S1; Dataset S1). Body mass affects all aspects of organismal biology and ecology (e.g., [41],[42]), including that of dinosaurs (e.g., [43]–[45]). Because of its relationship with animal energetics and first-order ecology, understanding the evolution of body mass is fundamental to identifying the macroevolutionary processes underlying biodiversity seen in both ancient and modern biotas. Therefore, by studying body mass evolution, we assess the broad pattern of niche filling in the assembly of dinosaur diversity through 170 Ma of the Mesozoic.
In many hypotheses of adaptive radiation, ecological speciation is an important process generating both morphological and taxonomic diversity (e.g., [2]; but see [46]), according to which ecological differentiation is essentially simultaneous with lineage splitting [12]. In consequence, many large-scale studies of adaptive radiation have focussed on diversification rates (e.g., [17],[21],[47]). A correlation between diversification rates and morphological rates is consistent with adaptive radiation (e.g., [15]). However, even when this can be demonstrated, the occurrence of ecological speciation is difficult (perhaps impossible) to test in clades even only a few Ma old [48]. Methods for estimating diversification rates on non-ultrametric trees (e.g., those including deep time data) have recently become available [49]. However, these methods require accurate estimates of sampling probability during discrete time intervals, and it is not clear that it is possible to obtain such estimates from the dinosaur fossil record, which contains many taxa known only from single occurrences. Therefore, our study focuses on the predictions of niche-filling models of morphological evolution during adaptive radiation, as done in some previous studies (e.g., [6],[13]).
Most of the earliest dinosaurs weighed 10–35 kg (Figure 1); Herrerasaurus was exceptionally large at 260 kg. Maximum body masses increased rapidly to 1,000–10,000 kg in sauropodomorphs, with especially high masses in early sauropods such as Antetonitrus (5,600 kg; Norian, Late Triassic) and Vulcanodon (9,800 kg; Early Jurassic), whereas minimum body masses of 1–4 kg were attained by Late Triassic ornithischians and theropods (Figure 1). Jurassic Heterodontosauridae (∼0.7 kg [50]), Middle Jurassic and younger Paraves (e.g., Epidexipteryx, 0.4 kg; Anchiornis, 0.7 kg), and Cretaceous Avialae (birds: 13–16 g to 190 kg [51]) extended this lower body size limit (Table 1). Archaeopteryx weighed 0.99 kg (the largest, subadult specimen [52]) and the Cretaceous sauropod Argentinosaurus weighed approximately 90,000 kg (Table 1). Our full set of mass estimates is available in Dataset S1 and a summary is presented in Table 1.
Our node height tests indicate that evolutionary rate estimates at phylogenetic nodes (standardised phylogenetically independent contrasts [39]) vary inversely with log-transformed stratigraphic age for most phylogenies (Figure 2). This relationship is significant (based on robust regression [14],[53]) for most phylogenies of non-maniraptoran dinosaurs, and for ornithischians and non-maniraptoran theropods when analysed separately (Figure 2B). This result is weakened, and becomes non-significant, when Triassic nodes are excluded (Figure S1).
Declining evolutionary rates through time are not found in any analyses including maniraptorans. Indeed, when maniraptorans are added to analyses of Dinosauria, a burst of high nodal rate estimates is evident in lowess lines spanning the Middle Jurassic–Early Cretaceous interval of maniraptoran diversification (Figure 2A). Maniraptorans have a weakly positive (non-significant) relationship between evolutionary rates and body mass, and do not show diminishing evolutionary rates through time (Figure 2B–C). This contrasts with non-maniraptoran dinosaurs, in which evolutionary rates vary inversely with body mass (Figure 2C).
Maximum-likelihood models [6],[38] were fitted to phylogenies calibrated to stratigraphy using the “equal” and “mbl” (minimum branch length) methods (see Materials and Methods), and complement the results of our node height tests in showing support for early burst models only in analyses excluding Maniraptora (Table 2; Figure S2). Note, however, that the maximum-likelihood method has less statistical power to detect early burst patterns than does the node height test when even a small number of lineages escape from the overall pattern of declining rates through time [14]. Two models that predict saturation of trait variance through a clade's history were commonly supported in our analyses: the early burst model of exponentially declining evolutionary rates through time, and the Ornstein–Uhlenbeck (OU) model of attraction to a “trait optimum” value. Other models (e.g., Brownian motion, stasis) had negligible AICc weights in all or most (directional trend model) analyses (AICc is Akaike's information criterion for finite sample sizes).
Early burst models received high AICc weights for analyses of ornithischians, non-maniraptoran theropods, and non-maniraptoran dinosaurs when using the “equal” branch length calibration method (Table 2; Figure S2). Early burst models had comparable AICc weights to Ornstein–Uhlenbeck models for sauropodomorphs when using the “equal” branch length calibration method, and for ornithischians and non-maniraptoran theropods when using the “mbl” method. Early burst models had generally lower AICc weights for non-maniraptoran dinosaurs and for sauropodomorphs when using the “mbl” branch length calibration method (Table 2; Figure S2). Support from some phylogenies for Ornstein–Uhlenbeck models of attraction to a large body size optimum from small ancestral body sizes [54],[55] in ornithischians [56], non-maniraptoran theropods, and especially sauropodomorphs and non-maniraptoran dinosaurs (Table 2; Figure S2), suggests the occurrence of Cope's rule in dinosaurs. All phylogenies provide strong support for this pattern in maniraptorans (Table 2).
Exceptionally high rates at individual nodes in our phylogenies were identified as down-weighted datapoints in robust regression analyses [14],[53]. Five sets of exceptional nodes in the Triassic–Early Jurassic represent rapid evolutionary shifts from primitive masses around 10–35 kg to large body masses in derived sauropodomorphs (>1,000 kg), armoured ornithischians (Thyreophora; Figure 1B) and theropods (Herrerasaurus, and derived taxa such as Liliensternus (84 kg) and Dilophosaurus (350 kg)), and to smaller body sizes in heterodontosaurid ornithischians (Figure 3; Table 3). Rapid body size changes were rare in later ornithischians and sauropodomorphs, which each show only one exceptional Jurassic node, marking the origin of body sizes greater than 1,000 kg in derived iguanodontians, and of island dwarfism in the sauropod Europasaurus [57]. By contrast, up to six exceptional Jurassic nodes occur in theropod evolution, with especially high contrasts at the origins of body sizes exceeding 750 kg in Tetanurae, and marking phylogenetically nested size reductions on the line leading to birds: in Coelurosauria (e.g., Ornitholestes, 14 kg; Zuolong, 88 kg) and in Paraves, which originated at very small body masses around 1 kg [58].
The contrast between theropods and other dinosaurs is even greater in the Cretaceous, when no exceptional nodes occur in Sauropodomorpha, and only two in Ornithischia: at the origins of large-bodied Ceratopsidae and island dwarf rhabdodontid iguanodontians (e.g., Mochlodon [59]). At least nine shifts occurred during the same interval of theropod evolution, including seven in maniraptorans (Figure 3; Table 3).
Patterns of dinosaur body size evolution are consistent with the niche-filling model of adaptive radiation [1],[4],[6]. Early dinosaurs exhibit rapid background rates of body size evolution, and a predominance of temporally rapid, order-of-magnitude shifts between body size classes in the Triassic and Early Jurassic. These shifts reflect radiation into disparate ecological niches such as bulk herbivory in large-bodied sauropodomorphs (e.g., [60]) and thyreophoran ornithischians, herbivory using a complex masticating dentition in small-bodied heterodontosaurids (e.g., [61],[62]), and increasing diversity of macropredation in large theropods (Table 3). Subsequently, rates of body size evolution decreased, suggesting saturation of coarsely defined body size niches available to dinosaurs in terrestrial ecosystems, and increasingly limited exploration of novel body size space within clades.
The early burst pattern of dinosaurian body size evolution is substantially weakened when Triassic data are excluded (Figure S1). This suggests that key innovations of Triassic dinosaurs (e.g., [63],[64]), and not the Triassic/Jurassic extinction of their competitors [37], drove the early radiation of dinosaur body sizes [34]. Indeed, phylogenetic patterns indicate that many basic ecomorphological divergences occurred well before the Triassic/Jurassic boundary.
It is not clear which innovations allowed dinosaurs to radiate [34], or whether the pattern shown here was part of a larger archosaurian radiation [65]. However, the evolution of rapid growth rates may have been important [64], especially in Sauropodomorpha [66], and the erect stance of dinosaurs and some other archosaurs [34] might have been a prerequisite for body size diversification via increased efficiency/capacity for terrestrial weight support [63].
Maniraptoran theropods are an exception to the overall pattern of declining evolutionary rates through time: exhibiting numerous instances of exceptional body size shifts, maintaining rapid evolutionary rates, and generating high ecological diversity [67],[68], including flying taxa. Although a previous study found little evidence for directional trends of body size increase in herbivorous maniraptoran clades [69], this does not conflict with our observation that some body size shifts in maniraptorans (and other coelurosaurs) coincide with the appearance of craniodental, or other, evidence for herbivory (Table 3; e.g., [67],[68],[70]).
Much of our knowledge of Late Jurassic and Early Cretaceous maniraptorans comes from a few well-sampled Chinese Lagerstätten, such as the Jehol biota. Without information from these exceptional deposits, we would have substantially less knowledge of divergence dates and ancestral body sizes among early maniraptorans. However, this is unlikely to bias comparisons between maniraptorans and other groups of dinosaurs for two reasons: (1) these deposits provide equally good information on the existence and affinities of small-bodied taxa in other clades, such as Ornithischia; and (2) exceptional information on early maniraptoran history should bias analyses towards finding an early burst pattern in maniraptorans. Inference of high early rates in Maniraptora would be more likely, due either to concentration of short branch durations at the base of the tree (especially using the “mbl” stratigraphic calibration method), or observation of additional body size diversity at the base of the tree that would remain undetected if sampling was poor. We cannot speculate as to the effects on our analyses of finding comparable Lagerstätten documenting early dinosaur history. However, there is currently little positive evidence that the general patterns of body size evolution documented here are artefactual.
Many stratigraphically younger dinosaurs, especially non-maniraptorans, exhibit large body size and had slow macroevolutionary rates, possibly due to scaling of generation times (e.g., [71],[72]). Scaling effects are observed across Dinosauria, but show substantial scatter (non-significant; Figure 2C) within Ornithischia and Sauropodomorpha, consistent with previous suggestions that scaling effects should be weak in dinosaurs because of the life history effects of oviparity [73]. Small dinosaurs (10–50 kg) had the highest evolutionary rates, and rates attenuated only weakly, or not at all, at sizes below 10 kg (Figure S3). This might have been key to maniraptoran diversification from small-bodied ancestors, and also explains the origins of fundamentally new body plans and ecotypes from small-bodied ancestors later in ornithischian history (Iguanodontia, Ceratopsidae; Figure 1).
Maniraptora includes Avialae, the only dinosaur clade to frequently break the lower body size limit around 1–3 kg seen in other dinosaurs. It is likely that more niches are available to birds (and mammals) around 100 g in mass [41],[74], so obtaining smaller body sizes might have contributed to the ecological radiation of Mesozoic birds (e.g., [27],[75]). If the K/Pg extinction event was ecologically selective, vigorous ecological diversification may have given maniraptoran lineages a greater chance of survival: Avialae was the only dinosaurian clade to survive, perhaps because of the small body sizes of its members. Although the fossil record of birds is inadequate to test hypotheses of K/Pg extinction selectivity, it is clear that smaller-sized squamates and mammals selectively survived this event [76],[77]. Therefore, our results suggest that rapid evolutionary rates within Maniraptora paved the way for a second great adaptive radiation of dinosaurs in the wake of the K/Pg extinction event: the diversification of neornithine birds [21].
Our findings complement recent studies of diversification rates in the avian crown group [3],[21], and suggest that birds, the most speciose class of tetrapods, arose from a long evolutionary history of continual ecological innovation. Our most striking finding is of sustained, rapid evolutionary rates on the line leading to birds (i.e., in maniraptorans) for more than 150 Ma, from the origin of dinosaurs until at least the end of the Mesozoic. Rates of evolution declined through time in most dinosaurs. However, this early burst pattern, which characterises the niche-filling model of adaptive radiation [6],[7], does not adequately describe evolution on the avian stem lineage. The recovered pattern of sustained evolutionary rates, and the repeated generation of novel ecotypes, suggests a key role for the maintenance of evolvability, the capacity for organisms to evolve, in the evolutionary success of this lineage. Evolvability might have also played a central role in the evolution of other major groups such as crustaceans [78] and actinopterygians [15], supporting its hypothesised importance in organismal evolution [79].
Rapid evolutionary rates observed during the early evolutionary history of Dinosauria, which decelerated through time in most subclades, indicate that much of the observed body size diversity of dinosaurs was generated by an early burst pattern of trait evolution. However, this pattern becomes difficult to detect when data from early dinosaurian history are not included in analyses (Figure S1), consistent with the observation that deep time data improve model inference in simulations [24]. The pruning of lineages by extinction might also overwrite the signals of ancient adaptive radiation in large neontological datasets. For example, Rabosky et al. [15] recovered slow evolutionary rates at the base of the actinopterygian tree, but the fossil record reveals substantial morphological and taxonomic diversity of extinct basal actinopterygian lineages [80],[81]. Although it has not yet been tested quantitatively, this diversity might have resulted from early rapid rates across Actinopterygii, as observed here across Dinosauria.
If our results can be generalised, they suggest that the unbalanced distribution of morphological and ecological diversity among clades results from the maintenance of rapid evolutionary rates over vast timescales in key lineages. These highly evolvable lineages may be more likely to lead to successful modern groups such as birds, whereas other lineages show declining evolutionary rates through time. Declining evolutionary rates in dinosaurian lineages off the line leading to birds indicate large-scale niche saturation. This might signal failure to keep pace with a deteriorating (biotic) environment (the Red Queen hypothesis [82],[83]), with fewer broad-scale ecological opportunities than those favouring the early radiation of dinosaurs. There is strong evidence for Red Queen effects on diversification patterns in Cenozoic terrestrial mammals [22], and it is possible that a long-term failure to exploit new opportunities characterises the major extinct radiations of deep time (and depauperate modern clades), whether or not it directly caused their extinctions.
We used phylogenetic comparative methods to analyse rates of dinosaur body mass evolution [6],[14],[38],[39] (Appendix S1). Body mass, accompanied by qualitative observations (Table 3), was used as a general ecological descriptor. Body mass was estimated for all dinosaurs for which appropriate data were available (441 taxa; Dataset S1) using the empirical scaling relationship of limb robustness (stylopodial circumference) with body mass, derived from extant tetrapods [40] (Appendix S1). We analysed log10-transformed data (excluding juveniles), which represent proportional changes in body mass.
Stylopodial shaft circumferences are infrequently reported in the literature, so many were taken from our own measurements, or were calculated from shaft diameters (Appendix S1). Previous large datasets of dinosaurian masses were based on substantially less accurate methods, using the relationship between linear measurements (e.g., limb bone lengths) and volumetric models of extinct dinosaurs ([84]–[86]; reviewed by [40]).
Quantitative macroevolutionary models were tested on composite trees compiled from recent, taxon-rich cladograms of major dinosaur groups (Appendix S1; Figure S4, Figure S5, Figure S6, Figure S7). Phylogenetic uncertainty was reflected by analysing alternative topologies and randomly resolved polytomies (Appendix S1). Tip heights and branch durations were stratigraphically calibrated, and zero-length branches were “smoothed” using two methods: (1) by sharing duration equally with preceding non-zero length branches (the “equal” method [87]); and (2) by imposing a minimum branch length of 1 Ma (the “mbl” method [88]).
We used maximum-likelihood model comparison [6],[38] and “node height” test [14],[39] methods (Appendix S1) to test the prediction of the niche-filling hypothesis: that rates of morphological evolution diminish exponentially through time after an adaptive radiation [1],[2],[4]. The node height test treats standardised independent contrasts [89] as nodal estimates of evolutionary rate [39] and tests for systematic deviations from a uniform rate Brownian model, using regression against log-transformed geological age (robust regression [14],[53]). We also regressed standardised contrasts against nodal body mass estimates (a proxy for generation time and other biological processes that might influence evolutionary rates). As well as testing for a “background” model of declining evolutionary rates through time, robust regression identifies and down-weights single nodes deviating substantially from the overall pattern [14],[53]. These nodes represent substantial, temporally rapid, niche-shift events [14], following the macroecological principle that organisms in different body size classes inhabit different niches and have different energetic requirements [41]. We used lowess lines to visualise non-linear rate variation with time and body mass.
Exponentially declining rates of evolution through time, predicted by the niche-filling model of adaptive radiation [1]–[3], were also tested by comparing the fit of an early burst model [6],[7] with other commonly used models: Brownian motion, directional evolution (“trend”), the Ornstein–Uhlenbeck model of evolution attracted to an optimum value, and stasis (“white noise”) [38],[56],[90] (Appendix S1). Explicit mathematical models of trait evolution on our phylogenies were fitted using the R packages GEIGER version 1.99–3 [91] and OUwie version 1.33 [55] (for Ornstein–Uhlenbeck (OU) models only), and compared using AICc [92],[93]. Unlike GEIGER, OUwie allows estimation of a trait optimum (θ) that is distinct from the root value (Z0) in OU models. Values from GEIGER and OUwie are directly comparable: identical log likelihood, AICc, and parameter estimates are obtained for test datasets when fitting models implemented in both packages (Brownian motion in all instances; and OU models when θ = Z0 for ultrametric trees); although note that comparable standard error values entered to the OUwie function of OUwie 1.33 are the square of those entered to the fitContinuous function of Geiger 1.99–3. The algorithm used to fit OU models in GEIGER 1.99–3 is inappropriate for non-ultrametric trees (personal communication, Graham Slater to R. Benson, December 2013). This problem is specific to OU models implemented by GEIGER 1.99–3, and does not affect the other models that we tested. GEIGER 1.99–3 fits models of trait evolution using independent contrasts, after rescaling the branch lengths of the phylogenetic tree according to the model considered [7]. For all models, except the OU model in the case of non-ultrametric trees, the covariance between two taxa i and j can be written as a function of the path length sij shared between the two taxa (e.g., [6],[7]). The tree can thus easily be rescaled by applying this function to the height of each node before computing independent contrasts. In the case of the OU model, the covariance between two taxa i and j is a function of both the shared (pre-divergence) portion of their phylogenetic history and the non-shared (post-divergence) portion [54]. In the case of an ultrametric tree, the non-shared portion can also be written as a function of sij (it is simply the total height T of the tree, minus sij [90],[94]), and the corresponding scaling function can be applied to the tree (this is what is performed in GEIGER 1.99.3). However, in the case of a non-ultrametric tree, the post-divergence portion of the covariance cannot be written as a function of sij, so there is no straightforward scaling function to apply. Instead, it is necessary to fit the model by maximum likelihood after computing the variance–covariance matrix. This is what is implemented in OUwie, and now in GEIGER 2.0 (personal communication, Josef Uyeda to R. Benson, January 2014).
Our data and analytical scripts are available at DRYAD [95].
|
10.1371/journal.pgen.1006566 | KLK5 and KLK7 Ablation Fully Rescues Lethality of Netherton Syndrome-Like Phenotype | Netherton syndrome (NS) is a severe skin disease caused by the loss of protease inhibitor LEKTI, which leads to the dysregulation of epidermal proteases and severe skin-barrier defects. KLK5 was proposed as a major protease in NS pathology, however its inactivation is not sufficient to rescue the lethal phenotype of LEKTI-deficient mice. In this study, we further elucidated the in vivo roles of the epidermal proteases in NS using a set of mouse models individually or simultaneously deficient for KLK5 and KLK7 on the genetic background of a novel NS-mouse model. We show that although the ablation of KLK5 or KLK7 is not sufficient to rescue the lethal effect of LEKTI-deficiency simultaneous deficiency of both KLKs completely rescues the epidermal barrier and the postnatal lethality allowing mice to reach adulthood with fully functional skin and normal hair growth. We report that not only KLK5 but also KLK7 plays an important role in the inflammation and defective differentiation in NS and KLK7 activity is not solely dependent on activation by KLK5. Altogether, these findings show that unregulated activities of KLK5 and KLK7 are responsible for NS development and both proteases should become targets for NS therapy.
| Netherton syndrome (NS) is a genetic skin disorder caused by the loss of protease inhibitor LEKTI, which leads to the dysregulation of epidermal proteases and severe skin-barrier defects. In this work, we aimed to explore the molecular mechanisms underlying this disease using a novel mutant mouse model for NS, which is based on mimicking a causative mutation known from human patients. This novel model reproduces the symptoms of NS and thus provides a useful tool to study the NS pathology in a complex in vivo environment. Most importantly, by combination of this NS-mouse model with mutant mice individually or simultaneously deficient for proteases KLK5 and KLK7, we elucidated the complex proteolytic networks that are dysregulated in the absence of LEKTI. We show that although the single ablation of KLK5 or KLK7 is not sufficient to rescue the lethal effect of LEKTI-deficiency, simultaneous deficiency of both KLKs completely rescues the epidermal barrier and the postnatal lethality. Our results also provide novel insights into the roles of KLK5 and KLK7 in the inflammation and differentiation defects that are associated with NS. Based on these findings, we propose that both, KLK5 and KLK7 should become targets for NS therapy.
| Netherton syndrome (NS) is a life-threatening autosomal recessive disorder that affects approximately one in 200 000 newborn children [1,2]. Newborns suffering from NS exhibit congenital ichthyosiform erythroderma with scaly and peeling skin, resulting in severe disruption of epidermal barrier, which in some cases is fatal. These conditions may improve with age and older patients often show less severe ichthyosis exhibiting erythematous plaques with double-edged scales at the periphery [1–3]. The hair of NS patients is usually thin, fragile and the patients often develops “bamboo hair”, a hair shaft defect where the distal part of the hair shaft is invaginated into its proximal part [4]. NS may be also associated with growth retardation, asthma, food allergies, and elevated serum levels of IgE [1,5].
NS is caused by mutations in SPINK5 gene (serine protease inhibitor Kazal-type 5) that encodes LEKTI (lympho-epithelial Kazal-type related inhibitor), an inhibitor of serine proteases expressed in the epidermis and other stratified epithelia [6]. Full length LEKTI consists of 15 inhibitory domains (D1—D15) and upon synthesis undergoes proteolytic processing into multiple bioactive fragments containing one to six domains with distinct inhibitory specificities [7,8]. LEKTI has been reported to inhibit several proteases including plasmin, trypsin, subtilisin A, cathepsin G, elastase, caspase-14 [9–11] as well as members of the family of kallikrein-related peptidases (KLK), mainly KLK5, KLK7 and KLK14 [12–14]. Unregulated activity of KLK5 and possibly also KLK7 is considered a major source of pathology in NS. Spink5 deficient mice show increased proteolytic activities of KLK5 and KLK7 [15], which corresponds to elevated tryptic and chymotryptic activities described in NS patients [16,17]. KLK5 also initiates a proteolytic cascade by proteolytic activation of KLK7 and KLK14, that leads to degradation of corneodesmosomal proteins desmoglein1 (DSG1), desmocollin1 (DSC1), and corneodesmosin (CDSN)[18]. Premature degradation of corneodesmosomes results in detachment of the stratum corneum (SC) and disruption of the epidermal barrier in NS patients [17]. Upregulated proteolytic activity can further contribute to skin barrier defects by abnormal processing of profilaggrin, a precursor protein which is proteolytically converted into physiologically active filaggrin monomers. Filaggrin is one of key players in maintaining skin hydratation and water retention of the epidermis [19]. Recently, KLK5 was shown to promote profilaggrin processing either via proteolytic activation of elastase 2, which cleaves filaggrin precursor proteins [20] or by direct degradation of profilaggrin [21]. In addition, previous studies using mouse models and human NS patients suggest that the unregulated activity of KLK5 contributes to the inflammatory response in the LEKTI-deficient epidermis by activation of protease-activated receptor 2 (PAR2)[22–24].
In this study, we revealed the in vivo roles of KLK5 and KLK7 using a set of mouse models that are simultaneously deficient for KLK5 and KLK7 on the genetic background of Netherton syndrome-like mouse model based on a mutation found in human patients. The close proximity of these genes (on the same locus) has so far prevented the generation of suitable animal models and therefore the in vivo roles of KLK5 and KLK7 could not be studied concurrently. Our study shows that individual functional ablation of KLK5 or KLK7 is not sufficient to rescue the lethal effect of Spink5 mutation. In contrast, simultaneous deficiency of both KLK5 and KLK7 completely rescues the lethality allowing adult mice to survive to adulthood with a fully functional skin barrier.
To study NS pathology in vivo, we generated a new mouse model mimicking a causative mutation of SPINK5 gene (398delTG; p.A134X) previously described in human patients [25]. Due to the similarity between the human and murine SPINK5 nucleotide sequences, deletion of TG nucleotides at positions 402 and 403 of murine Spink5 (402delTG) produces the premature termination codon (PTC) at a similar position as described in human patients (p.A135X) (Fig 1A). To introduce the mutation into mouse genomic DNA, we prepared TALE nucleases (TALENs) specific for the critical region of Spink5 in combination with a single stranded oligonucleotide (ssODN) carrying the desired mutation (Fig 1B). Founders were screened for targeted incorporation of ssODN by RFLP analysis using XbaI restriction site as a marker (Fig 1C). Heterozygous mice carrying A135X mutation (hereafter referred to as Sp5+/A135X) did not show any obvious phenotype and were used to obtain Sp5A135X/A135X mice, which showed dramatic downregulation of Spink5-RNA expression (Fig 1D). The presence of PTC in the Spink5 transcript was confirmed by sequencing (S1 Fig). Sp5A135X/A135X mice were born in normal Mendelian ratios, however they exhibited severe skin phenotype with exfoliating epidermis, predominantly localized in the abdominal and facial area (Fig 1E) and died within 12 hours after delivery. These phenotypical features mimic NS characteristics and correspond to previously published mouse models of Netherton syndrome [15,26–28].
To elucidate the roles of KLK5 and KLK7 in NS, we generated a set of KLK mutants (Klk5-/-, Klk7-/-, Klk5-/-Klk7-/-), which were crossed with Spink5+/A135X line (Fig 2A). Klk5-/- mice were generated by substitution of exon2 of Klk5 gene with a tm1a-type targeting vector [29] (S2A Fig). As Klk5 and Klk7 are located within close proximity on mouse chromosome 7, generation of Klk5 and Klk7 double-deficient mouse by cross-breeding of individual KO lines is not possible. We therefore applied TALEN-mediated mutagenesis to disrupt the Klk7 gene directly on the genetic background of Klk5-/- mice by introduction of a frame-shift mutation in exon3 of Klk7 gene (S3A Fig). Positively targeted mice were analysed by sequencing of genomic DNA and founder Klk7ex3-A containing 20 bp deletion in Klk7 coding sequence (S3B Fig) was used to establish Klk5-/-Klk7-/- double-deficient line. Targeting of Klk7 was confirmed by cDNA sequencing (S3C Fig) and ablation of KLK7 protein was verified by western blot analysis (S3D Fig). Klk5-/-Klk7-/- mice were further crossed to FLPe expressing mouse line to remove Klk5 KO cassette and re-constitute expression of Klk5 (S2B Fig), thus generating Klk7-/- mutant line (Fig 2A). Klk5-/-, Klk7-/- and Klk5-/-Klk7-/- newborn P0 mice were phenotypically indistinguishable from their control littermates and did not shown any obvious cutaneous phenotype. KLK-deficient mutants were further bred to Sp5+/A135X line in order to generate double Klk5-/-Sp5A135X/A135X, Klk7-/-Sp5A135X/A135X, and triple Klk5-/-Klk7-/-Sp5A135X/A135X mutant mice (Fig 2A). Analysis of mRNA levels confirmed the loss of the targeted gene’s expression (Fig 2B).
Similarly to Sp5A135X/A135X mutants, Klk5-/-Sp5A135X/A135X, Klk7-/-Sp5A135X/A135X, and Klk5-/-Klk7-/-Sp5A135X/A135X mice showed no embryonic lethality and the pups exhibited normal Mendelian ratio. Both Sp5A135X/A135X and Klk7-/-Sp5A135X/A135X pups had a fragile epidermis with numerous epidermal lesions located mainly on the abdomen and head (Fig 3A). Cutaneous defects were strongly improved in Klk5-/-Sp5A135X/A135X mice, which exhibited only minor skin lesions, while Klk5-/-Klk7-/-Sp5A135X/A135X mice showed no skin phenotype and the pups were visually indistinguishable from wt (Fig 3A). Sp5A135X/A135X P0 pups also showed underdeveloped or completely absent vibrissae hairs. These hair defects were partially rescued in Klk5-/-Sp5A135X/A135X and to a greater extent in Klk7-/-Sp5A135X/A135X newborn pups, which exhibited slightly shorter and irregularly distributed vibrissae hairs. Klk5-/-Klk7-/-Sp5A135X/A135X did not show any major abnormalities of whiskers (S4 Fig).
Inactivation of KLK7 did not affect the survival of Sp5A135X/A135X mice as Klk7-/-Sp5A135X/A135X pups died within 12 hours after birth. Interestingly, Klk5-/-Sp5A135X/A135X mice survived until postnatal day 5 (P5) when they exhibited reduced body-size and dry skin with severe scaling throughout the body surface (Fig 3B). In contrast, simultaneous inactivation of both KLK5 and KLK7 fully rescued the lethality and Klk5-/-Klk7-/-Sp5A135X/A135X survive to adulthood. At P5, the skin of Klk5-/-Klk7-/-Sp5A135X/A135X appears to be more stretched and shiny in comparison to wt mice, however they show no signs of scaling (Fig 3B). Nevertheless, Klk5-/-Klk7-/-Sp5A135X/A135X showed alopecia and growth retardation 2–4 weeks after birth (Fig 3C), which disappeared with age. Interestingly, scanning electron microscopy (SEM) analysis of vibrissae and pelage hairs revealed that Klk5-/-Klk7-/-Sp5A135X/A135X pups from P4 –P28 develop a specific hair shaft defect that strongly resembles bamboo hair in NS patients (Fig 3E). The hair defects and growth retardation in Klk5-/-Klk7-/-Sp5A135X/A135X improve with age (Fig 3D and 3F) and no major cutaneous phenotype was seen in adulthood, with the exception of minor scaling on the ears, shorter tail, and a lower body weight (Fig 3D and 3F).
Analysis of newborn P0 mice revealed that Sp5A135X/A135X and Klk7-/-Sp5A135X/A135X have significantly decreased weight in comparison to other mutant and wt lines (S5A Fig) while no differences where observed in E18.5 dpc embryos (S5B Fig). This suggests that the weight loss in Sp5A135X/A135X and Klk7-/-Sp5A135X/A135X lines is caused by severe epidermal barrier disruption followed by rapid dehydration.
The integrity of epidermal barrier in Sp5A135X/A135Xnewborn P0 pups was analysed using the toluidine blue (TB) penetration assay and showed severe skin barrier disruption marked by penetration of TB through large areas of the body, mainly the abdomen, paws, and head. Ablation of KLK7 on the Sp5A135X/A135X background did not improve the barrier and Klk7-/-Sp5A135X/A135X newborns showed similar barrier disruption to Sp5A135X/A135X. In contrast, Klk5-/-Sp5A135X/A135X newborns developed less severe barrier phenotype characterised by multiple small stained patches and in Klk5-/-Klk7-/-Sp5A135X/A135X pups, the barrier integrity was almost completely recovered and the mice showed TB staining only in the area of nostrils (Fig 4A).
To confirm that disruption of skin barrier leads to the dehydration of newborn mice, we assessed the trans-epidermal water loss (TEWL) in P0 pups over time. Consistently with previous analyses of epidermal barrier properties, we found significantly impaired water retention in Sp5A135X/A135X and Klk7-/-Sp5A135X/A135X. Water barrier was partially rescued in Klk5-/-Sp5A135X/A135X pups and completely restored in Klk5-/-Klk7-/-Sp5A135X/A135X (Fig 4B).
To understand why Klk5-/-Sp5A135X/A135X died at P5, mice were stained with TB at P5. Interestingly, Klk5-/-Sp5A135X/A135X mice showed clear penetration of the dye in the epidermis adjacent to hair shafts whereas this barrier defect was completely rescued in Klk5-/-Klk7-/-Sp5A135X/A135X (Fig 4C). Detailed analysis of the epidermis using SEM revealed that P5 Klk5-/-Sp5A135X/A135X mice had dramatic epidermal defects manifested by defective separation of hair shafts from the surrounding tissues and subsequent loss of infundibular epidermis in the upper part of hair follicles. These defects were not observed in Klk5-/-Klk7-/-Sp5A135X/A135X (Fig 4D) thus suggesting that KLK7 activity is responsible for the epidermal barrier defects contributing to the lethality of Klk5-/-Sp5A135X/A135X at P5. No barrier abnormalities were observed in P0 or P5 pups from control lines Klk5-/-, Klk7-/- and Klk5-/-Klk7-/-.
Histological analysis of the epidermis from NS patients together with previously published data on LEKTI-deficient models describe an abnormally differentiated epidermis [15]. In line with these observations, analysis of non-lesional skin of P0 pups showed a reduced granular layer, acanthosis and sporadic SC detachment and parakeratosis in Sp5A135X/A135X pups (Fig 5A and S6 Fig). Although Klk7-/-Sp5A135X/A135X pups showed a similar phenotype to Sp5A135X/A135X mice, no differentiation defects were observed in the epidermis of newborn pups apart from the occasional focal detachment of SC (Fig 5A and S6B Fig). Klk5-/-Sp5A135X/A135X and Klk5-/-Klk7-/-Sp5A135X/A135X newborn pups exhibited well differentiated epidermal layers.
To address the characteristics of epidermal differentiation defects, we analysed the expression of several differentiation markers in E18.5 dpc embryos to avoid the contribution of secondary effects following barrier disruption and exposure to the environment (Fig 5B). In accordance with previous results, Sp5A135X/A135X embryos exhibited a poorly defined basal layer, abnormal expression of keratin 14 (Krt14) and markedly increased expression of keratin 6 (Krt6), suggesting hyperproliferation of keratinocytes in Sp5A135X/A135X embryos (Fig 5B). Krt14 and Krt6 expression in Klk5-/-Sp5A135X/A135X, Klk7-/-Sp5A135X/A135X, and Klk5-/-Klk7-/-Sp5A135X/A135X embryos exhibited a similar pattern to wt animals, indicating that the differentiation defects of LEKTI-deficient epidermis are fully dependent on concurrent activities of both KLK5 and KLK7. In the epidermis of both Sp5A135X/A135X and Klk7-/-Sp5A135X/A135X embryos profilaggrin granules were absent whereas they were present in Klk5-/-Sp5A135X/A135X, Klk5-/-Klk7-/-Sp5A135X/A135X and in wt mice (Fig 5B).
As Klk5-/-Sp5A135X/A135X exhibited drastic epidermal defects before they die at P5 (Fig 4C and 4D), we performed histological analysis of skin from Klk5-/-Sp5A135X/A135X, Klk5-/-Klk7-/-Sp5A135X/A135X and wt mice at P5. We found that although Klk5-/-Sp5A135X/A135X pups showed normal differentiation at P0, over time they developed hyperplastic epidermis with acanthosis, severe intrafollicular hyperkeratosis and the skin was infiltrated by mast cells (Fig 6A and S7 Fig). In contrast, Klk5-/-Klk7-/-Sp5A135X/A135X showed no such defects and the epidermis was comparable to wt (Fig 5C). We also observed an increased expression of keratin6 in the epidermis of Klk5-/-Sp5A135X/A135X P5 pups, which was rescued in Klk5-/-Klk7-/-Sp5A135X/A135X (Fig 6B). Analysis of corneodesmosomal proteins revealed markedly decreased expression of CDSN at the stratum corneum/stratum granulosum interface in Klk5-/-Sp5A135X/A135X, but not in Klk5-/-Klk7-/-Sp5A135X/A135X pups (S8 Fig). These data suggest that abnormal epidermal differentiation of Klk5-/-Sp5A135X/A135X P5 pups is caused by KLK7 activity.
As the aggravated inflammatory response and allergic manifestations are symptomatic for NS, we assayed the expression of pro-inflammatory and pro-TH2 cytokines in the skin isolated from Spink5- and Klk-deficient E18.5 dpc embryos. As expected, Sp5A135X/A135X embryos showed elevated expression levels of TNFα, TSLP, Il-33, Il-1β as well as ICAM1 (Fig 7A). In contrast, these cytokines were not upregulated in Klk5-/-Sp5A135X/A135X animals, which is in line with the recent study of Furio et al. [24] indicating that KLK5 is responsible for triggering the inflammation in LEKTI-deficient epidermis. However, expression levels of TNFα, TSLP, Il-33, Il-1β and ICAM1 were also completely normal in mice with ablated KLK7, i.e. in Klk7-/-Sp5A135X/A135X (Fig 7A). Full rescue of cutaneous inflammation was also observed in Klk5-/-Klk7-/-Sp5A135X/A135X embryos (Fig 7A).
To analyse cutaneous inflammation in later stages of development, we assayed expression levels of TNFα, TSLP, Il-33, Il-1β and ICAM1 in the skin of P5 pups from surviving LEKTI-deficient mutant lines Klk5-/-Sp5A135X/A135X and Klk5-/-Klk7-/-Sp5A135X/A135X and wt controls. Although no signs of cutaneous inflammation were found in Klk5-/-Sp5A135X/A135X E18.5 dpc embryos, P5 pups showed significant upregulation of TNFα, TSLP, Il-33, Il-1β and ICAM1 (Fig 7B). Expression of these cytokines was normalized by inactivation of KLK7, in Klk5-/-Klk7-/-Sp5A135X/A135X P5 pups (Fig 7B).
NS is also associated with systemic inflammation, allergy and elevation of IgE levels. We examined the serum levels of TNFα, Il-1β, IL-9 and IL-17 in Klk5-/-Klk7-/-Sp5A135X/A135X 6 weeks old mice, however no signs of systemic inflammation were found when compared to wt mice (Fig 7C). These observations are in line with the rescue of other NS-like symptoms in adult Klk5-/-Klk7-/-Sp5A135X/A135X mice.
Netherton syndrome is a severe genetic disorder associated with unregulated proteolytic activity, caused by the absence of functional LEKTI, a protease inhibitor encoded by SPINK5 gene. In this study, we elucidated the roles of KLK5 and KLK7 proteases in the disease by genetic inactivation of these proteases on the background of a mouse model for NS. This novel model was generated by mimicking a SPINK5 p.A134X mutation found in human patients [25] and recapitulates the phenotype of previously described Spink5-deficient mouse models [15,26,27]. However, our Sp5A135X/A135X mice in combination with the ablation of KLK5 and KLK7 reveal the complexity of the LEKTI-KLK network. We showed, that although single inactivation of KLK5 or KLK7 rescues a number of NS-like pathological manifestations, only simultaneous ablation of both proteases fully rescues the lethal phenotype of Sp5A135X/A135X mice.
It has been proposed that the barrier defects observed in LEKTI-deficient skin are caused by proteolytic hyperactivity leading to premature degradation of corneodesmosomal proteins. In vitro assays showed that three putative LEKTI targets are able to promote corneodesmosome degradation namely KLK5, KLK7, and KLK14 [18]. The in vitro study of the KLK proteolytic activation cascade proposed that KLK5 acts upstream from KLK7 and KLK14 and therefore, KLK5 hyperactivity should contribute to barrier defects either directly or indirectly via activation of the remaining KLKs. Indeed, significant improvement of skin-barrier defects by inactivation of KLK5 in Sp5A135X/A135X mice was observed, however the rescue was incomplete as toluidine blue staining in Klk5-/-Sp5A135X/A135X mice revealed patches of disrupted-barrier distributed all over the body surface. This observation implicates the role of another protease whose activity contributes to barrier defects in the absence of LEKTI and does not depend on KLK5. This was identified as KLK7, since Klk5-/-Klk7-/-Sp5A135X/A135X newborn mice did not show any major barrier defects of epidermis. Interestingly, single inactivation of KLK7 on Sp5A135X/A135X background did not significantly improve the barrier defects. Therefore we assume that barrier properties of LEKTI-deficient neonatal epidermis are compromised mainly by direct activity of KLK5 and only to a lesser extent by KLK5-mediated activation of KLK7. The significant contribution of KLK5 to NS pathology is in line with a recent study of Furio et al. showing amelioration of skin barrier-phenotype in Spink5-deficent newborns upon KLK5 inactivation [24]. Nevertheless, the remaining activity of KLK7 still contributes to the defective barrier and further intensifies with age as Klk5-/-Sp5A135X/A135X show severe epidermal defects manifested by loss of infundibular epidermis at P5. Klk5-/-Klk7-/-Sp5A135X/A135X mice exhibit no skin-barrier defects at P5 and most importantly, in contrast to Klk5-/-Sp5A135X/A135X mice, the triple mutants survive to adulthood. The skin defects in Klk5-/-Sp5A135X/A135X P5 pups markedly resemble those observed in mice deficient for corneodesmosomal proteins CDSN and DSC1 [30,31]. Indeed, Klk5-/-Sp5A135X/A135X show reduced CDSN expression at P5, which indicates that unregulated activity of KLK7 results in degradation of corneodesmosomes. The mechanism of pro-KLK7 activation in the absence of KLK5 remains unclear. KLK7 can be activated by matriptase [32] and a recent study also suggests a role of mesotrypsin in pro-KLK7 activation [33].
Although the lethal phenotype is fully rescued in Klk5-/-Klk7-/-Sp5A135X/A135X mutants and the mice do not show any signs of skin barrier-defects leading to dehydration, we observed minor barrier disruptions in the nostril area of newborn pups, which suggests the activity of another protease physiologically inhibited by LEKTI. As the toluidine blue -stained area overlaps with the expression of KLK14 in late embryonic development (S9A Fig), we propose that KLK14 could be responsible for the remaining pathology of LEKTI-deficient mice even in the absence of KLK5 and KLK7. Moreover, we and others also observed expression of KLK14 in hair follicles (S9B Fig)[34], which makes KLK14 a candidate protease responsible for the development of the bamboo hair defect in Klk5-/-Klk7-/-Sp5A135X/A135X animals up to the age of 3 weeks. As reported, the defects of cell adhesion proteins in hair follicles result in “lanceolate hair”–a hair shaft phenotype in mice that strongly resembles the bamboo hairs of Klk5-/-Klk7-/-Sp5A135X/A135X mutants and NS patients [35,36]. This further supports a possible role of KLK14 in the formation of bamboo-hair, as KLK14 is linked to the degradation of desmosomal proteins in LEKTI-deficient epidermis [8]. Nevertheless, any targets of LEKTI inhibition present in hair follicles, such as caspase-14 [37] or other, currently unidentified proteases, should be considered as a potential cause of bamboo hairs.
Association of NS with abnormal epidermal differentiation accompanied by acanthosis, parakeratosis, and hyperproliferation of keratinocytes was previously reported in Spink5-deficient mouse models [15,26,27] and our Sp5A135X/A135X confirms the previous findings. We found clear overexpression of keratin6 in Sp5A135X/A135X E18.5 dpc epidermis, suggesting that events leading to hyperproliferation of keratinocytes are triggered prior to the exposure to the external environment and are a result of unregulated proteolytic activity in the epidermis. In light of the fact that single inactivation of either KLK5 or KLK7 completely rescues the differentiation defects in LEKTI-deficient embryos as well as in newborn mice, we believe that the signalling events resulting in keratinocyte hyperproliferation in neonates depend on the presence of both, KLK5 and KLK7 together. Moreover, we observed that aggravated cutaneous inflammation, which is found in E18.5 Sp5A135X/A135X embryos, fully depends on simultaneous activity of both, KLK5 and KLK7. KLK5 was previously shown to initiate inflammation in LEKTI-deficient epidermis by activation of PAR2, which results in the induction of pro-Th2 and pro-inflammatory cytokines [22–24]. In this study we show that KLK7 is also required for the induction of inflammation in LEKTI-deficient mice as P5 Klk5-/-Sp5A135X/A135X pups developed severe acanthosis together with significantly increased expression of TNFα, TSLP, Il-33, Il-1β and ICAM1 while Klk5-/-Klk7-/-Sp5A135X/A135X showed no major defects in the epidermis and increased levels of pro-inflammatory cytokines. Altogether, this suggests that inflammation and differentiation changes in older LEKTI-deficient pups (P5) are initiated by KLK7 activity which is independent of KLK5. Indeed, KLK7 was previously shown to induce inflammation and keratinocyte proliferation in the epidermis [38,39] and a recent study identified KLK7 as a proliferative factor in a mouse model of colon cancer and in human cells in vitro [40]. The mechanism by which KLK7 induces inflammation and differentiation changes remains to be elucidated. In contrast to KLK5, KLK7 cannot directly activate PAR2 as shown in vitro [41] and thus, the inflammation is likely to be triggered by a different mechanism. One possible pathway is the KLK7-mediated conversion of pro-IL1β to active IL1β [42], which could affect the inflammatory phenotype of NS- epidermis.
In summary, we show that the individual inactivation of KLK5 or KLK7 only partially rescues the defective skin barrier but not the lethal phenotype of Sp5A135X/A135X. Only the concurrent ablation of both KLK5 and KLK7 can fully rescue the lethal phenotype of Sp5A135X/A135X mice, therefore both proteases should be investigated as clinical targets. We show that KLK7 plays an important role in the inflammation and defective differentiation in NS and its activity is not dependent on activation by KLK5. We also show that the pathological effects of unregulated KLK activities are remarkably age dependent. Altogether, this study expounds the complexity of the proteolytic network and its regulation, which are especially important to understand Netherton syndrome and its treatment.
All animal studies were ethically reviewed and performed in accordance with European directive 2010/63/EU and were approved by the Czech Central Commission for Animal Welfare.
Knock-out first allele of Klk5 was produced by introduction of targeting construct (vector PRPGS00082_A_A10 obtained from NIH Knock-out Mouse Program, KOMP) via homologous recombination in embryonic stem cells (ESC). Positively targeted ESC were injected into developing wt embryos, to produce chimeric mice, which were used to establish Klk5-/- line.
TALENs targeting exon5 of Spink5 gene were designed using TAL Effector Nucleotide Targeter 2.0 (https://tale-nt.cac.cornell.edu/) [43,44], assembled using the Golden Gate Cloning system[43], and cloned into the ELD-KKR backbone plasmid as described previously [45]. DNA binding domains of TALENs specific for the desired target site within Spink5 gene (Fig 1B) consisted of following repeats: NI-HD-HD-NN-HD-NI-NN-NG-NI-NN-NI-NG-NN-NG-NN-NI-NI-HD-NG (5´ TALEN-Spink5) and HD-NG-NN-HD-NG-NG-NG-NI-NG-NI-NN-NN-NN-NG-NI-HD-NG-HD-NI-HD (3´ TALEN-Spink5). Both TALEN plasmids were used for production of TALEN encoding mRNA as described previously [46]. 5 μl of TALEN mRNA (with total RNA concentration of 40 ng/μl) was mixed with 100 μM of targeting single-stranded oligonucleotide (Sigma-Aldrich; Fig 1B) and the final solution was microinjected into C57BL6/N-derived zygotes. Genomic DNA isolated from tail biopsies of newborn mice were screened by PCR (primers F1: 5´-CCTGTCTCTGCCTTCAGACC-3´ and R1: 5´-GGCTGTGGTAACTGTCCAAAA-3´) and subsequent RFLP analysis using XbaI restriction enzyme (Thermo-Scientific).
TALENs were designed and synthesized as described above. DNA binding domains of TALENs specific for exon3 of murine Klk7 contained following repeats: NN-NG-NI-NI-NI-NN-NI-NI-NN-NN-HD-NG-HD-NN-HD (5´ TALEN-Klk7) and NN-NI-NG-NG-NN-HD-HD-NG-NG-NG-NN-NI-NN-HD-NI-NN (3´ TALEN-Klk7). TALEN mRNA with total RNA concentration of 40 ng/μl was microinjected into Klk5-/-oocytes. Genomic DNA isolated from tail biopsies of newborn mice was screened by PCR (primers F3: 5´- GGAGAAGGCCAGGGTCTGAA-3´ and R3: 5´- TGGTCAGAAACCCACGGAGA-3´) and subsequently analyzed by RFLP using NcoI restriction enzyme (Thermo-Scientific).
Newborn pups from at least two independent litters were separated from mothers to prevent fluid intake. The rate of water loss was analyzed by measuring the reduction of initial body weight at 1h, 2h, 3h and 4 h.
Newborn mice were euthanized and then dehydrated by incubation for 5 min in 25, 50, 75, and 100% methanol. After rehydration in PBS, mice were incubated for 4 hours in 0.1% toluidine blue (Sigma-Aldrich), washed in PBS and imaged.
Newborn pups or skin tissues were fixed in 3.6% formaldehyde for 24 h and embedded in paraffin. 5-μm sections were prepared using microtome were stained by hematoxylin/eosin (H&E) or by 0.5% toluidine blue using standard protocols. Images were obtained using Zeiss Axioscan Z1 (Carl Zeiss AG).
Dorsal skin of P5 pups was fixed in 3.6% formaldehyde for 24 hours and embedded in paraffin. 5-μm paraffin sections were used for antigen retrieval with Discovery Ultra automated IHC/ISH system (Ventana) and stained with antibodies against Dsg-1 (Santa Cruz, 1:100 dilution, retrieval at pH6) and CDSN (Abcam, 1:100 dilution, retrieval at pH6). After 1 hour incubation at room temperature, anti-rabbit peroxidase conjugated polymer (Zytomed GmBH) was applied for 30 min and the reaction was developed using DAB (DAKO) as a chromogen. Images were obtained using Zeiss Axioscan Z1 (Carl Zeiss AG).
In order to stain the cryosections, dorsal skin of E18.5 dpc embryos or P5 pups was isolated, embedded in Tissue-Tek O.C.T (Sakura), and frozen at -80°C. 6 μm-sections were stained as described previously [47], using antibodies against Keratin6 (Covance, 1:1000 dilution), Keratin14 (Covance, 1:2000 dilution) and Filaggrin (Covance, 1:1000 dilution). Nuclei were stained using DAPI (Roche). Images were obtained using Zeiss Axioimager Z2 (Carl Zeiss AG).
Dorsal skin was obtained from newborn pups, crushed in liquid nitrogen and total RNA was isolated using TRIzol (Thermo-Scientific) according to the manufacturer's instructions. Residuals of genomic DNA were removed using 1 U of DNAse I (Roche) per 1 μg of RNA by 15 min incubation at 37°C. 1 μg of total RNA was used for reverse transcription by M-MLV Reverse Transcriptase (Promega) using oligo (dT) primers. RT-PCR was performed in a 20-μl reaction mixture containing SYBR Green JumpStart™ Taq ReadyMix with MgCl2 (Sigma-Aldrich) and 0.25mM of each primer. Respective gene expression was normalized to the expression of TATA-binding protein (TBP). Normalized expression levels were then re-expressed relative to the mean expression level of the respective target in the wt mice. Primer sequences are detailed in S1 Table.
The samples on cellulose filter paper strips were fixed with 3% glutaraldehyde in cacodylate buffer overnight at 4°C. After fixation, extensively washed samples were dehydrated through ascending alcohol concentrations followed by absolute acetone and critical point drying from liquid CO2 in a K 850 unit (Quorum Technologies Ltd). The dried samples were sputter-coated with 20 nm of gold in a Polaron Sputter-Coater (E5100) (Quorum Technologies Ltd). The final samples were examined in a FEI Nova NanoSem 450 scanning electron microscope (FEI) at 5 kV using secondary electron detector.
The levels of TNFα, IL-1β, IL-9, and IL-17 in mouse serum were analyzed using Bio-Plex Pro Mouse Cytokine Assay (Bio-Rad Laboratories) with high sensitivity range standard settings according to manufacturer’s instructions.
|
10.1371/journal.pntd.0002904 | Late Onset of the Serological Response against the 18 kDa Small Heat Shock Protein of Mycobacterium ulcerans in Children | A previous survey for clinical cases of Buruli ulcer (BU) in the Mapé Basin of Cameroon suggested that, compared to older age groups, very young children may be less exposed to Mycobacterium ulcerans. Here we determined serum IgG titres against the 18 kDa small heat shock protein (shsp) of M. ulcerans in 875 individuals living in the BU endemic river basins of the Mapé in Cameroon and the Densu in Ghana. While none of the sera collected from children below the age of four contained significant amounts of 18 kDa shsp specific antibodies, the majority of sera had high IgG titres against the Plasmodium falciparum merozoite surface protein 1 (MSP-1). These data suggest that exposure to M. ulcerans increases at an age which coincides with the children moving further away from their homes and having more intense environmental contact, including exposure to water bodies at the periphery of their villages.
| Although M. ulcerans, the causative agent of Buruli ulcer (BU), was identified in 1948, its transmission pathways and environmental reservoirs remain poorly understood. The occurrence of M. ulcerans infections in endemic countries in West and Central Africa is highly focal and associated with stagnant and slow flowing water bodies. BU is often described as a disease mainly affecting children <15 years of age. However, taking the population age distribution into account, our recent longitudinal survey for BU in the Mapé Dam Region of Cameroon revealed that clinical cases of BU among children <5 years are relatively rare. In accordance with these findings, data of the present sero-epidemiological study indicate that children <4 years old are less exposed to M. ulcerans than older children. Sero-conversion is associated with age, which may be due to age-related changes in behavioural factors, such as a wider movement radius of older children, including more frequent contact with water bodies at the periphery of their villages.
| It has been established that the chronic necrotizing skin disease BU is caused by the emerging pathogen Mycobacterium ulcerans, however the mode(s) of transmission and environmental reservoirs are still unknown. Comparative genetic studies have revealed that M. ulcerans has diverged from the fish pathogen M. marinum. Through the acquisition of a plasmid, M. ulcerans has gained the ability to produce a cytotoxic and immunosuppressive macrolide toxin, referred to as mycolactone [1], [2]. In addition to M. ulcerans strains isolated from human lesions, which belong either to the classical or to the ancestral lineage [3], other mycolactone-producing mycobacteria (MPM) have been identified as fish and frog pathogens and given diverse species names [4]–[7]. However, recent comparative genomic analyses have shown that all MPM are genetically closely related and can be divided into three principal ecovars of M. ulcerans [8]. Extensive pseudogene formation and genome downsizing of the human M. ulcerans pathogen are indicative for an adaptation to a more stable ecological niche. In African endemic settings both the physical environment and organisms such as amoeba, insects, fish and frogs have been proposed as possible environmental reservoirs of the pathogen [9]. Accordingly, direct inoculation of bacteria into the skin from an environmental reservoir, but also bites from insects, such as mosquitos or water bugs have been suggested as route of infection. While possums have been identified as an animal reservoir in BU endemic areas of Southern Australia [10], no mammalian reservoir has so far been detected in Africa. The distribution pattern of lesions is not indicative for a particular route of infection [11] and a genetic fingerprinting study of M. ulcerans isolates has revealed a highly focal transmission pattern, which excludes certain modes of transmission [12].
While it has long been generalized that in African BU endemic areas children below the age of 15 are most affected by the disease [13], population age-stratified data from our previous survey for BU in the Mapé Basin of Cameroon showed that children less than five years old were underrepresented among cases [11]. One explanation for this observation may be a lower degree of exposure of very young children to M. ulcerans. Sero-epidemiological studies in Ghana have shown that screening blood sera of local populations for the presence of IgG specific for the 18 kDa shsp of M. ulcerans represents a tool to monitor exposure of populations to M. ulcerans [14]. However, in these investigations study participants were older than five years of age. Since a proportion of study participants of all age groups tested positive, it is still not known at which age immune responses against M. ulcerans start to emerge and hence where and at which age exposure to the pathogen begins.
In the present sero-epidemiological study the potential association between age and exposure to M. ulcerans was investigated by determining anti-18 kDa shsp IgG titres in 875 individuals from BU endemic sites in the Densu River Basin of Ghana and the Mapé Basin of Cameroon. In these cross-sectional surveys we included more than 100 children less than five years old allowing us to estimate the age of sero-conversion, which may provide another cornerstone in the search for the mode of M. ulcerans transmission.
Ethical clearance for the collection and testing of human blood samples from Ghana and Cameroon was obtained from the institutional review board of the Noguchi Memorial Institute for Medical Research (Federal-wide Assurance number FWA00001824) and the Cameroon National Ethics Committee (N°172/CNE/SE/201) as well as the Ethics Committee of Basel (EKBB, reference no. 53/11). Written informed consent was obtained from all individuals involved in the study. Parents or guardians provided written consent on behalf of children.
We investigated the association between age and exposure to M. ulcerans by determining serum antibody titres against the 18 kDa shsp in individuals living in two different BU endemic areas.
In Cameroon, serum samples were collected from inhabitants of the village of Mbandji 2. This village is located in the Bankim Rural Health Area of the Bankim Health District, where we conducted a cross-sectional house-by-house survey for BU in early 2010, including the collection of data on the population age structure. These data and the subsequent identification of BU cases until June 2012 were published in our previous study [11]. In the present study we provide updated information based on a continued monitoring of new BU cases in this area until May 2013. The age-specific incidence rates were calculated using the ages of the BU cases identified between March 2010 and May 2013 and the population age distribution as collected in the house-by-house survey in the Bankim Health District.
Sera were collected in January 2011 from all inhabitants of Mbandji 2, who agreed to participate (395 individuals with a nearly equal gender distribution). Re-sampling of 80 blood donors from Mbandji 2 was carried out one year after the first blood collection to analyze stability of anti-18 kDa shsp serum IgG levels over time.
The second study site comprised villages within the Obom sub-district of the Ga-South district in Ghana. This sub-district is one of the major BU endemic communities along the Densu River Basin. The villages from which the sera were collected, have active transmission on-going as they have continuously reported cases for the past five years. Study participants included 96 laboratory confirmed BU patients (57 females and 39 males) as well as 4 age-, sex-, and home village-matched controls for each patient (384 control individuals).
Demographic data as well as history of known previous mycobacterial infections were recorded for all participants at both sites. While the majority of individuals had no history of mycobacterial infections, eight study participants from Mbandji 2 reported to having had tuberculosis (2), leprosy (1) or BU (5). All control participants recruited in Ghana had no history of mycobacterial infection. The age distribution of study participants from Cameroon and Ghana is shown in Figure 1A and 1B, respectively.
Blood sera from the 875 individuals were tested for the presence of anti-18 kDa shsp antibodies in an ELISA format. In addition, 96 sera from children <5 living in Mbandji 2 were tested by Western Blot analysis for the presence of antibodies against this protein, as well as against a Plasmodium falciparum MSP-1 protein domain in order to assess the exposure and immune responses of child study participants to this mosquito transmitted parasite.
96-well Nunc-Immuno Maxisorp plates (Thermo Scientific) were coated with 0.25 µg recombinant M. ulcerans 18 kDa shsp per well in 100 µl phosphate-buffered saline (PBS) and incubated over night at 4°C. Plates were washed four times with washing buffer (dH2O, 2.5% Tween 20) before being incubated with blocking buffer 1 (5% skim milk in PBS) for 2 hours at room temperature (RT). After washing as described above, 50 µl of 1∶100 diluted human blood sera in blocking buffer 2 (1% skim milk in PBS) was added to each well and incubated for 2 hours at RT. Following a further washing step, 50 µl of 1∶8000 diluted goat anti-human IgG (γ-chain specific) antibodies coupled to horseradish peroxidase (HRP, SouthernBiotech) in blocking buffer 2 was added to each well and incubated for 1.5 hours at RT. Plates were washed and 50 µl TMB Microwell Peroxidase Substrate (KPL) was added per well. The reaction was stopped after 5 minutes using 0.16 M sulfuric acid. The absorbance was measured at 450 nm in a Tecan Sunrise microplate reader.
15 µg of recombinant M. ulcerans 18 kDa shsp or 5 µg of a Plasmodium falciparum MSP-1 protein domain (amino acids 34-469 of strain K1) were separated on NuPAGE Novex 4–12% Bis-Tris ZOOM Gels with 1.0 mm IPG well (Invitrogen) using NuPAGE MES SDS Running Buffer (Invitrogen) under reducing conditions. After electrophoresis the proteins were transferred onto nitrocellulose membranes using an iBlot Gel Transfer Device (Invitrogen). Membranes were blocked with blocking buffer 3 (5% skim milk in PBS containing 0.1% Tween 20) and cut into thin strips. Membrane strips were then incubated with human blood sera at a 1∶1000 dilution in blocking buffer 3 for 2 hours at RT. Strips were repeatedly washed with 0.3 M PBS containing 1% Tween 20 and after that incubated with 1∶20'000 diluted goat anti-human IgG (γ-chain specific) antibodies coupled to HRP (SouthernBiotech) for 1 hour at RT. After a second washing step, bands were visualized by chemiluminescence using ECL Western Blotting substrate (Pierce).
ELISA results were analyzed using GraphPad Prism version 6.0 (GraphPad Software, San Diego California USA) and R version 3.0.1 [15].
The distribution of antibody titres and the differences between two successive antibody titres are presented as box plots. These comprise a line for the median, edges for the 25th and 75th percentiles and traditional Tukey whiskers showing 1.5 times the interquartile distance. Dots on the graph represent individual points that lie outside that range.
We compared changes in OD between age categories in the Cameroon dataset using the Kruskal-Wallis test. Levene's test for homogeneity of variances was used to compare the degree of variation by age category. We compared the OD values for the Ghana matched cases and controls using conditional logistic regression.
The overall bias and variation between the first and second Ghanaian serum samples was estimated using the Bland-Altman method [16].
The age-specific BU incidence rates for the population in the Mapé Basin were calculated using 76 BU cases identified between March 2010 and May 2013. Based on these cases, a low incidence rate of BU was detected for children less than 4 years of age (Figure 2A).
The age-distribution of IgG titres against the M. ulcerans 18 kDa shsp for a cross-sectional survey of 395 individuals from the village Mbandji 2 is shown in Figure 2B. While high antibody titres were detected in individuals of all age groups over 4 years, none of the children younger than 4 years showed an ELISA IgG titre above the background, which was determined by Western Blot analysis as OD < 0.35. Analysis of the sera sampled from children less than 7 years old by Western Blot analysis showed no specific bands representing IgG antibodies against the 18 kDa shsp for sera from children <4 years of age (Figure 3). In contrast, Western Blot positive sera were found in all tested age groups >4 years old. Since very weak IgG titres were recorded for some of the sera from 4 year olds, sero-conversion may start in some children around this age.
IgG titres against a recombinant fragment of MSP-1 were determined by Western Blot analysis.
In contrast to the lack of antibody responses against the 18 kDa shsp in children <4 years old, serum IgG responses against a P. falciparum malaria parasite MSP-1 domain were detected in all age groups tested. Strong staining of the MSP-1 band was observed for the majority of sera collected from children between one and seven years of age as well as for one of the infants (Figure 4).
One year after the first serum collection in Mbandji 2, 80 of the 395 study participants were re-sampled. While only minimal changes in antibody titres against the 18 kDa shsp were recorded overall, more individuals had a decreased than an increased serum IgG level after one year (Figure 5A).
Increases in OD tended to be small and confined to the older children and young adults (Figure 5B). The most distinct changes, characterized by a marked decrease of antibody titres between the two surveys, occurred in young adults.
There was a significant association between age group and the absolute change in OD (Kruskal-Wallis test p = 0.01) and borderline evidence of an association between the variation in changes in OD and age group (Levene's test for homogeneity of variances, p = 0.08).
M. ulcerans 18 kDa shsp specific IgG titres were also determined in sera from 96 BU patients and 384 healthy matched control individuals living in a second BU endemic site in West Africa, the Densu River Valley in Ghana. Each serum sample was tested twice, once in each of two independent experiments (Figure S1). Negligible overall bias between experiments was observed with the mean difference (OD1-OD2) of 0.024. There was also a reasonably small variation in the individual differences with the 95% limits of agreement from −0.0796 to 0.1278.
There was no evidence of a difference in the ELISA OD values between the cases and controls (p = 0.99) (Figure 6A).
While sero-responders were identified in all age groups of individuals more than 6 years old, none of the sera from children younger than 5 years exhibited a distinct anti-18 kDa shsp IgG titre (Figure 6B). Western Blot analysis of sera from 2-year-old children confirmed the absence of anti-M. ulcerans 18 kDa shsp IgG in these samples (Figure 6C). Results of representative subsets of sera which tested negative, moderately positive or highly positive by ELISA were reconfirmed by Western Blot analysis, showing good agreement between ELISA OD values and Western Blot band intensities (Figure S2).
A high degree of antigenic cross-reactivity among mycobacterial species complicates investigations on M. ulcerans-specific humoral immune responses. However, the immunodominant 18 kDa shsp [17], which is overexpressed in M. ulcerans [18], represents a suitable serological marker for exposure to M. ulcerans [14]. Diverse outcomes of infection with other mycobacteria, such as M. tuberculosis and M. leprae have been associated with both host and pathogen factors. While only one study has investigated a possible association between BU and host genetics [19], various behavioural factors that may lead to increased risk to develop the disease have been reported, with poor wound care, failure to wear protective clothing, and living or working near water bodies being the most common risk factors identified [20]. While the generalization persists that children <15 years old are most affected by the disease [13], our recent survey for BU in the Mapé Basin [11] and continued monitoring of new BU cases in this region have revealed that the risk of BU is as high in individuals above the age of 50 as in young teenagers and that very young children below the age of four are underrepresented among cases when adjusting for the population age distribution. Data of our previous sero-epidemiological investigations revealed that the proportion of individuals from a BU endemic area showing serum IgG titres against the 18 kDa shsp of M. ulcerans is comparable for all age groups >5 years [14]. Results of the present study, including for the first time a substantial number of serum samples from children <5 years of age, showed that children of this age group have not yet sero-converted. Hence, young children appear to be considerably less exposed to M. ulcerans. This reduced exposure may be explained by the smaller movement radius away from the house of these very young children. Although, these small children do leave the house, they usually do so being carried by a caregiver and are therefore not in direct contact with the environment, at more distant places from their homes. No significant difference could be observed when comparing anti-M. ulcerans 18 kDa shsp antibody titres between BU patients and controls. This may be related to the immune-suppressive effect of mycolactone and concurs with the lack of a serological response in experimentally infected mice (unpublished data).
The results of a case-control study carried out in a BU endemic region of south-eastern Australia indicated reduced odds of having BU for individuals who frequently used insect repellent and increased odds for those who were bitten by mosquitoes [21]. In African BU endemic settings, the highly focal transmission of M. ulcerans haplotypes [12], [22], [23], as well as the distribution pattern of BU lesions on the body [11], speak against an exclusive role of mosquito vectors in transmission. Here we observed in children <5 years frequent sero-conversion for the MSP-1 antigen of the mosquito-transmitted malaria parasites in the absence of an IgG response against the M. ulcerans 18 kDa shsp. The age distribution of BU cases and the relatively abrupt changes in this risk of contracting BU with age do not speak for transmission of BU by mosquito species commonly found within the small movement radius of very young children.
Within the framework of our analyses, blood was collected for a second time from a limited number of participants one year after the first sample. Results of this pilot study showed that anti-18 kDa shsp IgG titres were relatively stable in older adults. Future studies of the age-related changes in behaviour of three to six year old children, monitoring of their movement radius and water contact patterns in combination with larger longitudinal serological and environmental studies may have the potential to shed further light onto the mode of transmission and relevant environmental reservoirs of M. ulcerans.
|
10.1371/journal.pcbi.1000674 | Minimization of Biosynthetic Costs in Adaptive Gene Expression Responses of Yeast to Environmental Changes | Yeast successfully adapts to an environmental stress by altering physiology and fine-tuning metabolism. This fine-tuning is achieved through regulation of both gene expression and protein activity, and it is shaped by various physiological requirements. Such requirements impose a sustained evolutionary pressure that ultimately selects a specific gene expression profile, generating a suitable adaptive response to each environmental change. Although some of the requirements are stress specific, it is likely that others are common to various situations. We hypothesize that an evolutionary pressure for minimizing biosynthetic costs might have left signatures in the physicochemical properties of proteins whose gene expression is fine-tuned during adaptive responses. To test this hypothesis we analyze existing yeast transcriptomic data for such responses and investigate how several properties of proteins correlate to changes in gene expression. Our results reveal signatures that are consistent with a selective pressure for economy in protein synthesis during adaptive response of yeast to various types of stress. These signatures differentiate two groups of adaptive responses with respect to how cells manage expenditure in protein biosynthesis. In one group, significant trends towards downregulation of large proteins and upregulation of small ones are observed. In the other group we find no such trends. These results are consistent with resource limitation being important in the evolution of the first group of stress responses.
| Although different environmental stresses trigger specific sets of protective changes in the gene expression of yeast, the adaptive responses to these stresses also share some common features. We hypothesize that minimization of metabolic costs may contribute to shaping such adaptive responses. If this is so, then such pressure should be more noticeable in the costliest biosynthetic processes. One of these is protein synthesis. Thus, we analyze the set of genes and proteins whose expression changes during the responses and look for evidence to support or falsify our hypothesis. We find that protein properties that are indicative of protein cost correlate to changes in gene expression in a way that is consistent with that hypothesis for a large number of adaptive responses. However, if changes in gene expression are small during the adaptive response, we find no evidence of protein cost as a factor in shaping the adaptive response.
| Unicellular organisms are sensitive to environmental challenges. Their internal milieu acts as a buffer against such changes by mounting an adaptive response involving modifications at different cellular levels. Appropriate adaptive responses require intracellular signaling, changes in the conformation and activity of proteins, changes in transcription and translation of genes, etc. [1]. Many of the cellular modifications that characterize any adaptive response are due to the need for acquiring new protein functionalities while shutting down other protein functionalities that are not required in the new conditions. These changes ultimately fine tune the mechanisms and processes that allow the cell to function appropriately and survive under changing environments.
Such fine tuning is shaped by various functional requirements and physiological constraints. The functional requirements are a result of the specific demands that are imposed on cell survival by the environment. On the other hand, the physiological constraints are defined by the limits within which the cell is physically capable of changing the activity of its component parts to meet the functional requirements. From a global point of view, adaptive responses can be seen as a multi-optimization problem because cells evolved appropriate responses to cope with different types of stress, while optimizing different parts of its metabolism for each of those responses [2],[3]. For example, cells simultaneously have to increase the concentration of specific metabolites and proteins, while decreasing the concentration of other components to prevent an increase in the concentration of unneeded metabolites. Such an increase could strain cell solubility capacity or increase spurious reactivity to dangerous levels. These and other functional constraints are likely to provide sustained evolutionary pressures that ultimately select a specific gene expression profile that leads to suitable adaptive responses.
With these arguments in mind, it is thus important to identify the functional requirements and quantitative physiological constraints that may significantly shape adaptive responses. Among others, minimization of energetic expenditure plays an important role in cells growing exponentially in a rich medium. Several signatures that are consistent with minimization of metabolic cost have already been identified in the properties of the set of proteins that is expressed when cells are growing in rich media (basal conditions).
For example, genes coding for proteins that are highly abundant under basal conditions have a pattern of synonymous codon usage that is well adapted to the relative abundance of synonymous tRNAs in the yeast S. cerevisiae and in Escherichia coli [4],[5].
Another signature that is found in genes that are highly expressed under basal conditions is a sequence bias that minimizes transcriptional and translational costs [6]. This minimization of metabolic cost is further observed in the relative amino acid composition of abundant proteins under the same conditions. These proteins are enriched with metabolically cheaper amino acids [7].
A final example of a general signature is the codon bias of long genes. This bias is such that the probability of missense errors is reduced during translation [6],[8],[9]. These biases suggest that reducing overall costs in metabolism, whenever possible, may significantly increase cellular fitness. This view is consistent with the observation that small changes in gene expression affecting the levels of protein synthesis influence the fitness of specific E. coli strains [10].
This body of results strongly supports the notion that metabolic cost acts as a selective pressure in shaping the properties of cells growing in a rich medium, in absence of environmental stresses. Thus, one might ask if minimization of metabolic cost is also an important factor in the evolution of adaptive responses to stress conditions. It is predictable that this evolutionary pressure might leave stronger signatures in adaptive responses that require the use of higher ATP amounts by the cell, such as adaptation to heat, weak organic acids, or NaCl. In these three cases, it has been reported that ATP concentrations decrease due to a high energy demand [11].
Given that protein synthesis is one of the costliest biosynthetic efforts for the cell [12], the minimization of metabolic cost might have biased the properties of proteins whose expression change during adaptation. Therefore, here we ask the following questions. Is there a signature that is consistent with a selective pressure for minimizing metabolic cost in proteins synthesis during adaptive responses to stress? Can one find general signatures in the physicochemical properties proteins and in the expression patterns of genes that are involved in the adaptive response to different environmental challenges? If so, what physiological constraints are consistent with those signatures?
We address these questions by investigating how is the value of several properties of proteins (size and molecular weight of proteins, codon adaptation index, aromaticity, average cost per amino acid, etc.) related to changes in gene expression levels during various environmental changes.
We find that genes whose expression is upregulated during different types of adaptive responses tend to code for proteins that are small, while genes whose expression is downregulated during the same responses tend to code for proteins that are large. This is a signature that is consistent with a selective pressure for minimizing metabolic cost in proteins synthesis. It is more significant in adaptive responses where changes in gene expression levels affect a large fraction of the genome. To our knowledge, this is the first general and global signature that has been identified for the properties of proteins involved in adaptive responses to stress.
The microarray data we analyze provide information regarding relative up and downregulation (UpCF and DownCF, respectively) of gene expression with respect to a pre-stress control condition. To facilitate comparison between upregulated and downregulated genes, we use the inverse of the ratio for downregulated genes. Thus, all values for the ratios of changes in gene expression discussed below are greater than 1.
Changes in gene expression during stress responses are dynamic and, for the most part, transient. Because of this, we take the maximum value of up or downregulation as an approximated measure of the maximal change in gene expression during the transient stress response.
Changes in gene expression are underestimated for genes that undergo very strong up or downregulation, due to intrinsic limitations of the microarray technology [26]. To minimize any errors that may come from this limitation we use the 98th quantile of all the ratio values for a given gene as a proxy of its maximum UpCF or DownCF.
As discussed in the Introduction, previous authors report clear trends between different properties of proteins and their basal abundance in yeast growing exponentially on a rich medium (basal conditions) [6],[7],[8],[9]. In this work we evaluate the existence of similar signatures in the proteins that are involved in the adaptive response of yeast to stress. Before presenting the specific results, and due to the complexity of the analysis, it is worth it to briefly outline the strategy we follow and its rationale. There are four main steps:
We now discuss the results of the analysis in detail.
Some of the protein properties we consider are strongly correlated (see Table S1). For example, different measures of codon preference towards the major tRNA isoacceptors, such as CAI, CB, and FOP, are highly correlated (r = 0.83–0.97). Length and molecular weight of proteins are, in practice, equivalent. Protein and mRNA abundance show a correlation of r = 0.56. Protein abundance is also positively correlated to CAI, to CB and to FOP (r = 0.53–0.54), and to mRNA abundance under basal growth conditions (r = 0.60–0.64). Similarly, average amino acid cost (ACPA) is highly correlated to aromaticity (r = 0.84), because the most expensive amino acids are aromatic. Thus, if a protein has a high percentage of aromatic amino acids (which is proportional to aromaticity) it will have a larger average cost per amino acid than proteins with lower percentage of aromatic amino acids.
The only type of data that is available for both, the entire genome and a comprehensive set of yeast adaptive responses, is gene expression data from microarray experiments [13],[27]. Thus, in order to search for trends between the adaptive responses and the protein properties, we analyze how the value of those properties is related to the changes in gene expression during the response. We analyze microarray data for fourteen stress responses and two control conditions (change in carbon source — C Source — and return to basal conditions after osmotic shock — ↓Sorbitol).
Short proteins with a high relative composition of metabolically cheaper amino acids are highly abundant under basal conditions, which is consistent with the hypothesis that lowering protein cost is a driving force in shaping the protein complement of yeast in those conditions [7]. It is also well known that the process leading up to protein synthesis is one of the costliest components of cellular metabolism [12] and that during response to many environmental stress signals yeast shuts down gene expression and decreases the number of ribosomes [13],[14].
It has been proposed that gene expression profiles have signatures that are specific to the conditions under which they have evolved [2],[28]. If metabolic cost in general, and cost of protein synthesis in particular, is a significant factor in shaping adaptive profiles, then one might expect that the stronger the resource limitation is, the larger its signature will be. It would then be reasonable to expect that adaptive responses where a resource limitation exists may have similar qualitative bulk expenditure in protein synthesis.
To find support for this hypothesis we must estimate that cost for the different stress responses. Changes in protein levels can be roughly estimated over the whole genome by the changes in the levels of gene expression [29]. Thus, an index that approximately estimates the changing costs of protein synthesis during a given adaptive response i can be defined as:(1)
In this equation Ak is the basal abundance of protein k and Lk is the primary sequence length of that protein UpCFik and DownCFik represent the change-fold of up- or downregulation of the gene that codes for protein k.
It is likely that specific functional requirements during any given stress response will lead to the synthesis of new proteins whose functionality is required for survival under the new conditions. By calculating a cost index for each of the twenty five Gene Ontology (GO) categories of cellular components defined in the SGD Slim Mapper Tool, we can analyze if the requirement for new functions is restricted to specific categories of the GO classification or not. Such a discriminating cost index can be defined as:(2)The index refers to stress condition and GO category . For each protein within the GO category , the up- (UpCFijk) or down-change fold (DownCFijk) is multiplied by its length . If in the GO category the expression of genes coding for small proteins is preferably upregulated and the expression of genes coding for large proteins is preferably downregulated, the index will be negative. This index provides a rough bulk estimate of how much a cell invests in synthesizing new proteins (the upregulation term) subtracting how much the cell saves by decreasing the synthesis of other proteins (the downregulation term). The index under basal conditions is 0 because the difference between up- and down-expression is null in that case.
A cluster analysis of the twenty five dimensional vectors built for each adaptive response with the index calculated for each GO category is shown in Figure 3. Four clusters can be distinguished from this analysis. Responses to ↓Sorbitol, C source, Menadione and Acid cluster together with basal conditions (Basal Cluster) and apart from the other responses. Interestingly, this Basal Cluster includes stress responses in which the previous analytical methods find a low correlation between protein cost and changes in gene expression (Tables 1–2 and Figures 1–2 and S1). Because we could not find an accurate bootstrap statistical test to calculate significance for the clusters in Figure 3 we further tested similitude between the conditions using a discriminant analysis of the data used to build the clusters. Two dimensions explain 99.9% of the variance in the data and separate all four groups found in the cluster analysis (Figure S4).
The normalized values of each component of the vector for each type of adaptive response plotted in Figure 4 show the similarity between the different responses. For reference purposes, the basal condition is represented by a dashed circle in each of the panels of that figure. Lines below that circle indicate negative values for while any lines above the basal condition circle indicate positive values for . Conditions included in the Basal Cluster show low absolute values for this index in all GO categories. This is different for the other groups that, overall, have larger negative values for in categories “Cytoplasm”, “Nucleus”, and “Ribosome”. The four clusters are also consistent with the gradation observed in the moving-quantile plots for length and abundance (Figure 2 and S1).
Altogether, the results presented in this section, suggest two broad types of adaptive responses. In one type, corresponding to responses in the Basal Cluster, the changes in gene expression are small. In this group of responses, we find no correlation between protein properties and gene expression. In the other type of stress, responses have evolved in a way that is consistent with a significant pressure to minimize the metabolic cost of the response.
The previous results suggest that the stress conditions considered can be classified in two broad types with respect to metabolic economy. On one hand, we have the Basal Cluster in Figures 3 and 4. This cluster includes the adaptive responses to Menadione, Acid and the two controls, C Source and ↓Sorbitol. The results do not support a significant pressure by metabolic economy in shaping these responses. On the other hand, all other responses can be clustered into three subgroups. Nevertheless, they all appear to be shaped to some degree by metabolic economy. Therefore, we lump together change-folds for gene expression of all these later stress responses. By doing this we create a data set that has a stronger signal than that found in responses to the individual stresses when we relate properties and gene expression changes. The stronger signal in this combined data set also allows us to analyze patterns within each GO category for function, biological processes and cellular location of the proteins.
What type of general selective pressure might explain the correlations we find between changes in gene expression and protein abundance or length during stress response? One answer to this question is that minimizing the cost of protein synthesis is a significant pressure that shapes changes in gene expression during adaptive responses. Why would minimizing metabolic costs improve fitness of S. cerevisiae? As the cell optimizes the expenditure of resources for metabolic maintenance, it will have more resources available for survival and reproduction, thus out competing organisms. This seem logical, but it also raises another question, which is why would one expect this pressure to be felt at the level of proteins?
Calculations based on the typical cellular composition of yeast and bacteria predict that protein synthesis uses more metabolic resources and ATP molecules than the formation of other macromolecules and it is a limiting step for yield [12],[34],[35],[36],[37]. As proteins of a shorter size use less amino acids, evolving fully functional short proteins leads to faster protein synthesis with less usage of cellular resources. It must be stressed that this argument cannot be seen as defending that cell will, over time, simply loose all large proteins and use smaller proteins to perform all necessary molecular functions. Evolution is constrained by life history. Specifically, the evolutionary unit of proteins is the functional domain [38]. Such functional domains have on average appeared only once in evolution and examples of domain convergent evolution are rare. Protein function often depends on how a small number of amino acids are located within the 3D structure of these domains. Therefore, a shorter protein may not have a 3D structure that will allow for the maintenance of an appropriate biological activity. This will constrain the amount of resources that can be saved by evolving shorter proteins.
Under stress, availability of resources may be significantly limited, and the cell must adapt quickly in order to survive. For challenging stress conditions, resource limitation may impose severe limitations to the adaptive response. Exposure to these kinds of stresses causes the cell to deviate considerable resources from its steady state metabolism towards the adaptive response and imposes important constraints to cell economy [11],[39]. For example during exposure to high NaCl concentrations, additional energy expenditure for growth increased between 14% and 31% [40], and the activity of the plasma membrane H+-ATPase (highest consumer of ATP) is repressed during heat shock or in the presence of a weak acid [11],[41]. Another situation that has been put forward as supporting a cellular energetic shortage during stress response is the hypersensitivity to oxidative stress of mutants that lack mitochondrial function and of yeast treated with mitochondrial inhibitors [42]. These authors suggest that the oxidative sensitivity is due to a defect in an energy-requiring process that is needed for detoxification of ROS or for the repair of oxidative molecular damage.
Further support for the importance of protein cost as a selective pressure in the evolution of adaptive changes in gene expression is found in different studies. For example, pathways appear to have evolved to maximize flux for a minimum amount of protein, because the enzyme concentration may be limited by both the protein synthesizing capacity and the solvent capacity of a cell [43]. In fact, theoretical studies suggest that adaptive responses of yeast to environmental changes trigger a gene expression profile that is optimal under the constraint of minimal total enzyme production [2],[44],[45].
There are three aspects that the cell can tune to decrease cost of protein synthesis. First, it can decrease the amount of protein that it synthesizes per time units. If we take changes in gene expression as a proxy of changes in protein synthesis, we find that, in many cases the overall protein synthesis during stress response is decreased (the yij index defined above is negative). Second, the cell may decrease cost of protein synthesis by expressing at higher levels proteins that are small. This would decrease the biosynthetic cost per protein chain and is consistent with our results. Finally, the cell may decrease the cost of protein synthesis by increasing the half life of proteins. We find no evidence for this strategy.
In summary, if decreasing the cost of protein synthesis significantly contributes to shaping the gene expression profile of an adaptive response, we should find trends in the composition of the changing protein complement that are consistent with the following predictions:
The results of our analysis are broadly consistent with these predictions (see Figure 3 for a summary) and support the hypothesis that response to the various stresses has evolved under a selective pressure for minimizing the cost of protein synthesis. GO analysis show that the results are not biased by a specific type of proteins and that the hypotheses are consistent with the results over a wide variety of GO categories. We also see that proteins involved in molecular complexes have changes in gene expression that are similar to proteins that are very large. A more detailed analysis of this later result would require an accurate knowledge about the stoichiometry of the complexes.
Further analysis that would directly establish whether there are limitations on resources and energy usage during a given adaptive response would require data about ATP usage and production under each relevant condition. Such data would allow us to better understand which constraints are important in shaping the evolution of those responses.
|
10.1371/journal.pbio.1001791 | Pre-B Cell Receptor Signaling Induces Immunoglobulin κ Locus Accessibility by Functional Redistribution of Enhancer-Mediated Chromatin Interactions | During B cell development, the precursor B cell receptor (pre-BCR) checkpoint is thought to increase immunoglobulin κ light chain (Igκ) locus accessibility to the V(D)J recombinase. Accordingly, pre-B cells lacking the pre-BCR signaling molecules Btk or Slp65 showed reduced germline Vκ transcription. To investigate whether pre-BCR signaling modulates Vκ accessibility through enhancer-mediated Igκ locus topology, we performed chromosome conformation capture and sequencing analyses. These revealed that already in pro-B cells the κ enhancers robustly interact with the ∼3.2 Mb Vκ region and its flanking sequences. Analyses in wild-type, Btk, and Slp65 single- and double-deficient pre-B cells demonstrated that pre-BCR signaling reduces interactions of both enhancers with Igκ locus flanking sequences and increases interactions of the 3′κ enhancer with Vκ genes. Remarkably, pre-BCR signaling does not significantly affect interactions between the intronic enhancer and Vκ genes, which are already robust in pro-B cells. Both enhancers interact most frequently with highly used Vκ genes, which are often marked by transcription factor E2a. We conclude that the κ enhancers interact with the Vκ region already in pro-B cells and that pre-BCR signaling induces accessibility through a functional redistribution of long-range chromatin interactions within the Vκ region, whereby the two enhancers play distinct roles.
| B lymphocyte development involves the generation of a functional antigen receptor, comprising two heavy chains and two light chains arranged in a characteristic “Y” shape. To do this, the receptor genes must first be assembled by ordered genomic recombination events, starting with the immunoglobulin heavy chain (IgH) gene segments. On successful rearrangement, the resulting IgH μ protein is presented on the cell surface as part of a preliminary version of the B cell receptor—the “pre-BCR.” Pre-BCR signaling then redirects recombination activity to the immunoglobulin κ light chain gene. The activity of two regulatory κ enhancer elements is known to be crucial for opening up the gene, but it remains largely unknown how the hundred or so Variable (V) segments in the κ locus gain access to the recombination system. Here, we studied a panel of pre-B cells from mice lacking specific signaling molecules, reflecting absent, partial, or complete pre-BCR signaling. We identify gene regulatory changes that are dependent on pre-BCR signaling and occur via long-range chromatin interactions between the κ enhancers and the V segments. Surprisingly the light chain gene initially contracts, but the interactions then become more functionally redistributed when pre-BCR signaling occurs. Interestingly, we find that the two enhancers play distinct roles in the process of coordinating chromatin interactions towards the V segments. Our study combines chromatin conformation techniques with data on transcription factor binding to gain unique insights into the functional role of chromatin dynamics.
| B lymphocyte development is characterized by stepwise recombination of immunoglobulin (Ig), variable (V), diversity (D), and joining (J) genes, whereby in pro-B cells the Ig heavy (H) chain locus rearranges before the Igκ or Igλ light (L) chain loci [1],[2]. Productive IgH chain rearrangement is monitored by deposition of the IgH μ chain protein on the cell surface, together with the preexisting surrogate light chain (SLC) proteins λ5 and VpreB, as the pre-B cell receptor (pre-BCR) complex [3]. Pre-BCR expression serves as a checkpoint that monitors for functional IgH chain rearrangement, triggers proliferative expansion, and induces developmental progression of large cycling into small resting Ig μ+ pre-B cells in which the recombination machinery is reactivated for rearrangement of the Igκ or Igλ L chain loci [3],[4].
During the V(D)J recombination process, the spatial organization of large antigen receptor loci is actively remodeled [5]. Overall locus contraction is achieved through long-range chromatin interactions between proximal and distal regions within these loci. This process brings distal V genes in close proximity to (D)J regions, to which Rag (recombination activating gene) protein binding occurs [6] and the nearby regulatory elements that are required for topological organization and recombination [5],[7],[8]. The recombination-associated changes in locus topology thereby provide equal opportunities for individual V genes to be recombined to a (D)J segment. Accessibility and recombination of antigen receptor loci are controlled by many DNA-binding factors that interact with local cis-regulatory elements, such as promoters, enhancers, or silencers [7]–[9]. The long-range chromatin interactions involved in this process are thought to be crucial for the regulation of V(D)J recombination and orchestrate changes in subnuclear relocation, germline transcription, histone acetylation and/or methylation, DNA demethylation, and compaction of antigen receptor loci [5],[10].
The mouse Igκ locus harbors 101 functional Vκ genes and four functional Jκ elements and is spread over >3 Mb of genomic DNA [11]. Mechanisms regulating the site-specific DNA recombination reactions that create a diverse Igκ repertoire are complex and involve local differences in the accessibility of the Vκ and Jκ genes to the recombinase proteins [12]. Developmental-stage-specific changes in gene accessibility are reflected by germline transcription, which precedes or accompanies gene recombination [13]. In the Igκ locus, germline transcription is initiated from promoters located upstream of Jκ (referred to as κ0 transcripts) and from Vκ promoters [14]. Deletion of the intronic enhancer (iEκ), located between Jκ and Cκ, or the downstream 3′κ enhancer (3′Eκ), both containing binding sites for the E2a and Irf4/Irf8 transcription factors (TFs), diminishes Igκ locus germline transcription and recombination [15]–[19]. On the other hand, the Sis (silencer in intervening sequence) element in the Vκ–Jκ region negatively regulates Igκ rearrangement [20]. This Sis element was shown to target Igκ alleles to centromeric heterochromatin and to associate with the Ikaros repressor protein that also colocalizes with centromeric heterochromatin. Sis contains a strong binding site for the zinc-finger transcription regulator CTCC-binding factor (Ctcf) [21],[22]. Interestingly, deletion of the Sis element or conditional deletion of the Ctcf gene in the B cell lineage both resulted in reduced κ0 germline transcription and enhanced proximal Vκ usage [21],[23]. Very recently, a novel Ctcf binding element located directly upstream of the Sis region was shown to be essential for locus contraction and recombination to distal Vκ genes [23]. In addition, the Igκ repertoire is controlled by the polycomb group protein YY1 [24].
Induction of Igκ rearrangements requires the expression of the Rag1 and Rag2 proteins, the attenuation of the cell cycle, and transcriptional activation of the Igκ locus, all of which are thought to be crucially dependent on pre-BCR signaling [4],[25]. At first, pre-BCR signals synergize with interleukin-7 receptor (IL-7R) signals to drive proliferative expansion of IgH μ+ large pre-B cells [4]. In these cells, transcription of the Rag genes is low and the Rag2 protein is unstable due to cell-cycle-dependent degradation [26]. Subsequently, signaling through the pre-BCR downstream adapter Slp65 (SH2-domain-containing leukocyte protein of 65 kDa, also known as Blnk or Bash) switches cell fate from proliferation to differentiation [4]. Importantly, Slp65 (i) induces the TF Aiolos, which down-regulates λ5 expression [27]; (ii) binds Jak3 and thereby interferes with IL-7R signaling [28]; and (iii) reduces inhibitory phosphorylation of Foxo TFs [29]. All these changes result in attenuation of the cell cycle and thus Rag protein stabilization. Moreover, Rag gene transcription is induced by Foxo proteins [30].
Although rearrangement and expression of the Igκ locus can occur independently of IgH μ chain expression [31],[32], several lines of evidence indicate that pre-BCR signaling is actively involved in inducing Igκ and Igλ locus accessibility and gene rearrangement. First, surface IgH μ chain expression correlates with germline transcription in the Igκ locus [33]. Second, in the absence of Slp65, κ0 germline transcription is reduced [34]. Third, mice deficient for Bruton's tyrosine kinase (Btk), which is a pre-BCR downstream signaling molecule interacting with Slp65, show reduced Igλ L chain germline transcription and reduced Igλ usage [35]. Fourth, transgenic expression of the constitutively active E41K-Btk mutant in IgH μ chain negative pro-B cells induces premature rearrangement and protein expression of Igκ L chain [34]. Based on fluorescence in situ hybridization (FISH) studies, it has been proposed that in pro-B cells distal Vκ and Cκ genes are separated by large distances and that the Igκ locus specifically undergoes contraction in small pre-B and immature B cells actively undergoing Vκ-Jκ recombination [36]. However, it remains unknown how pre-BCR-induced signals affect the accessibility, contraction, and topology of the Vκ region, or how they affect the long-range interactions of the κ regulatory elements involved in organizing these events.
In this study, we identified the effects of pre-BCR signaling on germline Vκ transcription and on the expression of TFs implicated in the regulation of Igκ gene rearrangement. We found that the decrease in pre-BCR signaling capacity in wild-type, Btk-deficient, Slp65-deficient, and Btk/Slp65 double-deficient pre-B cells was paralleled by a gradient of decreased expression of many TFs including Ikaros, Aiolos, Irf4, and (to a lesser extent) E2a, as well as by a decreased Igκ locus accessibility for recombination. Several of these factors can mediate long-range chromatin interactions and are known to occupy κ regulatory elements that regulate locus accessibility [37]–[40]. We therefore sought to analyze the effect of pre-BCR signaling on the higher order chromatin structure organized by these regulatory sequences at the Igκ locus. To this end, we performed chromosome conformation capture and sequencing (3C-seq) analyses [41] on pro-B cells and pre-B cells from mice single or double deficient for Btk or Slp65 to evaluate the effects of this pre-BCR signaling gradient on Igκ locus topology. These 3C-seq experiments demonstrated that already in pro-B cells the κ enhancers robustly interact with the ∼3.2 Mb Vκ region and its flanking sequences, and that pre-BCR signaling induces accessibility by a functional redistribution of enhancer-mediated chromatin interactions within the Vκ region.
Whereas mice deficient for the pre-BCR signaling molecules Btk and Slp65 have a partial block at the pre-B cell stage [42],[43], in Btk/Slp65 double-deficient mice, only very few pre-B cells show progression to the immature B cell stage characterized by functional IgL chain gene recombination [44]. To enable analysis of the effects of pre-BCR signaling on (i) the expression of genes involved in Igκ gene rearrangement and on (ii) long-distance chromatin interactions in the Igκ locus in pre-B cells in the absence of Igκ gene recombination events, we bred Btk and Slp65 single- and double-deficient mice on the Rag1−/− background. In these mice, progression of B cell progenitors to the pre-B cell stage was conferred by the transgenic, functionally rearranged VH81x IgH μ chain, which ensures pre-BCR expression and cellular proliferation. The absence of functional Rag1 protein precludes IgL chain gene rearrangement and cells are completely arrested at the small pre-B cell stage (Figure 1A).
We performed genome-wide expression profiling of FACS-purified B220+CD19+ pre-B cell fractions from wild-type (WT), Btk, and Slp65 single- and double-deficient VH81x transgenic Rag1−/− mice (Figure 1A). In these experiments non-VH81x transgenic Rag1−/− pro-B cells served as controls. One-way ANOVA analysis using MeV software (p<0.01) [45] revealed that 266 genes were differentially expressed between the five groups of pro-B/pre-B cells (Figure 1B). When compared with WT VH81x transgenic Rag1−/− pre-B cells, 174 genes were up-regulated, whereby the average values of the fold increase were ∼1.70, ∼3.28, ∼3.36, and ∼3.47 for Btk−/−, Slp65−/−, Btk−/−Slp65−/− VH81x transgenic Rag1−/− pre-B cells and non-VH81x transgenic Rag1−/− pro-B cells, respectively (see Table S1). A similar gradient of gene expression changes was apparent from the average values of the fold change for the 192 significantly down-regulated genes, which were ∼1.65, ∼2.29, ∼3.79, and ∼4.15 in the four groups of pre-B/pro-B cells, respectively (see Table S2). In a hierarchical clustering analysis of the five groups of B cell precursors, the expression profiles of Btk−/−Slp65−/− VH81x transgenic Rag1−/− pre-B cells and non-VH81x transgenic Rag1−/− pro-B cells were very similar (Figure 1B). This implies that expression of the 266 genes is not substantially influenced by pre-BCR-mediated proliferation, which is still induced in pre-B cells lacking both Btk and Slp65 [44],[46] but not in Rag1−/− pro-B cells. Consistent with these findings, gene distance matrix analysis revealed a clear gene expression gradient among the five groups of pre-B/pro-B cells, in which Btk−/−Slp65−/− pre-B and Rag1−/− pro-B cells again showed highly comparably expression signatures (Figure S1).
In agreement with previous findings [34],[43],[46], pre-BCR signaling-defective pre-B cells manifested increased expression of Dntt, encoding terminal deoxynucleotidyl transferase and the SLC components Vpre (Vpreb1) and λ5 (Igll1), as well as decreased expression of the cell surface markers Cd2, Cd22, Cd25(IL-2R), and MHC class II (Table 1). Btk and Slp65 single-deficient and particularly double-deficient pre-B cells failed to up-regulate various genes known to be involved in IgL chain recombination, such as Ikzf3 (Aiolos), Ikzf1 (Ikaros), Irf4, Spib, Pou2f2 (Oct2), polymerase-μ [47], as well as Hivep1 encoding the Mbp-1 protein, which has been shown to bind to the κ enhancers [48]. In addition, pre-BCR signaling influenced the expression levels of many other DNA-binding or modifying factors that were not previously associated with IgL chain recombination, including Lmo4, Zfp710, Arid1a/3a/3b, the lysine-specific demethylases Aof1 and Phf2, Prdm2 (a H3K9 methyltransferase), the sik1 gene encoding a histon deacetylase (HDAC) kinase, Hdac5, Hdac8, and the DNA repair protein gene Rev1 (Table S2). We did not find significant differences in the expression of several other TFs implicated in Ig gene recombination—for example, Obf1/Oca-B, Pax5, E2a, and Irf8 (Table 1). In addition, in signaling-deficient pre-B cells, we found reduced transcription of genes encoding several signaling molecules (e.g., Rasgrp1, Rapgefl1, Ralgps2, Blk, Traf5, Hck, Nfkbia (IκBα), Syk, Csk), cell surface markers (Cd38, Cd72, Cd74, Cd55, and Notch2), or genes regulating cell survival (Bmf and Bcl2l1 encoding BclXL) (Table S2). Interestingly, we observed concomitant up-regulation of signaling molecules that are also associated with the T cell receptor (Lat, Zap70, and Prkcq (PKCθ); Table S1).
Next, we used quantitative RT-PCR to confirm the observed differential expression of several TFs. Expression levels of these genes were indeed significantly reduced in a pre-BCR signaling-dependent manner, especially for Aiolos, Ikaros, and Irf4, with residual expression levels in Btk−/−Slp65−/− VH81x transgenic Rag1−/− pre-B cells that were ∼1%, ∼20%, and ∼9% of those observed in WT VH81x Rag1−/− mice, respectively (Figure 1C). In addition, we found moderate effects on Obf1 (Oca-B) and E2a with residual expression levels of ∼28% and ∼44%, respectively. In chromatin immunoprecipitation (ChIP) assays, we observed in pre-B cells substantial binding of E2a protein to the intronic and 3′ κ enhancer regions and to the three Vκ regions analyzed. Under conditions of reduced pre-BCR signaling activity, E2a binding to the enhancers was essentially maintained (3′Eκ) or reduced (iEκ), but E2a binding to the Vκ regions was lost (Table S3). Consistent with the significant reduction of Ikaros expression in Slp65−/− pre-B cells, Ikaros binding to both κ enhancers and Vκ regions was undetectable in these cells (Table S3).
Taken together, from these findings we conclude that the five groups of pro-B/pre-B cells, representing a gradient of progressively diminished pre-BCR signaling, show in parallel a gradient of diminished modulation of many genes that signify pre-B cell differentiation, including key genes implicated in Igκ gene recombination.
In these expression profiling studies, we only detected limited differences in germline transcription (GLT) over unrearranged Jκ and Vκ gene segments, which is thought to reflect locus accessibility [12]. However, we previously showed by serial-dilution RT-PCR that the levels of κ0 0.8 and κ0 1.1 germline transcripts, which are initiated in different regions 5′of Jκ and spliced to the Cκ region [49], are apparently normal in Btk−/− pre-B cells, modestly reduced in Slp65−/− pre-B cells, and severely reduced in Btk−/−Slp65−/− pre-B cells [34]. We could confirm these findings for κ0 GLT by quantitative RT-PCR assays on FACS-purified B220+CD19+ pro-B/pre-B cell fractions (Figure 2A). In agreement with our reported findings [34], we also found that Btk−/−and Slp65−/− pre-B cells have defective λ0 transcription, which is initiated 5′ of the Jλ segments (Figure 2B) [49].
GLT across the Vκ region showed a similar pattern of sensitivity to pre-BCR signaling: decreased transcription of six individual Vκ regions tested (Vκ3–7, Vκ8–24, Vκ4–55, Vκ10–96, Vκ1–35, and Vκ2–137) correlated with decreased pre-BCR signaling activity (Figure 2C) in the pre-B cells of the four groups of mice. GLT over unrearranged Vλ1 and Vλ2 segments was strongly reduced in the absence of Btk or Slp65, as detected by the expression arrays (Table 1).
These observations indicate that Igκ locus accessibility, a hallmark of recombination-competent antigen receptor loci, is progressively reduced under conditions of diminishing pre-BCR signaling.
Accessibility of antigen receptor loci for V(D)J recombination is thought to be initiated by enhancers, in part through long-range chromatin interactions with promoters of noncoding transcription, resulting in the activation of germline transcription [8]. Because pre-BCR signaling affects the expression of GLT and various nuclear proteins that mediate long-range chromatin interactions and bind the κ enhancers, it is conceivable that pre-BCR signaling induces changes in the enhancer-mediated higher order chromatin structure of the Igκ locus that facilitates Vκ gene accessibility.
We therefore performed 3C-Seq analyses on FACS-purified B220+CD19+ fractions from the same five groups of mice (WT, Btk−/−, Slp65−/−, and Btk−/−Slp65−/− VH81x transgenic Rag1−/− pre-B cells, as well as Rag1−/− pro-B cells). Erythroid progenitors were analyzed in parallel as a nonlymphoid control, in which the Igκ locus was not contracted. Genome-wide chromatin interactions were measured for three regulatory elements involved in the control of Igκ locus accessibility and recombination: the iEκ and 3′Eκ enhancers [50]–[52] and the Sis element [20], which contain binding sites for Ikaros/Aiolos, E2a, and Irf4 [16],[17],[20],[38],[53].
In WT pre-B cells, all three regulatory elements showed extensive long-range chromatin interactions within the Vκ region and substantially less interactions with regions up- or downstream of the ∼3.2 Mb Igκ domain (Figure 3A; see Figure S2, Figure S3, and Figure S4 for line graphs), confirming previous observations [21]. Under conditions of reduced pre-BCR signaling activity, the three Igκ regulatory elements still showed strong interactions with the Vκ region. Surprisingly, even in the complete absence of pre-BCR signaling in Rag1−/− pro-B cells, long-range interactions were still observed at frequencies well above those seen in nonlymphoid cells, suggesting that a contracted Igκ locus topology is not strictly dependent on pre-BCR signaling (Figure 3A, Figure S2, Figure S3, and Figure S4). Next, we used 3D DNA FISH analyses using BAC probes hybridizing to the distal Vκ and Cκ/enhancer regions to confirm that Igκ locus contraction was similar in Rag-1−/− pro-B cells and VH81x transgenic Rag-1−/− pre-B cells (both showing a contracted topology, compared with noncontracted pre–pro-B cells deficient for the TF E2a; Figure 3B).
Nevertheless, we did observe that pre-BCR signaling induced clear differences in interaction frequencies. Whereas an increase in pre-BCR signaling was associated with a decrease in the interaction frequencies between the two κ enhancers and regions flanking the Igκ locus (as also revealed by more detailed images of selected regions upstream and downstream of the Igκ domain; see Figure S5), the overall interaction frequency within the Igκ domain appeared unchanged (Figure S3, Figure S4, and Figure S5). Remarkably, interactions with the Sis element showed quite an opposite pattern: pre-BCR signaling correlated with increased overall interactions within the Igκ domain and did not substantially affect interaction frequencies in the Igκ flanking regions (Figure S2 and Figure S5).
Taken together, these analyses show that (i) the Igκ locus is already contracted at the pro-B cell stage and that (ii) pre-BCR signaling induces changes in long-range chromatin interactions, both within the Igκ locus and in the flanking regions.
The differential effects of pre-BCR signaling on long-range chromatin interactions of the iEκ, 3′Eκ, and Sis elements clearly emerged in a quantitative analysis of the 3C-seq datasets (Figure 4A; see Materials and Methods for a detailed description of the quantification methods used). When pre-BCR signaling was absent (Rag1−/− pro-B cells) or very low (Btk−/−Slp65−/− pre-B cells), the average interaction frequencies were similar within the ∼3.2 Mb Vκ region and the ∼3.2 Mb downstream flanking region, for all three regulatory elements. Interaction frequencies with the upstream flanking region were lower, consistent with the larger chromosomal distance to the three viewpoints. The presence of increasing levels of Btk/Slp65-mediated pre-BCR signaling was associated with reduced interaction of iEκ and 3′Eκ with the Igκ flanking regions and with increased interaction of the Sis element and (to a lesser extent) 3′Eκ with the Vκ region (Figure 4A). As a result, for all three regulatory elements pre-BCR signaling resulted in a preference for interaction with fragments inside the Vκ region over fragments outside the Vκ region (Figure S7).
We next focused our analysis on the Vκ region and compared fragments that harbor a functional Vκ gene (Vκ+ fragment) and those that do not (Vκ− fragment). When pre-BCR signaling was absent (Rag1−/− pro-B cells) or very low (Btk−/−Slp65−/− pre-B cells), the average interaction frequencies of the Sis or iEκ elements with Vκ+ fragments were higher than with Vκ− fragments. The average interaction frequencies of 3′Eκ with Vκ+ and Vκ− fragments, however, were similar (Figure 4B). Upon pre-BCR signaling, the Sis element showed an increase in interaction frequencies with both Vκ+ and Vκ− fragments, with nevertheless an interaction preference for Vκ+ fragments. In contrast, interaction frequencies between the iEκ element and Vκ+ or Vκ− fragments were not modulated by pre-BCR signaling at all (Figure 4B). The 3′Eκ element exhibited yet another profile: pre-BCR signaling induced increased interaction frequencies specifically with Vκ+ fragments, while interactions with Vκ− fragments were not notably modulated by pre-BCR signaling (Figure 4B). When we separately analyzed nonfunctional pseudo-Vκ genes, we found for the Sis and 3′Eκ elements that the interaction patterns with functional and nonfunctional Vκ genes were similar (Figure S8). In contrast, the iEκ enhancer did show an overall increased interaction frequency with Vκ functional genes, compared with nonfunctional Vκ genes, a phenomenon which was again independent from pre-BCR signaling (Figure S8).
The finding that interactions of Vκ genes with the intronic enhancer are already robust in pro-B cells, while those with the 3′κ enhancer are dependent on pre-BCR signaling, suggested that for individual Vκ genes pre-BCR signaling may result in more similar interaction frequencies with the two enhancers. To investigate this, we examined for all individual Vκ genes the correlation between their 3C-seq interaction frequencies with the iEκ and 3′κ elements and found that these were highly correlated in WT pre-B cells (R2 = 0.68; Figure 4C). Correlation was severely reduced when pre-BCR signaling was low in Btk−/−Slp65−/− pre-B cells (R2 = 0.26; Figure 4C). Similar pre-BCR signaling-dependent correlations were observed between Vκ-interactions with the Sis element and those with the two enhancers (Figure S9). As the Sis element particularly suppresses recombination of the proximal Vκ3 family, we investigated interaction correlations specifically for this Vκ family. Similar to our findings for all Vκ genes, a subanalysis showed strong correlations for the interactions of Vκ3 family genes with iEκ, 3′κ, and Sis in WT pre-B cells, which were diminished when pre-BCR signaling was low, except for iEκ–Sis correlations, which were pre-BCR signaling-independent (Figure S9).
In summary, we conclude that pre-BCR signaling induces a redistribution of long-range interactions of the iEκ, 3′Eκ, and Sis elements, thereby restricting interactions towards the Vκ gene region. Moreover, upon pre-BCR signaling the long-range interactions mediated by 3′Eκ and Sis—but not those mediated by iEκ—become enriched for fragments harboring a Vκ gene, demonstrating increased proximity of 3′Eκ and Sis to Vκ genes. Finally, for individual Vκ genes, the interactions with iEκ, 3′Eκ, and Sis become highly correlated upon pre-BCR signaling, indicating that pre-BCR signals result in regulatory coordination between these three elements that govern Igκ locus recombination. In contrast, interactions between genes of the proximal Vκ3 family, Sis and iEκ—but not 3′κ—appear to be coordinated already in the absence of pre-BCR signaling.
Next, we investigated the effects of pre-BCR signaling on the interaction frequencies of individual functional Vκ genes with the three κ regulatory elements (Figure 5A,B). The 3C-seq patterns of the majority (∼91%) of the 101 individual Vκ+ fragments showed evidence for interaction with one or more of the κ regulatory elements (>25 average counts). When comparing Btk−/−Slp65−/− with WT pre-B cells, we observed that for a large proportion (∼38–52%) of Vκ+ fragments, interaction frequencies increased upon pre-BCR signaling (Figure 5B). Smaller proportions of Vκ+ fragments showed a decrease (∼12–29%) or were not significantly affected by pre-BCR signaling (∼17–25% with <1.5-fold change). The observed increase or decrease was not related to proximal or distal location of the Vκ genes, nor to their sense or antisense orientation (not shown). Distributions of the three different classes of Vκ+ fragments showed substantial differences between the κ regulatory elements. For the Sis and 3′Eκ elements, more Vκ+ fragments showed increased than decreased interactions (Figure 5B), in agreement with the signaling-dependent increase in average interaction frequencies of all Vκ+ fragments (Figure 4B). In contrast, for the iEκ viewpoint, Vκ+ fragments showing increased and decreased interactions were more equal in number, consistent with the limited effects of pre-BCR signaling on overall iEκ interaction frequencies of all Vκ+ fragments (Figure 4B).
Although antigen receptor recombination is in principle regarded as a random process, a significant skewing of the primary Igκ repertoire of C57BL/6 mice was recently reported: one third of the Vκ genes was shown to account for >85% of the Vκ segments used by B cells [54]. To assess whether a correlation exists between usage of Vκ genes and their interaction frequencies with κ regulatory elements, we divided the Vκ genes into four usage categories (<0.1%, 0.1–0.3%, 0.3–0.5%, and >0.5%) and calculated their average 3C-Seq interaction frequencies with Sis, iEκ, and 3′κ (Figure 5C). In WT pre-B cells, Vκ usage showed a strong positive correlation with 3C-Seq interaction frequencies for all three regulatory elements (R2 = ∼0.7–0.9; Figure 5C). These correlations were pre-BCR signaling-dependent, since in Btk−/−Slp65−/− pre-B cells, they were reduced (for iEκ; R2 = 0.33) or absent (for Sis and 3′κ; R2<0.10 and R2<0.16, respectively) (Figure 5C).
Collectively, our results indicate that specifically the most frequently used Vκ genes are the main interaction targets of κ regulatory elements, whereby pre-BCR signaling completely underlies this specificity for the Sis and 3′Eκ elements, and to a lesser extent for iEκ.
Next, we investigated whether long-range interactions between κ regulatory elements and the Vκ region correlated with the presence of the TFs Ctcf [21], Ikaros [55], and E2a [56], which have been implicated in Igκ locus recombination [21],[37],[55],[57],[58]. Notably, Ikaros and E2a both strongly bind all three κ regulatory elements, while the Sis element is also occupied by Ctcf ([21]; unpublished data).
Remarkably, we found similar striking correlations between the presence of in vivo binding sites for each of these TFs (as determined by ChIP experiments; see Materials and Methods for the relevant references) and long-range chromatin interactions with the κ regulatory elements (Figure 6A–C), even though Ctcf sites are mostly located in between Vκ genes [21] and Ikaros/E2a sites were frequently found close to Vκ gene promoter regions ([2]; Figure 7A). Even when pre-BCR signaling was absent (Rag1−/− pro B cells) or very low (Btk−/−Slp65−/− pre-B cells), the average interaction frequencies of the κ regulatory elements with fragments containing Ctcf, Ikaros, or E2a bindings sites were higher than those without binding sites. Irrespective of the presence or absence of bindings sites for these TFs, we found that upon pre-BCR signaling interaction frequencies with the Sis element increased and those with the iEκ did not change. In contrast, for the 3′Eκ we found that pre-BCR signaling specifically increased interaction frequencies with fragments occupied by Ctcf, Ikaros, or E2a.
Finally, we found that the presence of di- or trimethylation of histone 3 lysine 4 (H3K4Me2/3), an epigenetic signature associated with locus accessibility [59] and Rag-binding [60],[61], also correlated with increased interaction frequencies with κ regulatory elements, revealing a similar pre-BCR signaling dependency as seen for the TFs analyzed (Figure 6D).
We conclude that the presence of essential TFs or H3K4Me2/3 in the Vκ region strongly correlates with the formation of long-range chromatin interactions with the κ regulatory elements, and that for the Sis and 3′Eκ elements this interaction preference is further enhanced by pre-BCR signaling.
Since the long-range interactions with κ regulatory elements correlated with the presence of TFs implicated in Igκ recombination, we next asked whether the κ regulatory elements preferentially interacted with Vκ genes that are in close proximity to binding sites for Ctcf, Ikaros, or E2a.
Strikingly, the majority of functional Vκ genes (95/101) was found to have an Ikaros binding site in close proximity—that is, located on the same 3C-seq restriction fragment (average length of ∼3 kb, unpublished data) (Figure 7A). Proximity of Vκ genes to an E2a binding site (37%) or H3K4Me2/3 positive region (∼28%) is more selective, while only a small fraction of Vκ genes are close to Ctcf binding sites (∼12%) ([22]; Figure 7A). All Vκ genes marked by E2a, Ctcf, H3K4Me2/3, or a combination of these also contain an Ikaros binding site. Frequently used Vκ genes (>1.0% usage; 33/101 genes) were located in two separate regions, a proximal and a distal region, which also contained virtually all E2a and H2K4Me2/3-marked Vκ genes (Figure 7A).
We found that Vκ genes marked by both Ikaros and E2a were used substantially more often than those only bound by Ikaros (Figure 7B), suggesting that these Vκ genes are preferentially targeted for Vκ-to-Jκ gene rearrangement. Our 3C-seq analyses showed that in WT pre-B cells, interaction frequencies with the three κ regulatory elements were higher for Ikaros/E2a-marked Vκ genes compared to genes marked by Ikaros binding alone (Figure 7C). In fact, Vκ+ restriction fragments containing an Ikaros binding site but not an E2a binding site showed interaction frequencies similar to Vκ− restriction fragments. Under conditions of very low pre-BCR signaling (in Btk−/−Slp65−/− pre-B cells), we observed strongly reduced interaction frequencies of Vκ+ E2a binding restriction fragments with the Sis and 3′Eκ elements. These interaction frequencies were in the same range as those of Vκ− fragments or Vκ+ fragments that harbored an Ikaros site only (Figure 7C). Interaction frequencies with the iEκ enhancer, however, were independent of pre-BCR signaling. As shown in Figure 7D, for the majority of Ikaros/E2a-marked Vκ+ fragments (65%), pre-BCR signaling was associated with increased interactions with the Sis and 3′Eκ elements (comparing wild-type and Btk−/−Slp65−/− pre-B cells). In these analyses, only ∼13.5% and ∼5.4% of Ikaros/E2a-marked Vκ+ fragments showed a decreased interaction frequency upon pre-BCR signaling. In contrast, almost equal proportions of Ikaros/E2a-marked Vκ+ fragments showed increased (∼37%) and decreased (∼30%) interactions with iEκ upon pre-BCR signaling.
Taken together, these data reveal strong positive correlations between the presence of E2a binding sites, Vκ usage, and long-range chromatin interactions with κ regulatory elements in pre-B cells. Remarkably, for the iEκ element, these correlations are largely independent of Btk/Slp65-mediated pre-BCR signaling, whereas for the 3′Eκ they are completely dependent on signaling.
During B-cell development the pre-BCR checkpoint is known to regulate the expression of many genes, part of which control the increase in Igκ locus accessibility to the V(D)J recombinase complex. However, it remained unknown how pre-BCR signaling events affect accessibility in terms of Igκ locus contraction and topology.
Here we identified numerous genes involved in IgL chain recombination, chromatin modification, signaling, and cell survival to be aberrantly expressed in pre-B cells lacking the pre-BCR signaling molecules Btk and/or Slp65. We found that GLT over the Vκ region, reflecting Vκ accessibility, is strongly reduced in these cells. We used 3C-Seq to show that in pro-B cells both the intronic and the 3′ κ enhancers frequently interact with the ∼3.2 Mb Vκ region, as well as with Igκ flanking sequences, indicating that the Igκ locus is already contracted at the pro-B cell stage. 3C-Seq analyses in wild-type and Btk/Slp65 single- and double-deficient pre-B cells demonstrated that pre-BCR signaling significantly affects Igκ locus topology. First, pre-BCR signaling reduces the interactions of the intronic and 3′κ enhancers with Igκ flanking regions, effectively focusing enhancer action towards the Vκ region to facilitate Vκ-to-Jκ recombination. Second, pre-BCR signaling strongly increases nuclear proximity of the 3′κ enhancer to Vκ genes, whereby this increase is more substantial for more frequently used Vκ genes and for Vκ genes close to a binding site for the basic helix-loop-helix protein E2a. Third, pre-BCR signaling augments interactions between κ regulatory elements and fragments within the Vκ region bound by the key B-cell TFs Ikaros and E2a and the architectural protein Ctcf. Fourth, pre-BCR signaling has limited effects on interactions of the intronic κ enhancer with fragments within the Igκ locus, as this enhancer already displays interaction specificity for functional Vκ genes and TF-bound regions in pro-B cells. Fifth, pre-BCR signaling has limited effects on the interactions between the intronic or 3′κ enhancers and fragments that do not contain a Vκ gene or an Ikaros, E2a, or Ctcf binding site, emphasizing the specificity of pre-BCR signaling-induced changes in Igκ locus topology. Sixth, pre-BCR signaling appears to induce mutual regulatory coordination between the three regulatory elements, as their interaction profiles with individual Vκ genes become highly correlated upon signaling. Finally, pre-BCR signaling increases interactions of the Sis element with DNA fragments in the Igκ locus, irrespective of the presence of a Vκ gene or TF. Collectively, our findings demonstrate that pre-BCR signals relayed through Btk and Slp65 are required to create a chromatin environment that facilitates proper Igκ locus recombination. This multistep process is initiated by up-regulation of key TFs like Aiolos, Ikaros, Irf4, and E2a. These proteins are then recruited to or further accumulate at the Igκ locus and its regulatory elements, resulting in a specific fine-tuning of enhancer-mediated locus topology that increases locus accessibility to the Rag recombinase proteins.
Importantly, the presence of strong lineage-specific interaction signals between the Cκ/enhancer region and distal Vκ genes in pro-B cells indicates that the Igκ locus is already contracted at this stage. In contrast to a previous microscopy study indicating that Igκ locus contraction did not occur until the small pre-B cell stage [36], our 3D DNA FISH analysis indeed detected similar nuclear distances between distal Vκ and the Cκ/enhancer region in cultured pro-B and pre-B cells. Recently Hi-C was employed to study global early B cell genomic organization whereby substantial interaction frequencies were found between the intronic κ enhancer and the Vκ region in pro-B cells [40]. E2a-deficient pre–pro-B cells, which are not yet fully committed to the B-cell lineage [62], showed very few interactions among the iEκ and the distal part of the Vκ region [40], resembling the interactions we observed in nonlymphoid cells (Figure 3A). Accordingly, 3D-FISH analysis showed that the Igκ locus adopted a noncontracted topology in these pre–pro-B cells (Figure 3B). These data indicate that Igκ locus contraction is already achieved in pro-B cells and depends on the presence of E2a. Supporting this notion, active histone modifications and E2a were already detected at the κ enhancers and Vκ genes at the pro-B cell stage [56],[63], whereby E2a was frequently found at the base of long-range chromatin interactions together with Ctcf and Pu.1, possibly acting as “anchors” to organize genome topology [40]. The observed correlation between E2a binding, Vκ gene usage and iEκ proximity in pro-B cells (Figure 5C, Figure 7C) further strengthens an early critical role for E2a in regulating Igκ locus topology, Vκ gene accessibility, and recombination.
Our 3C-seq experiments revealed that pre-BCR signaling is not required to induce long-range interactions between the κ regulatory elements and distal parts of the Vκ locus, indicating that TFs strongly induced by signaling—that is, Aiolos, Ikaros, and Irf4—are not strictly necessary to form a contracted Igκ locus. Prime candidates for achieving Igκ locus contraction at the pro-B cell stage are E2a and Ctcf, as they have been implicated in regulating Ig locus topology [21],[40],[64],[65] and E2a already marks frequently used Vκ genes at the pro-B cell stage (Figure 7), although we did observe reduced E2a expression and binding to the iEκ enhancer and Vκ genes when pre-B cell signaling was low (Figure 1 and Table S3), suggesting that pre-BCR signaling is required for high-level E2a occupancy of the Vκ genes. We previously reported that Igκ gene recombination can occur in the absence of Ctcf and that Ctcf mainly functions to limit interactions of the κ enhancers with proximal Vκ regions and to prevent inappropriate interactions between these strong enhancers and elements outside the Igκ locus [21]. Because at the pro-to-pre–B cell transition Aiolos, Ikaros, and Irf4 are recruited to the Igκ locus and histone acetylation and H3K4 methylation increases [17],[38],[63],[66], we hypothesize that pre-BCR–induced TFs act upon an E2a/Ctcf-mediated topological scaffold to further refine the long-range chromatin interactions of the κ regulatory elements. Hereby, these TFs mainly act to focus and to coordinate the interactions of the two κ enhancers to the Vκ gene segments, in particular to frequently used Vκ genes, thereby increasing their accessibility for recombination (see Figure 7E for a model of pre-BCR signaling-induced changes in Igκ locus accessibility).
In this context, our 3C-seq data show that the two κ enhancer elements have distinct roles. Both 3′Eκ and iEκ elements manifest interaction specificity for highly used, E2a-marked, Vκ genes. However, whereas iEκ already shows this specificity in pro-B cells (although pre-BCR signaling does augment this specificity), 3′Eκ only does so in pre-B cells upon pre-BCR signaling. These observations indicate that iEκ is already “prefocused” at the pro-B cell stage and that pre-BCR signals are required to fully activate and focus the 3′Eκ to allow synergistic promotion of Igκ recombination by both enhancers (see Figure 7E) [52]. In agreement with such distinct sequential roles, iEκ and not the 3′Eκ was found to be required for the initial increase in Igκ locus accessibility, which occurred upon binding of E2a only [37],[38],[67]. The 3′Eκ on the other hand requires binding of pre-BCR signaling-induced Irf4 to promote locus accessibility [19],[38], followed by further recruitment of E2a to both κ enhancers and highly used Vκ genes (Table S3 and [38],[57]).
The Sis regulatory element was shown to dampen proximal Vκ–Jκ rearrangements and to specify the targeting of Igκ transgenes to centromeric heterochromatin in pre-B cells [20]. As Sis is extensively occupied by the architectural Ctcf protein and deletion of Sis or Ctcf both resulted in increased proximal Vκ usage [21],[23], it was postulated that Sis functions as a barrier element to prevent the κ enhancers from too frequently targeting proximal Vκ genes for recombination. In this context, we now provide evidence that interactions between the proximal Vκ genes, Sis, and iEκ—but not 3′κ—are already coordinated before pre-BCR signaling occurs (Figure S9). Perhaps not surprisingly, Sis-mediated long-range chromatin interactions displayed a pattern and pre-BCR signaling response that was different from the κ enhancers. Unlike for the enhancers, upon pre-BCR signaling, Sis-mediated interactions with regions outside the Igκ locus were maintained and interaction within the Vκ region increased, irrespective of the presence of Vκ genes or TF binding sites. Because Sis is involved in targeting the nonrecombining Igκ allele to heterochromatin [20], the observed interaction pattern of the Sis element might reflect its action in pre-B cells to sequester the nonrecombining Igκ locus and target it towards heterochromatin. This might also explain the increased interaction frequencies of Sis with highly used Vκ genes upon pre-BCR signaling (Figures 5C and 7C), as such highly accessible genes likely require an even tighter association with Sis and heterochromatin to prevent undue recombination.
Surprisingly, we observed a striking correlation between Ikaros binding and Vκ gene location (94% of Vκ genes were in close proximity to an Ikaros binding site; Figure 7A). Although Ikaros and Aiolos have a positive role in regulating gene expression during B-cell development [55],[58] and Ikaros is required for IgH and IgL recombination [39],[58], Ikaros has also been reported to silence gene expression through its association with pericentromeric heterochromatin [68] or through recruitment of repressive cofactor complexes [69],[70]. Recruitment of Ikaros to the Igκ locus was found increased in pre-B cells as compared to pro-B cells [63], in agreement with its up-regulation in pre-B cells (Figure 1). Furthermore, Ikaros binds the Sis element, where it was suggested to mediate heterochromatin targeting of Igκ alleles by the Sis region [20]. Aiolos, although not essential for B-cell development like Ikaros [58],[71], is strongly induced by pre-B cell signaling and has been reported to cooperate with Ikaros in regulation gene expression [27]. Although their synergistic role during IgL chain recombination has not been extensively studied, the Ikaros/Aiolos ratio changes upon pre-BCR signaling (Figure 1). Increased recruitment of Ikaros/Aiolos to Vκ genes and the κ enhancers likely increases Igκ locus accessibility and contraction (see Figure 6), as Ikaros was very recently shown to be essential for IgL recombination [58]. On the other hand, it is conceivable that on the nonrecombining allele, increased recruitment of Ikaros/Aiolos to Vκ genes and the Sis region could facilitate silencing of this allele. Further investigations using allele-specific approaches [72] will be required to clarify the allele-specific action of the Sis element during Igκ recombination.
In summary, by investigating the effects of a pre-BCR signaling gradient—rather than deleting individual TFs—we have taken a more integrative approach to study the regulation of Igκ locus topology. Our 3C-Seq analyses in wild-type, Btk, and Slp65 single- and double-deficient pre-B cells show that interaction frequencies between Sis, iEκ, or 3′ Eκ and the Vκ region are already high in pro-B cells and that pre-BCR signaling induces accessibility through a functional redistribution of long-range chromatin interactions within the Vκ region, whereby the iEκ and 3′Eκ enhancer elements play distinct roles.
VH81x transgenic mice [73] on the Rag-1−/− background [74] that were either wild-type, Btk−/− [75], Slp65−/− [42], or Btk−/−Slp65−/− have been previously described [34]. Mice were crossed on the C57BL/6 background for >8 generations, bred, and maintained in the Erasmus MC animal care facility under specific pathogen-free conditions and were used at 6–13 wk of age. Experimental procedures were reviewed and approved by the Erasmus University Committee of Animal Experiments.
Preparation of single-cell suspensions and incubations with monoclonal antibodies (mAbs) were performed using standard procedures. Bone marrow B-lineage cells were purified using fluorescein isothiocyanate (FITC)-conjugated anti-B220(RA3-6B2) and peridinin chlorophyll protein (PCP)-conjugated anti-CD19, together with biotinylated mAbs specific for lineage markers Gr-1, Ter119, and CD11b and APC-conjugated streptavidin as a second step to further exclude non-B cells. Cells were sorted with a FACSARia (BD Biosciences). The following mAbs were used for flow cytometry: FITC-, PerCP–anti-B220 (RA3-6B2), phycoerythrin (PE)–anti-CD2 (LFA-2), PCP-, allophycocyanin (APC)- or APC–Cy7–anti-CD19 (ID3), PE-, or APC anti-CD43 (S7). All these antibodies were purchased from BD Biosciences or eBiosciences. Samples were acquired on an LSRII flow cytometer (BD Biosciences) and analyzed with FlowJo (Tree Star) and FACSDiva (BD Biosciences) software.
Extraction of total RNA, reverse-transcription procedures, design of primers, and cDNA amplification have been described previously [21]. Gene expression was analyzed using an ABI Prism 7300 Sequence Detector and ABI Prism Sequence Detection Software version 1.4 (Applied Biosystems). All PCR primers used for quantitative RT-PCR of TFs or κ0, λ0, and Vκ GLT are described in [21], except for Obf1 (forward 5′-CCTGGCCACCTACAGCAC-3′, reverse 5′-GTGGAAGCAGAAA CCTCCAT-3′, obtained from the Roche Universal Probe Library).
Biotin-labeled cRNA was hybridized to the Mouse Gene 1.0 ST Array according to the manufacturer's instructions (Affymetrix); data were analyzed with BRB-ArrayTools (version 3.7.0, National Cancer Institute) using Affymetrix CEL files obtained from GCOS (Affymetrix). The RMA approach was used for normalization. The TIGR MultiExperiment Viewer software package (MeV version 4.8.1) was used to perform data analysis and visualize results [45]. One-way ANOVA analysis of the five experimental groups of B cells was used to identify genes significantly different from wild-type VH81X Tg Rag1−/− pre-B cells (p<0.01).
ChIP experiments were performed as previously described [76] using FACS sorted bone marrow pre-B cell fractions (0.3–2.0 million cells per ChIP). Antibodies against E2a (sc-349, Santa Cruz Biotechnology) and Ikaros (sc-9861, Santa Cruz Biotechnology) were used for immunoprecipitation. Purified DNA was analyzed by quantitative RT-PCR as described above. Primer sequences are available on request.
3C-Seq experiments were essentially carried out as described previously [21],[41]. For 3C-Seq library preparation, BglII was used as the primary restriction enzyme and NlaIII as a secondary restriction enzyme. 3C-seq template was prepared from WT E13.5 fetal liver erythroid progenitors and FACS-sorted bone marrow pro-B cell or pre-B cell fractions (see above) from pools of 4–6 mice. In total, between 1 and 8 million cells were used for 3C-seq analysis. Primers for the Sis, iEκ, and 3′Eκ viewpoint-specific inverse PCR were described previously [21]. 3C-seq libraries were sequenced on an Illumina Hi-Seq 2000 platform. 3C-Seq data processing was performed as described elsewhere [41],[77]. Two replicate experiments were sequenced for each genotype and viewpoint, and normalized interaction frequencies per BglII restriction fragment were averaged between the two experiments.
For quantitative analysis, the Igκ locus and surrounding sequences were divided into three parts (mm9 genome build): a ∼2 Mb upstream region (chr6:65,441,978–67,443,029; 759 fragments), a ∼3.2 Mb Vκ region (chr6:67,443,034–70,801,754; 1,290 fragments) and a downstream ∼3.2 Mb region (chr6:70,801,759–73,993,074; 1,143 fragments). For each cell type (as described above) sequence read counts within individual BglII restriction fragments were normalized for differences in library size (expressed as “reads per million”; see [74]) and averaged between the two replicates before further use in the various calculations. Very small BglII fragments (<100 bp) were excluded from the analysis. Fragments in the immediate vicinity of the regulatory elements (chr6:70,659,392–70,693,183; 10 fragments) were also excluded because of high levels of noise around the viewpoint, a characteristic of all 3C-based experiments. Vκ gene coordinates (both functional genes and pseudogenes) were obtained from IMGT [11] and NCBI (Gene ID: 243469) databases. Vκ gene usage data (C57BL/6 strain, bone marrow) were obtained from [54]. ChIP-seq datasets were obtained from [21] (Ctcf), [55] (Ikaros), and [56] (E2a, H3K4Me2, and H3K4Me3). Vκ genes were scored positive for TF binding sites or for a histone modification, if they were located on the same BglII restriction fragment (corresponding to the 3C-Seq analysis).
Rag-1−/− pro-B and Rag-1−/−;VH81X pre-B cells were isolated from femoral bone marrow suspensions by positive enrichment of CD19+ cells using magnetic separation (Miltenyi Biotec). Cells were cultured for 2 wk in Iscove's Modified Dulbecco's medium containing 10% fetal calf serum, 200 U/ml penicillin, 200 mg/ml streptomycin, 4 nM L-glutamine, and 50 µM β-mercaptoethanol, supplemented with IL-7 and stem cell factor at 2 ng/ml. E2a−/− hematopoietic progenitors were grown as described previously [78]. Prior to 3D-FISH analysis, cells were characterized by flow cytometric analysis of CD43, CD19, and CD2 surface marker expression to verify their phenotype (Figure S6).
3D DNA FISH was performed as described previously [79] with BAC clones RP23-234A12 and RP23-435I4 (located at the distal end of the Vκ region and at the Cκ/enhancer region, respectively; Figure 3A) obtained from BACPAC Resources (Oakland, CA). Probes were directly labeled with Chromatide Alexa Fluor 488-5 dUTP and Chromatide Alexa Fluor 568-5 dUTP (Invitrogen) using Nick Translation Mix (Roche Diagnostics GmbH).
Cultured primary cells were fixed in 4% paraformaldehyde, and permeabilized in a PBS/0.1% Triton X-100/0.1% saponin solution and subjected to liquid nitrogen immersion following incubation in PBS with 20% glycerol. The nuclear membranes were permeabilized in PBS/0.5% Triton X-100/0.5% saponin prior to hybridization with the DNA probe cocktail. Coverslips were sealed and incubated for 48 h at 37°C, washed, and mounted on slides with 10 µl of Prolong gold anti-fade reagent (Invitrogen).
Pictures were captured with a Leica SP5 confocal microscope (Leica Microsystems). Using a 63× lens (NA 1.4), we acquired images of ∼70 serial optical sections spaced by 0.15 µm. The datasets were deconvolved and analyzed with Huygens Professional software (Scientific Volume Imaging, Hilversum, the Netherlands). The 3D coordinates of the center of mass of each probe were transferred to Microsoft Excel, and the distances separating each probe were calculated using the equation: √(Xa−Xb)2+(Ya−Yb)2+(Za−Zb)2, where X, Y, and Z are the coordinates of object a or b.
Statistical significance was analyzed using a nonparametric Mann–Whitney U test (IBM SPSS Statistics 20). The p values<0.05 were considered significant.
3C-seq and microarray expression datasets have been submitted to the Sequence Read Archive (SRA, accession number SRP032509) and Gene Expression Omnibus (GEO, accession number GSE53896), respectively.
|
10.1371/journal.pntd.0003124 | Use of Humanised Rat Basophilic Leukaemia Cell Line RS-ATL8 for the Assessment of Allergenicity of Schistosoma mansoni Proteins | Parasite-specific IgE is thought to correlate with protection against Schistosoma mansoni infection or re-infection. Only a few molecular targets of the IgE response in S. mansoni infection have been characterised. A better insight into the basic mechanisms of anti-parasite immunity could be gained from a genome-wide characterisation of such S. mansoni allergens. This would have repercussions on our understanding of allergy and the development of safe and efficacious vaccinations against helminthic parasites.
A complete medium- to high-throughput amenable workflow, including important quality controls, is described, which enables the rapid translation of S. mansoni proteins using wheat germ lysate and subsequent assessment of potential allergenicity with a humanised Rat Basophilic Leukemia (RBL) reporter cell line. Cell-free translation is completed within 90 minutes, generating sufficient amounts of parasitic protein for rapid screening of allergenicity without any need for purification. Antigenic integrity is demonstrated using Western Blotting. After overnight incubation with infected individuals' serum, the RS-ATL8 reporter cell line is challenged with the complete wheat germ translation mixture and Luciferase activity measured, reporting cellular activation by the suspected allergen. The suitability of this system for characterization of novel S. mansoni allergens is demonstrated using well characterised plant and parasitic allergens such as Par j 2, SmTAL-1 and the IgE binding factor IPSE/alpha-1, expressed in wheat germ lysates and/or E. coli. SmTAL-1, but not SmTAL2 (used as a negative control), was able to activate the basophil reporter cell line.
This method offers an accessible way for assessment of potential allergenicity of anti-helminthic vaccine candidates and is suitable for medium- to high-throughput studies using infected individual sera. It is also suitable for the study of the basis of allergenicity of helminthic proteins.
| Infection with parasitic helminths is characterised by a marked elevation of total and parasite-specific Immunoglobulin E (IgE). It is widely believed that this IgE response has evolved to protect hosts against large metazoan parasites. Such a protective function has been well characterised in particular against members of the genus Schistosoma. However, with a few notable exceptions, the molecular targets of the IgE response and the downstream immunological mechanisms leading to host protection are not well understood. The molecular targets of a specific IgE response are by definition called allergens. While almost 3,000 different allergens, contained in e.g. plant pollen or seeds, moulds or animal materials, have been characterised at the molecular level, and are listed and described in databases such as the Allergome database (www.allergome.org), only a few dozen allergens have been characterised in parasitic helminths. A more detailed understanding of the molecular targets of the anti-helminth IgE response can not only be expected to further our basic understanding of protective immune responses and allergy in general–such knowledge can also be expected to have important repercussions on the production of safe and effective anti-helminthic vaccines. This research describes a novel approach suitable for genome-wide functional identification of allergens in S. mansoni and other parasites, paving the way for the identification of the Schistosoma allergome.
| Helminthic parasites are well known to induce a strong Th2-biased response in their hosts, characterised by elevated levels of total and parasite-specific IgE, IL-4, IL-5 and IL-13, with concomitant expansion and mobilization of specific effector cells [1], [2]. This IgE response is widely believed to have evolved to protect against ectoparasites and parasitic helminths, and Schistosoma in particular [3], although this widespread view has been recently challenged [4].
Human infection with the trematode Schistosoma mansoni is well known to correlate with a progressive increase of serum IgE levels [5]. S. mansoni infection usually peaks in early adolescence and declines in adulthood, a pattern that suggests that individuals in endemic areas can gradually acquire an age-related resistance to reinfection [6], [7]. Progressive acquisition of anti-schistosome immunity coincides with natural death of worms (averaging 10–15 years of life), an event during which the parasites release and expose previously inaccessible antigens to the immune system [8]. Similarly, repeated treatment with praziquantel can speed up the process of immunity, resulting (in some individuals) in so-called post-treatment resistance to infection [7], [9]–[11]. A Th1-type or mixed Th-1/Th2-type response is associated with putative natural resistance in ‘endemic normal’ individuals [12]. However, post-treatment resistance is associated with a stronger Th2-type response dominated by IgE and IgG4 [5], with the higher IgE/IgG4 ratio, rather than their absolute levels, best predicting resistance [13]–[15]. A group of antigens related to the different infection status of endemic area residents in Brazil was recently identified by a serological proteomic analysis which may be related to susceptibility or resistance to infection [14]. However, despite recent progress and decades of research, the targets of this protective antibody response and the basis of its ‘inefficient’ acquisition are still unknown.
The occurrence of natural and post-treatment resistance suggests that immunity could be conferred by appropriately formulated vaccines, possibly using mixtures of antigens. Strategies used for vaccine development have changed as the genomic data for schistosomes have become increasingly available and post-genomic technologies have matured [15]. The traditional approach has been to identify immunogenic antigens using immunological screening (i.e. Western Blots), followed by cloning, expression and case-by-case testing for protection in murine or other animal models. To date, even the best vaccine candidates have achieved protection levels below 70% in animal models, with higher protection only achieved by using high doses of irradiated cercariae [16], [17]. The most promising vaccine candidate (SmTSP-2) achieved 57% and 64% reduction (adult worm and egg burden, respectively) and importantly was recognized by IgG1 and IgG3, but not IgE, in sera of naturally resistant, but not uninfected or chronically infected individuals [18].
There is however a major unsolved conundrum specific to the development of anti-helminthic vaccines. While the bulk of the evidence points to a major protective role of the parasite-specific IgE response against the parasite, vaccinating with an allergen bears the inherent risk of potentially inducing hazardous allergic reactions in sensitised individuals, as recently reported during clinical trials for an anti-hookworm vaccine using Na-ASP-2, where adult volunteers experienced generalized urticarial reactions immediately after vaccination [19]. It could be shown that individuals who displayed urticarial reactions possessed high levels of IgE against Na-ASP-2. This led to testing of specific IgE levels for other candidate vaccine antigens such as Necator americanus GST1 and APR1 using sera from individuals resident in helminth-endemic areas [20]. Thus it would be beneficial to identify such allergens at an early stage during vaccine development.
We have previously shown that human basophils are sensitised within 6 weeks of a single, low-dose infection with N. americanus infective stage larvae [21]. Basophil activation could be detected by flow cytometry in the absence of measurable parasite-specific IgE levels in the serum. This suggests that basophils may offer a sensitive way of measuring the presence of parasite antigen-specific IgE in infected individuals, and, perhaps more importantly in the context of vaccination, to demonstrate the ability of allergens to induce basophil or mast cell activation, in contrast to measuring allergen binding by specific IgE alone. Therefore, we recently developed a new detection system for antigen-specific IgE based on the NFAT-dependent luciferase expression in a humanised rat basophilic leukaemia cell line (RS-ATL8) [22], [23]. When sensitised with egg white-allergic patient's serum, this cell line detected at least 1 fg/mL of egg white extract proteins as a luciferase expression [23].
The sensitivity of this detection method makes it possible to study the potential allergenicity of a protein using only minute amounts of protein. The lack of requirement for a high yield allows the use of a fast and easy cell-free expression system such as wheat germ lysate, which allows for expression of microgram amounts of protein [24] in less than two hours. This in turn makes it possible to produce many correctly folded antigens of interest in short time [25].
Here, we demonstrate proof-of-principle of how the RS-ATL8 cell line, in combination with a cell-free in vitro translation system and a set of stringent quality controls, can be used for assessment of allergenicity of S. mansoni antigens. This technology paves the way for high-throughput, genome-wide assessment of S. mansoni antigen allergenicity - the Schistosoma allergome.
Samples of schistosomiasis patients were from a Ugandan study. All EDTA plasma samples were obtained from a male cohort from the village of Musoli on Lake Victoria. Samples used in this study are from a subgroup of infected people described in Fitzsimmons et al. [26]. The blood samples were selected based on their known content of SmTAL1-specific IgE as measured by isotype specific ELISA, as described by Naus and co-authors [27]. Details of the plasma samples are summarised in Table S1 in the Supplementary data.
Ethical clearance was obtained from the Uganda National Council of Science and Technology (ethics committee for Vector Control Division, Ugandan Ministry of Health) who approved the age of consent as 15 y at the time of sample collection (2004/2005). Consent forms were translated into the local language and informed written consent was obtained from all adults and from the parents/legal guardians of all children. Parental consent was not sought for individuals 15–18 y old.
Sera from patients allergic to Parietaria judaica (commonly known as spreading pellitory in the Mediterranean area, and sticky weed or asthma weed in Australia) were collected after informed consent from the patients, and under a study protocol approved by the institutional ethical committee to establish the sera bank. Institutional Review Board of IDI-IRCCS, Rome, Italy (n. 106-CE-2005). The sera obtained from nine patients were pooled in equal amounts and the levels of specific IgE, IgG4, and total IgG against 104 103 different allergens measured by ImmunoCAP ISAC multiplexing analysis (ThermoScientific) following the protocol previously described [28]. The pooled sera showed high levels of specific IgE to Par j 2 (55 U/ml). The complete ISAC characterisation of the pooled sera is shown in the Supplementary data (Table S2).
Fifteen chosen genes representing proteins from a diverse families such as major egg allergens, troponin, fatty acid binding protein (complete list described in supplementary data in Table S3) were amplified from relevant cDNA libraries (adult worm λZAP cDNA library generous donation by K. Hancock, CDC Atlanta, USA; egg stage λZAP cDNA library kindly contributed by Helmut Haas and Gabi Schramm, Research Centre Borstel, Germany), or available cDNA clones (kindly donated by Alan Wilson, University of York, UK) using 25 µL JumpStart REDTaq ReadyMix Reaction Mix (Sigma-Aldrich), 2.5 µL of each forward and reverse custom made gene specific primer (Sigma Aldrich, final concentration 0.5 µM), 2 µL of S. mansoni cDNA library in 50 µL final volume. For longer sequences, Q-5 polymerase (New England Biolabs) was used to take advantage of its high proofreading activity. Forward primers were constructed by adding a SgfI restriction site and a start ATG (where not available, i.e. when expressing the mature protein sequence after leader peptide cleavage at the 5′ end), and reverse primers by adding a His6-tag followed by a valine (for facilitation of subcloning into other expression vectors) and a stop codon, followed by a PmeI restriction site at the 3′ end. The complete primer sequences are listed in Table S3 in Supplementary materials.
Temperature gradients were run initially to optimise the annealing temperature for the different genes; however an annealing temperature of 54°C worked for most of the genes. Successful PCR was confirmed by 1% agarose gel electrophoresis, using 100 bp Tridye DNA ladder (New England Biolabs) for reference. The final products were purified using a Promega Wizard SV Gel and PCR Clean-Up System, as described by the manufacturer.
The concentration of the amplified genes was measured using a NanoDrop 1000 Spectrophotometer (Thermo Scientific). The purified PCR products were inserted into pF3A WG (BYDV) Flexi Vector (Promega) using the Flexi Vector System (Promega). The ligations were heated at 65°C for 5 min for T4 DNA ligase (HC) inactivation, before transformation, which substantially increased the number of colonies obtained. Transformation was achieved by employing electro-competent DH5α Escherichia coli cells using a BioRad Micropulser electroporation apparatus, following standard molecular biology procedures. Plasmids were purified from the transformed cells using a QIAGEN QIAprep Spin Miniprep Kit, as per the manufacturer's protocol. All purified plasmids were verified by DNA sequencing (Source Bioscience, Nottingham, UK).
Plasmids were used to produce proteins through coupled in vitro transcription-translation, using TnT SP6 High-Yield Wheat Germ Protein Expression System (Promega), as per the manufacturer's instructions. Protein synthesis was initiated by mixing the appropriate DNA template (2–3 µg), 30 µL of the TnT SP6 High-Yield Wheat Germ Master Mix and water for a 50 µL final volume, and then incubating the reaction at 25°C for 2 hours. Protein expression was analyzed by the incorporation of labelled lysine residues (FluoroTect GreenLys, Promega) in a 10 µl aliquot of the plasmid/wheat germ lysate mixture (WGL) as directed in the instructions. Samples were heated for 3 minutes at 70°C, run on 4–20% SDS-PAGE gradient gels (BioRad, UK), under reducing conditions and imaged with a laser-based fluorescent gel scanner (Fujifilm LAS-4000 319 Imaging System). Molecular weights (MW were estimated using the Kaleidoscope Precision plus marker (BioRad, UK) which contains several fluorescently labelled components.
SmTAL1 and Sm-TAL2 were expressed in E. coli as described previously [31].
IPSE/alpha-1, SmTAL1, and SmTAL2 proteins produced in the wheat germ lysate system were transferred to a 0.45 µm nitrocellulose membrane (NCM) (Sigma-Aldrich) and sections of membrane were incubated with mouse anti-IPSE/alpha-1 monoclonal antibody (1∶2000) or rabbit anti-sera (1∶500) against SmTAL1 or SmTAL2, using the method described by Burnette [32]. The anti-IPSE/alpha-1 monoclonal antibody used is from clone 74 2G4 [33], [34]which recognises both monomeric and dimeric IPSE/alpha-1. Both antibodies were diluted in TBS with 3% Tween, 30% Wheat Germ Extract and 5% skimmed milk powder.
The mouse or rabbit primary antibodies were detected with an HRP-conjugated secondary antibody using chemiluminescence (ECL Plus Western Blotting Detection System, GE Healthcare) diluted 1∶5000 and visualised using a Fujifilm LAS-4000 imaging system.
RS-ATL8 cells were cultured in 75 cm2 flasks, with 0.2 µm vent caps (Corning, USA), in an incubator set at 37°C with 5% carbon dioxide with a humidified atmosphere [22], [23]. The flasks contained 10 mL MEM (GIBCO, USA), supplemented with 5% v/v heat-inactivated FCS (GIBCO, USA), 100 U/mL penicillin, 100 µg/mL streptomycin (Sigma, UK) and 2 mM L-glutamine (Sigma, UK), with medium change every 2–3 days. Cells were detached by washing the flasks twice with calcium/magnesium-free DPBS, followed by incubation with 2 mL trypsin-EDTA (GIBCO, USA) for 10 minutes. Alternatively, cells were scraped using cell scrapers (TPP, Switzerland). 1 mg/mL G418 (Fisher ThermoScientific, UK) and 600 µg/mL hygromycin B (Invitrogen, Paisley, UK) were used to maintain expression of human FcεRI genes and NFAT-luciferase, respectively. Prior to testing, cells were incubated overnight in culture medium with various dilutions of pooled serum from Parietaria judaica patients or S. mansoni infected individuals, and washed once prior to addition of the stimulus (recombinant Par j 2 from Bial, Zamudio, Spain, accession: R-17). The following positive control was used for all RS-ATL8 experiments: sensitization with 1 µg/mL of human IgE (AbD Serotec) followed by stimulation with 1 µg/mL of goat anti-human IgE polyclonal IgG (Vector Labs).
ONE-Glo Luciferase Assay System (Promega, UK) was used for all luciferase assays, following the manufacturer's instructions. Half volume (50 µL) reactions were used. Chemiluminescence was measured on an Infinite M200 microplate reader (Tecan, Männedorf, Switzerland) not later than 30 minutes after the addition of the luciferase substrate.
One-way ANOVA followed by Dunnett's or Tukey's post hoc test was performed using GraphPad Prism 6 software for Figures 3, 4, 6, 7. Spearman's rank correlation test was performed to compare IgE titres with the luminescence response in activated RS-ATL8 cells (Figure 8).
All the genes described herein were successfully amplified and ligated into the p3FA (BYDV) WG vector. The annealing temperature of 54°C was efficient for all the genes and DNA sequences were confirmed for the clones before in vitro expression. All fifteen reported parasitic genes reported herein were then successfully translated in vitro using the coupled transcription/translation WGL mixture and an aliquot of the translation mix was used for monitoring of translation by incorporation of fluorescently-tagged lysine. An example of 5 genes obtained by this method is shown in Figure 1. As seen in the negative control, protein translation with WGL results in two endogenous fluorescent components in the 15–20 kDa range (lane three in Fig. 1) which have the potential to interfere with detection of the translated parasitic protein if of this size. More examples can be seen in Supplementary data Figure S1.
In order to assess whether cell-free translation in wheat germ lysates results in proteins with unaltered antigenic properties, three unlabelled S. mansoni genes (IPSE/alpha-1, SmTAL1, SmTAL2), for which either polyclonal antisera or monoclonal antibodies were available, were expressed using wheat germ lysates and tested by immunoblotting. As shown in Figure 2, all three antigens were specifically recognised by the corresponding antibodies, demonstrating that the method chosen for expression does not appear to alter antigenicity of the parasitic proteins.
Initial experiments using cells sensitised with human monoclonal myeloma IgE (with unknown specificity) overnight and stimulated with an anti-human IgE antibody had demonstrated a high sensitivity of 10 pg IgE per assay (Supplementary Figure S2). To determine the sensitivity of the assay using polyclonal IgE in serum, a dose response curve of RS-ATL8 cell sensitised overnight with pooled sera diluted 1∶50 obtained from Parietaria judaica allergic patients, stimulated with the matching allergen Par j 2, was performed. These experiments showed that activation of basophils sensitised with this pooled serum displayed a characteristic bell-shaped curve over a wide range of allergen concentrations ranging from 10 µg/mL to 1 pg/mL, with an optimum at 100 pg/mL. An example for such dose-response curve from 1 ng/mL to 10 fg/mL is shown in Figure 3 (see Figure S3 in Supplementary data for higher concentration range).
Higher concentrations of human serum can display strong cytotoxic activities towards RBL cells, requiring the serum to be sufficiently diluted [35]. This dilution however also leads to dilution of the IgE present in the serum, potentially limiting sensitivity of the assay by reducing efficiency of sensitisation, particularly when using the traditional beta-hexosaminidase enzymatic assay for detection of degranulation and sera with low IgE titres [36].
Therefore in order to test the potential cytotoxic activity of the Par j 2-specific serum pool with and without anti-viral treatment the RS-ATL8 cells were sensitized overnight with different serum dilutions, challenging the cells with the previously determined optimal 100 pg/mL Par j 2 concentration and measuring luciferase activity after 4 hours. As shown in Figure 4, the highest concentration of untreated serum (10-fold dilution) reduced the measured luciferase activity by approximately two thirds compared with the optimal 50-fold dilution, which can be ascribed to the well documented cytotoxic effects of some human sera on RBL cells. The anti-viral treatment completely abrogates the luminescent response at higher serum concentrations but can be used efficiently in dilutions higher than 1∶100.
We assessed whether it is possible to heat sera in the presence of 2M glucose or 1M MgSO4, conditions which were shown by Binaghi [37] to prevent denaturation of IgE, without affecting successful inactivation of complement. The results are shown in Figure 5. Heating of serum in the presence of 2M glucose resulted in complete protection of IgE as judged by its ability to sensitise RS ATL8 cells and their ability to produce luciferase upon stimulation with anti-IgE or recombinant Par j 2 allergen.
Our next step was to test the ability of the reporter cell line sensitized with a serum pool from S. mansoni-infected individuals virally inactivated with Tween-80 to report engagement of the IgE receptor when stimulated with SmTAL-1 and SmTAL-2 produced in wheat germ lysates. A sensitisation to wheat allergens as a source of activation could be categorically ruled out as a source of error as activation only occurred in the presence of translated allergen but not with wheat germ lysate controls.
Figure 6 shows that stimulation of cells sensitized with a serum pool from S. mansoni-infected individuals with an optimal dilution of SmTAL-1 results in marked elevation of luciferase production, in contrast to stimulation with SmTAL-2. Interestingly, antigen expressed in wheat germ lysate resulted in stronger activation of the reporter system compared with the E. coli expressed equivalent, despite using optimal concentration from full dilution curves (Supplementary data Figure S4).
As many secretory proteins will rely on correct folding and formation of intra- or intermolecular disulphide bridges, which may affect their recognition by specific IgE, it is important to assess the ability of the wheat germ expression system to produce correctly folded parasitic antigens.
Assessing the ability of this particular protein to activate RS-ATL8 in this manner would provide further validation of the luciferase system to report IgE dependent activation events, while at the same time informing of the ability of the wheat germ lysate to produce disulphide-bridged homodimers. We therefore assessed the basophil activating properties of IPSE/alpha-1, an IgE-binding factor from S. mansoni which is known to rely on dimeric structure for its biological activity [33]. Figure 7 shows a comparison of the ability of wheat germ expressed IPSE/alpha-1 with the same protein expressed and refolded in E. coli to induce RS-ATL8 activation.
Both forms of IPSE/alpha-1 were able to dose-dependently induce reporter gene expression. However the effect was much more prominent with the bacterially-expressed refolded recombinant protein, which induced luciferase to an extent similar to the positive controls IgE/anti-IgE and Par j 2-specific serum/Par j 2. The luciferase induction by wheat germ-derived IPSE/alpha-1 was modest, suggesting that it is mainly present in its monomeric form, while the E. coli recombinant protein occurs as a mixture of monomers and dimers after refolding [33], and we have previously shown that basophil activation requires IPSE/alpha-1 in its dimeric structure [34].
As specific IgE titres will vary between individuals, we assessed the ability of 11 individual sera from infected individuals (ranging from 3.8 to 15.27 ng/mL SmTAL-1 specific IgE, as determined by ELISA) to sensitise the RS ATL8 cell line. The sera ranged from RAST class 2 (moderate IgE, 07.70–3.49 IU/mL or 1.68–8.39 ng/mL) to RAST class 3 (high IgE, 3.50–17.49 IU/mL or 8.40–41.97 ng/mL). As can be seen from Figure 8A, all sera except serum 372-01 gave a positive response upon stimulation with an optimal concentration of SmTAL-1, showing that the RS ATL8 assessment works not only with sera with high levels of allergen-specific IgE, but also with moderate specific IgE levels. All sera gave vigorous responses upon polyclonal stimulation with 1 µg/mL anti-IgE (Fig. 8 B). Non-parametric analysis (Spearman rank correlation test) indicated a statistically significant positive correlation (Fig. 9) between levels of SmTAL-1-specific IgE and the amount of luciferase produced 4 hours after stimulation.
Our aim was to establish a workflow which is suitable for medium- to high-throughput screening of potential allergenicity of S. mansoni antigens and includes a set of three steps for stringent quality controls. The first two were implemented at the DNA level (correct size of the PCR amplicon and correct sequence after cloning into expression vector), while the third (successful translation and appropriate protein size by incorporation of fluorescently-tagged lysine) was at the protein level. We chose to incorporate fluorescently labelled Lys in a separate aliquot as this fluorophore may potentially affect the antigenicity or allergenicity of the in vitro-translated product.
Wheat germ lysate is a complex biological mixture and as such it has the potential to interfere with the cellular readout. In particular, the lectin wheat germ agglutinin (WGA) may play a critical role in the context of IgE receptor cross-linking. WGA is a 18 kDa lectin contained in wheat germ [38] which naturally occurs as a 36 kDa homodimer in which the two chains are linked with 16 disulfide bonds [39]. WGA is specific for N-acetyl-D-glucosamine and the chitin oligomers chitobiose and chitotriose [39]. WGA has been shown to inhibit RBL cell activation by engaging the high affinity IgE receptor FcεRI independently of its occupancy by IgE, leading to its down regulation and inhibition of allergen-induced signal transduction [40]. In contrast, in human peripheral blood basophils, WGA can lead to basophil activation and cytokine induction (IL-4 and IL-13) by cross-linking FcεRI-bound IgE via its carbohydrate side chains [41]. In line with this finding, the addition of undiluted wheat germ lysate used in this study resulted in non-specific, high luciferase signals. The WGA can be removed from the lysate by incubation with chitin beads either prior to addition of plasmid, or by the purification of the in vitro translated antigen via its incorporated His-Tag and immobilized metal-ion affinity chromatography. However, additional steps are to be avoided in a high-throughput procedure as they increase the overall cost, duration and introduce additional sources of errors. Due to the well-known bell-shaped curve of basophil activation and the high yield of the chosen in vitro translation system, the wheat germ lysate containing the translated parasitic protein has to be diluted 1000-fold to reach the concentration range around 100 pg/mL in which stimulation is optimal (as demonstrated by the use of Par j 2 allergen and matching sera in Figure 3). With this dilution, there is no inhibition of IgE receptor crosslinking or basophil activation by WGA. This also means that only small volumes of wheat germ lysate are needed for repeated testing, reducing the overall cost of a high throughput operation.
Thus while the potential disadvantages of choosing wheat germ lysate for translation (such as the lack of glycosylation) are not relevant for this assay when assessing anti-peptide IgE, there are multiple advantages. This system is amenable to high-throughput protein synthesis because it can produce sufficiently large amounts of properly folded proteins, and it bypasses many time-consuming steps of conventional expression systems. It also allows expression of proteins that are toxic to their host organism chosen for expression. Wheat germ lysates, in contrast to e.g. E. coli lysates, have very low endogenous mRNA levels, and most of the newly translated protein is thus of parasite (or other target) rather than plant origin. This allows for very efficient incorporation of non-radioactive labels for detection or other purposes. A limitation of the wheat germ system is the inability to provide a non-reducing environment together with all necessary compartmentalised components for disulphide bridge formation. This is demonstrated by our results obtained for IPSE/alpha-1. IPSE/alpha-1 is a secretory protein produced exclusively by S. mansoni egg stage [42], which naturally occurs as a homodimer with three intramolecular disulphide bridges and one intermolecular bridge formed by the most C-terminal cysteine in position C132 [33]. We have previously shown that this homodimeric molecule is able to activate human basophils by binding IgE molecules [43], and that IgE-dependent human basophil activation is dependent on its dimerization status [34]. However this limitation does not appear to affect this molecule's antigenicity (Fig. 2).
Using a humanised rat basophil cell line also has several advantages [22], [23]. It is easier to obtain than e.g. human peripheral blood basophils, which have been notoriously difficult to purify until recently [44], and are still difficult to obtain in sufficient amounts despite these advances due to their rarity. The cells are sensitized only with the desired sera and do not require difficult IgE stripping protocols [45]. Furthermore, known potential issues such as non-responder status [46] due to down regulation of key signalling molecules such as spleen tyrosine kinase (Syk) caused by chronic exposure to low level of allergens [47], which therefore might also be occurring in helminth infection, are avoided. Non-responder status is an issue we came across in a subset of individuals when assessing peripheral blood basophil sensitization status in a cohort experimentally infected with a single dose of ten Necator americanus infective stage larvae [21], and could lead to false negative results. Also as human IgG is not thought to bind to rat immunoglobulin receptors, and the rat high affinity IgE receptor does not bind human IgE [48], [49], there is no potential for confounding factors which could mask the potential allergenicity of the studied antigens, such as competing IgG4 [50], [51], inhibitory co-crosslinking of FcεRIα and FcγRIIB [52], or activation due to IgG-IgE immune complexes [53], as these factors are removed by washing the reporter cell line prior to allergen stimulation. Finally, using the NFAT luciferase reporter for detection of activation, rather than the traditionally used β-hexosaminidase biochemical assay, results in considerably increased sensitivity. The ability of the RS-ATL8 to detect sensitisation with less than 100 pg IgE (Supplementary data S2) compares favourably with previously reported limit of 10 ng/ml upon polyclonal stimulation with an anti-IgE antibody using β-hexosaminidase activity for detection [54]. We have previously assessed alternative methods of measuring activation induced by IgE crosslinking in RBL cells (traditional beta-hexosaminidase assays, Annexin V measurements, CD63 and CD107a levels by FACS, Calcium Influx using Oregon Green or Alexa488 BAPTA-1, as well Lucifer yellow uptake, but none of these methods worked or offered any advantage over luciferase measurements.
Serum samples which had undergone anti-viral treatment with a mixture of detergents had to be diluted at least 100-fold, as higher concentrations of detergents led to complete destruction of cells as assessed by microscopy and the complete lack of luciferase induction. Tween-80 is the major constituent in the viral inactivation detergent used here and has a critical micellar concentration (CMC) of 0.012 mM [55] which lies precisely in between the concentration of Tween-80 in the 50-fold and 100-fold dilutions of treated sera. Thus a 100-fold dilution of sera will work to reduce serum cytotoxicity for both untreated and virally inactivated sera. As previously shown, serum cytotoxicity in this system is probably in part due to complement activation, as heating the serum at 56°C for 30 min reduced its cytotoxicity [35]. This treatment however irreversibly denatures IgE, abrogating its binding to the high affinity IgE receptor [56]. Our results demonstrate that 2M glucose treatment might represent a suitable way of inactivating complement in sera without leading to loss of IgE functionality. However, subsequent experiments with glucose-treated sera clearly showed that these also had to be diluted 20–50-fold to avoid deleterious effects of the high glucose concentration in the cellular assay, or required dialysis-based methods. As this would complicate the workflow unnecessarily, we did not pursue these attempts, and used serum dilutions of 1∶50 with virally non-inactivated sera or 1∶100 to 1∶200 with virally inactivated, detergent-treated sera. The work described in Figures 3–5 and S2, was performed using a well-characterised (ISAC UniCAP, clinical history) serum pool from P. judaica allergic individuals, rather than sera from parasite-infected individuals and helminthic allergens. This was mainly due to the unavailability of large amounts of infection serum required for such studies, but also as it allowed us to validate the technology against the current gold standard for specific IgE determination. The results are equally relevant for IgE/allergen combinations studied in a tropical parasite infection context, as the underlying mechanisms of sensitisation, cellular activation and allergenicity are fundamentally the same [57].
The SmTAL proteins are a family of 13 closely related allergen-like molecules. Of these SmTAL1 and 2 have the greatest similarity in amino acid sequence (48% identity). However, in populations from endemic areas, SmTAL1 is reported to be the dominant IgE-inducing antigen whilst an IgE response to SmTAL2 is rare [26], [31]. It has been proposed that this is because the response to egg antigen – SmTAL2 – is desensitized by continuous exposure (as eggs die in the host tissues every day); whilst internal adult worm antigen SmTAL1 is only exposed infrequently when adult parasites die [26]. An interesting observation was made when comparing the allergenicity of SmTAL-1 and SmTAL-2 between the bacterial recombinant forms and their wheat germ expressed counterparts. We carried out full titration curves with all four but found that in both cases the WG-expressed form induced significantly higher reporter cell activation than the E. coli expressed SmTALs (Supplementary data S4). The reasons for this difference are not clear, but would appear to rule out that LPS contamination of bacterially expressed recombinant allergens could be a source of basophil activation in the used assay.
Taken together, this method offers a robust way for assessing potential allergenicity of S. mansoni (or any other parasite) in a format suitable for high-throughput analysis. The novelty of the method presented here lies in the combination of a fast cell-free expression system and an equally fast reporter system which allows expression of candidate allergens in a few hours and detection of activation within three hours, all up-scalable to high-throughput format. This method can be used as an additional safety test when assessing potential vaccine candidates. Perhaps more importantly, when used at the whole genome level, it could be used to unravel the entire allergome of S. mansoni and other medically important parasites. Ultimately this could lead to a better understanding of the basis of allergenicity, and in combination with additional cellular studies, to a better understanding of the relationship between parasite-specific IgE and host protection mechanisms at the molecular level [57].
|
10.1371/journal.pntd.0005445 | Geospatial distribution of intestinal parasitic infections in Rio de Janeiro (Brazil) and its association with social determinants | Intestinal parasitic infections remain among the most common infectious diseases worldwide. This study aimed to estimate their prevalence and provide a detailed analysis of geographical distribution of intestinal parasites in the metropolitan region of Rio de Janeiro, considering demographic, socio-economic, and epidemiological contextual factors.
The cross-section survey was conducted among individuals attending the Evandro Chagas National Institute of Infectious Diseases (FIOCRUZ, RJ) during the period from April 2012 to February 2015. Stool samples were collected and processed by sedimentation, flotation, Kato-Katz, Baermann-Moraes and Graham methods, iron haematoxylin staining and safranin staining. Of the 3245 individuals analysed, 569 (17.5%) were infected with at least one parasite. The most common protozoa were Endolimax nana (28.8%), Entamoeba coli (14.8%), Complex Entamoeba histolytica/Entamoeba dispar (13.5%), Blastocystis hominis (12.7%), and Giardia lamblia (8.1%). Strongyloides stercoralis (4.3%), Schistosoma mansoni (3.3%), Ascaris lumbricoides (1.6%), and hookworms (1.5%) were the most frequent helminths. There was a high frequency of contamination by protozoa (87%), and multiple infections were observed in 141 participants (24.8%). A positive association between age (young children) and gender (male) with intestinal parasites was observed. Geospatial distribution of the detected intestinal parasitic infections was not random or homogeneous, but was influenced by socioeconomic conditions (through the material deprivation index (MDI)). Participants classified in the highest levels of deprivation had higher risk of having intestinal parasites.
This study provides the first epidemiological information on the prevalence and distribution of intestinal parasitic infections in the Rio de Janeiro metropolitan area. Intestinal parasites, especially protozoa, are highly prevalent, indicating that parasitic infections are still a serious public health problem. MDI showed that intestinal parasites were strongly associated with the socioeconomic status of the population, thus making it possible to identify social vulnerable areas.
| Intestinal parasitic infections are considered indicators of health and socio-environmental vulnerability, and are associated with precarious sanitation and water quality of a country. They continue to pose a serious public health problem, especially in developing countries where sanitation is not expanded in line with population growth, such that access to basic services becomes more difficult. Although Brazil is a country with a high prevalence of intestinal parasitic infections, the prevalence in the metropolitan region of Rio de Janeiro (the second largest metropolitan area in the country) has not been estimated. Based on the identification of social determinants (income, education and sanitation), our group was able to identify vulnerable areas for intestinal parasitic infection in the metropolitan region of Rio de Janeiro. Infections caused by intestinal parasites are not included in the list of diseases compulsory notification in Brazil. However, special attention should be focused on this topic, and information on the geographic distribution and prevalence of intestinal parasites, as well as the recognition of vulnerable areas, are the first steps, and a prerequisite for development of appropriate control strategies by the government.
| Neglected tropical diseases, including intestinal parasitic infections, are a significant cause of morbidity and mortality in endemic countries [1]. Intestinal parasitic infections have particular relevance as they affect the poorest and most deprived areas in tropical and subtropical regions [1]. It is increasingly recognized that both protozoan and helminthic diseases are common among children under the age of five years. Children are more vulnerable to soil-transmitted helminths (STHs) than adults, and the nutritional impairment caused by the parasite can lead to iron-deficiency anaemia, malnutrition, and a negative impact on growth and cognitive development [2,3].
Despite all the medical and pharmaceutical advances and developments in sanitary engineering, intestinal parasitic infections remain among the most common infectious diseases worldwide, particularly in developing countries, where inadequate water treatment, poor sanitation and lack of adequate health services are common. Additionally, it is more difficult to implement enteric parasite-control actions in these regions due to the high cost of improvements in infrastructure, and the lack of educational projects offered to the population [1,4,5].
Water is essential to life, but is also a major vehicle for pathogen dissemination. The potential for waterborne parasite transmission is high since infective helminth eggs and protozoa (oo)cysts are distributed through water in the environment. Pathogens like Giardia lambia and Cryptosporidium spp. are recognized as important waterborne disease pathogens and are associated with severe gastrointestinal illness. Amoebiasis, balantidiosis, cyclosporidiosis and microsporidiosis outbreaks have been reported throughout the world [6,7]. It is well documented that conventional water and sewage treatment process are not completely effective in destroying protozoa (oo)cysts and helminth eggs [8–10]. Improper disposal of human and animal waste has also been identified as a source of infection, contaminating water sources [11] and recreational waters such as swimming pools, water parks and lakes [9]. Occasionally, sewer overflows also contribute to contamination of surface water and agricultural lands, which leads to potential human infection. Food contamination is also important and can occur directly in the handling process (contaminated equipment, infected food handlers or wash water), or indirectly through contaminated irrigation water [12].
The lack of sanitary conditions to which the population is exposed favours the acquisition of various pathogens, and patients are often multiply infected (polyparasitized). Recently, a systematic review and meta-analysis showed that sanitation facilities and water treatment are associated with lower risks of infection with intestinal protozoa, and could also prevent diarrhoeal diseases [1]. The same relationships were observed by Strunz et al. [13] for soil-transmitted helminths.
In Brazil, intestinal parasite infections persist, although their frequency has decreased due improvement of sanitary conditions [14–16]. Up until now, studies of enteric parasites in Brazil have been limited, isolated and fairly rare, generally reflecting the situation in small towns. Mariano and colleagues [17] observed 77.2% of positive cases, and a polyparasitism of 51.2% in children from Itabuna (Bahia). Similar results were observed in two localities of São Paulo, where 65.9% of the individuals were positive for at least one parasite [18]. In Rio de Janeiro, previous studies have shown intestinal parasite prevalence ranging from 18.3% to 66% [19–24].
The aim of this study was to estimate the number of individuals infected with intestinal parasites who attended a referral hospital located in Rio de Janeiro (Brazil), and to provide a detailed analysis of the geographical distribution. The study also looked at the influence of demographic variables, socio-economic status and environmental factors on the intestinal parasitic infections. This knowledge will be essential for the development of effective prevention and control strategies to eliminate or reduce intestinal parasitic infection.
The Research Ethics Committee Evandro Chagas National Institute of Infectious Diseases (INI/FIOCRUZ) approved the study (protocol number: 127.542). This project was in accordance with the Brazilian Ethical Resolutions, especially Resolution CNS 196/1996 and its complementary and the Code of Medical Ethics of 1988 (articles 122–1307). Study individuals provided a written signed informed consent prior to sample collection and for participants younger than 18 years, informed consent was provided by parents or guardians after a detailed explanation of the objectives of the work. A term of privacy and confidentiality was signed by the researches for individuals to whom it was not possible to obtain informed consent beforehand.
The cross-section survey was carried out from April 2012 to February 2015 in Evandro Chagas National Institute of Infectious Diseases (INI/FIOCRUZ), a reference hospital in infectious diseases in Brazil, located in Rio de Janeiro (RJ). Despite it being an infectious disease referral hospital, individuals also attend for routine consultations (cardiology, dermatologist, gynecology, neurology, ophthalmology, otolaryngologist, infectious disease speciality) or emergency situations. As the prevalence of intestinal parasites in Brazil remains high, it is common the doctor´s submit requests for parasitological analysis in faeces, regardless of age or genera and of having or not symptoms suggestive of intestinal infections. The INI/FIOCRUZ hospital receives individuals from all municipalities, mainly the metropolitan area.
Rio de Janeiro State is composed of 92 municipalities. The metropolitan region of Rio de Janeiro is composed of 21 municipalities: Belford Roxo, Cachoeira de Macacu, Duque de Caxias, Guapimirim, Itaboraí, Itaguaí, Japeri, Magé, Maricá, Mesquita, Nilópolis, Niterói, Nova Iguaçu, Paracambi, Queimados, Rio Bonito, Rio de Janeiro, São Gonçalo, São João de Meriti, Seropédica and Tanguá (Fig 1). It is the second largest metropolitan area in Brazil with 11.812.482 inhabitants in an area of 8.147.356 km2. This region has 2.746 slums, with a resident population of 1.702.073 inhabitants (14.4% from the total population) occupying 123.627km2 [25]. The main characteristics of each municipality of the metropolitan region of Rio de Janeiro State are summarized in Table 1.
According to the last census conducted in 2010, Rio de Janeiro municipality has a population of 6.320.446 inhabitants (Table 1) in an area of 1.197.463 km2. The municipality has 2.227 slums, with a resident population of 1.393.314 inhabitants (11.8% from the total population) occupying 54.213 km2 [25].
Municipal human development index (MHDI) is a summary measure of average achievement in key dimensions of human development (a long and healthy life, being knowledgeable and have a decent standard of living), and gini index is a measure of statistical dispersion whose value ranges from zero (perfect equality) to one (perfect inequality). The MHDI of Rio de Janeiro is 0.799 according to the United Nations Development Programme [26] and gini index is 0.6391 [25]. Most of the population (91.2%) has access to potable water and 70.1% has sanitation coverage [27].
The study population included individuals (n = 3245), of both genders and all age groups, attended in Evandro Chagas National Institute of Infectious Diseases, between April 2012 and February 2015. Stool samples were collected by the participant in plastic disposable flasks with or without preservatives and maintained at 4°C until laboratory analysis on the same day. Flasks were labelled with the name, collection date and the hospital number. The parasitological tests were conducted at the Parasitology Laboratory of INI by experienced laboratory technologists and College of American Pathologist certifies the Laboratory. Moreover, participant’ data (sex, age, educational level and residence) were obtained from the hospital’s database.
For laboratory diagnosis of intestinal parasites, the fresh specimens were analysed by means of centrifugation sedimentation [28], centrifugal flotation in zinc sulphate solution [29], Kato-Katz (Helm-TEST kit, Fiocruz, Brazil) and Baermann-Moraes method [28,30]. All these techniques were routinely performed on all fresh stool samples. Specimens preserved in MIF solution (mertiolate-iodine-formaldehyde) were processed by the centrifugation sedimentation method [28]. The Graham method, faecal occult blood test, the iron haematoxylin staining and the safranin staining procedure was carried out depending on doctor request [28]. The slides were then observed under the optical microscope.
All individuals attended in INI/FIOCRUZ are dewormed when diagnosed (drugs are provided by the institution itself).
The zip code for each participant was obtained from the hospital’s database and through Brazilian Institute of Statistics and Geography (IBGE) converted into geographic coordinates (latitude and longitude). IBGE was the source of data in respect of geography, demography and socioeconomic conditions of the studied population (National Census of 2010).
The spatial distribution of the participants was assessed through a Kernel Density Function that allows to estimate the intensity of events across a surface by calculating the overall number of cases within a given search radius from a target point. To identify if the participants were spatially clustered or dispersed the Average Nearest Neighbor test was used.
To evaluate the social and economic conditions of the place of residence a material deprivation index (MDI) was constructed, at the census tract level, to the metropolitan region of Rio de Janeiro. The MDI is based upon the following indicators: (1) illiteracy rate/education (percentage of population older than 10 years that can read or write); (2) water supply/sanitation (percentage of permanent households without public water treatment plant); and (3) family income (percentage of households with per capita monthly income ≤1 minimum wage). Based on the Carstairs and Morris method, the indicators considered in each index were standardised (using the z-score method) so that each indicator has a weighted mean of zero and a variance of one, and exerted the same influence upon the final result [31]. The MDI was analysed in quintiles: q1, lowest level of deprivation; q5, highest level of deprivation.
To address the potential effects of the socioeconomic conditions of the place of residence on the incidence of intestinal parasites, the proportion of participants living in each deprivation quintile was assessed. Simultaneously, the proximity to slums was analysed through geographical buffers of 50m and 100m. The spatial analysis was performed through the ArcMap 10.x software of ESRI.
The data entry was carried out using Excel software and analysed using Statistical Package for the Social Sciences (SPSS) version 16. Percentages were used to perform the exploratory analysis of the categorical variables and quantitative variables are presented as mean ± standard deviation (SD). Pearson´s chi-squared and Fisher’s Exact Test were used for categorical data. The level of statistical significance was set as p<0.05, an odds ratio and 95% confidence interval (CI) was computed. Logistic regression was used to identify a potential contribution of each of the variables for the acquisition of intestinal parasite infections.
Between April 2012 and February 2015, a total of 3245 individuals (1564 female and 1681 male) had the parasitological tests done (Table 2). In 2012 a total of 995 samples were collected, with 193 positive samples; in 2013, 1189 individuals were collected being 187 positive samples; in 2014, 938 individuals with 168 positive samples; and in 2015, 123 individuals with 21 positive samples. Summarizing, we had 569 individuals (17.5%) with positive stool examination for one or more enteric parasite and 2676 individuals (82.5%) with negative results.
The ages of the participants ranged from 1 to 93 years with an average of 41.34±15.54 (Mean±SD; median = 41). The adults between 26–65 years were the majority of participants (n = 2130) (Table 3). There were more male than female parasitized (64.5% versus 35.5%, respectively) and seventy-five percent of participants (n = 427) were educated above the primary grade (Table 3).
Endolimax nana was the most common enteric parasite, present in 216 samples (28.8%) followed by Entamoeba coli in 111 samples (14.8%), Complex Entamoeba histolytica/Entamoeba dispar in 101 samples (13.5%), Blastocystis hominis in 95 samples (12.7%), Giardia lamblia in 61 samples (8.1%), Iodamoeba butschilii in 33 samples (4.4%), Strongyloides stercoralis in 32 samples (4.3%), Schistosoma mansoni in 25 samples (3.3%), Cryptosporidium sp. in 14 samples (1.9%), Ascaris lumbricoides in 12 samples (1.6%), Cystoisospora belli in 12 samples (1.6%), hookworms in 11 samples (1.5%), Trichuris trichiura in 10 samples (1.3%), Entamoeba hartmani in 9 samples (1.2%), Enterobius vermicularis in 6 samples (0.8%) and Hymenolepis nana in one sample (0.1%) (Table 4).
The number of samples with one parasite (monoparasitism) is higher (428 positive samples, 57.1%) than those samples with various parasites (polyparasitism) (321 positive samples, 42.9%). Interesting, the frequency of the amoebae (Complex E.histolytica/E.dispar, E. coli and E. hartmani) as well of some geohelminths (A. lumbricoides and T. trichiura) is higher on samples with various parasites (polyparasitism) (Table 4).
We observed a very high frequency of protozoan infections (87%), occupying the first six positions; E. nana was the predominant, followed by E. coli and Complex E. histolytica/E. dispar. The most frequent helminths were S. stercoralis and S. mansoni; only appearing in seventh position. Of the 16 species of intestinal parasites detected, 11 were pathogenic (Complex E. histolytica/E. dispar, Cryptosporidium sp., C. belli, G. lamblia, A. lumbricoides, E. vermicularis, H. nana, hookworms, S. mansoni, S. stercoralis and T. trichiura) and 5 were non-pathogenic (B. hominis, E. nana, E. coli, E. hartmani and I. butschilii). The pathogenic species comprises 38.1% of the studied participants (285 of 749), while the non-pathogenic reached 61.9% (464 of 749).
Most of the participants (428 of 569; 75.2%) did not present any co-infection, whereas 141 (24.8%) had two or more parasites simultaneously. Among the multiple infected, 109 individuals were infected with two parasites (19.2%), 26 were infected with three parasites (4.6%), 5 had four parasites (0.9%) and 1 had five (0.1%). Regarding parasitic associations, only 11.8% (67 of 569) were co-parasited by helminths, 84.3% (480 of 569) by protozoa and only 3.9% (22 of 569) by both.
Age and gender were examined as potential associations for intestinal parasitic infections. A positive association between gender and intestinal parasites (p<0.0001), as well as protozoa (p<0.0001), helminths (p<0.0001) and poliparasitism (p<0.0001) were detected. Male were more likely to be infected with intestinal parasites (OR = 1.9; 95%CI of 1.56 to 2.27), protozoa (OR = 1.8; 95%CI of 1.50 to 2.20), helminths (OR = 2.8; 95%CI of 1.75 to 4.51) and have multiple parasites (OR = 3.4; 95% CI of 2.28 to 5.05) compared to female (Table 5).
No statistical significant difference was found between intestinal parasites and age (p = 0.166). However, when we analyse the parasite species separately we observed that children (0–14 years) were more likely to be infected with A. lumbricoides (p = 0.031; OR = 8.5; 95% CI = 1.8; 39.4), E. vermicularis (p = 0.005; OR = 28.2; 95% CI = 4.6; 171.6), B. hominis (p = 0.002; OR = 3.9; 95% CI = 1.8; 8.4), and G. lamblia (p = 0.011; OR = 4.1; 95% CI = 1.6; 10.7) as compared to the older participants (S1 Table). Moreover, there were no cases of multiple parasitic infections in children under 5 years old (S1 Table).
The prevalence of intestinal parasites varies by municipalities, most of participants (2847 of 3245; 87.7%) live in metropolitan region and 1748 (53.9%) live in Rio de Janeiro municipality (Tables 6 and S2). The metropolitan region of Rio de Janeiro had 532 positive cases (16.4%) and the others municipalities had 21 positive cases (0.6%) (Table 6). As expected, Rio de Janeiro municipality had a greater number of participants infected with intestinal parasites (332; 10.2%) since it has the larger population (S2 Table, S1 Fig). In 16 participants (0.5%) positive for intestinal parasites was not possible to identify the residence.
The distribution of parasites species also varied among the municipalities (Table 7). The metropolitan region had 93.7% (702 of 749) of the enteric parasites observed: in Rio de Janeiro it was possible to detect 434 enteric parasites (57.9%), Duque de Caxias was the second municipality with 81 (10.8%), followed by Nova Iguaçu (57; 7.6%), Belford Roxo (33; 4.5%), São João de Meriti (25; 3.4%), São Gonçalo (18; 2.4%), Nilópolis (15; 2%), Magé (12; 1.6%), Cachoeira de Macacu (5; 0.7%), Itaboraí (4; 0.5%), Niterói (3; 0.4%), Queimados (3; 0.4%), Itaguaí (2; 0.3%), Maricá (2; 0.3%), Mesquita (2; 0.3%), Seropédica (2; 0.3%), and Japeri (4; 0.5%). We did not have positive samples from participants of Guapimirim, Paracambi, Rio Bonito and Tanguá. Others municipalities amounted 28 (3.7%) enteric parasites, and 19 (2.5%) was not possible to identify the municipality (Table 7).
The current study estimated the prevalence of intestinal parasitic infections among individuals from Rio de Janeiro State (Brazil), in addition to evaluating some epidemiological aspects. Spatial analysis was applied for the first time to the case of Rio de Janeiro to describe the geographical distribution of individuals with enteric parasites infections. The study also looked at socio-economic indicators (social vulnerability indicator) for intestinal infections, in particular family income, education and sanitation (access to safe drinking water). The construction of a material deprivation index allowed us to identify the most vulnerable regions for intestinal parasitic infections in the metropolitan area of Rio de Janeiro State.
The mean prevalence of intestinal parasitic infections remains high in Rio de Janeiro State (17.5%) and also in the metropolitan region and the municipality (18.7% and 19%, respectively). Previous studies suggest that we may observe a decrease in the prevalence of intestinal parasites in Rio de Janeiro with time. A parasitological survey carried out in 1984 on children from day-care centres detected a prevalence of 35% [19]. Further studies carried out on pregnant women [20], children living in low income communities [21] and day-care centres located in slums in the municipality [22] showed a prevalence ranging from 37.6% to 54.5%. A survey made in 2007 in a paediatric hospital [23] detected values of 18.3%. Although our results indicate that the mean prevalence is similar to this last study, it should be noted that individuals attending the Evandro Chagas National Institute of Infectious Diseases (INI/FIOCRUZ) were mainly adults, where it was expected that prevalence would be lower when compared to studies on children.
Age is an important risk factor for intestinal parasitic infections. Children are more susceptible to intestinal infectious diseases than adults because of their poor hygiene habits; they are often in contact with contaminated soil and their immune system is immature [2,32]. In spite of our small number of samples from young participants, we observed a positive association between infections with A. lumbricoides, E. vermicularis, B. hominis and G. lamblia and the younger age.
The distribution of intestinal parasites varied among the municipalities that compose Rio de Janeiro State, with the highest incidence density of intestinal parasites in municipalities with larger population (Rio de Janeiro, Duque de Caxias, Nova Iguaçu, Belford Roxo and São João de Meriti). These results could be explained by the ease of access to the INI hospital, since these areas have the main road corridors of the municipality (Brazil Avenue, Governador Carlos Lacerda Avenue, Presidente João Goulart Avenue and Presidente Dutra highway), and also because many of the infected population lives near INI hospital. Despite São Gonçalo is the second largest municipality, only 2.4% (18 of 749) of the intestinal parasites were detected there. This municipality is located across the Guanabara Bay, such that access by participants to the INI hospital is probably limited by poor public transportation.
The prevalence of enteric parasites varies between regions of Brazil, and contrasting data are observed: 11.3% in Sergipe [33]; 42% reported from São Paulo (southeast) [34]; 73.5% in Mato Grosso do Sul (midwest) [35]; 75.3% in Paraná (south) [36]; 77.2% in Bahia (northeast) [17]. However, data extracted from previous studies in Brazil should be analysed with some caution, once they were limited, isolated, and usually reflect the results from small towns and/or of restricted groups (day-care centres, schools, indigenous tribes, small hospitals, fishing villages, etc.). Attention should also be given to studies conducted in other countries: Argentina (78.3%) in children living in a poor area [37]; Peru (66.3%) in orphanages [38]; Honduras (43.5%) in school going children [39]; Pakistan (52.8%) in children residing in slum areas [5]; and India (68%) in school going children [40].
In the present work, the most common pathogenic species detected were Complex E. histolytica/E. dispar (13.5%) and G. lamblia (8.1%). These two parasites are frequently found in Brazil [17,18,35,41]. However, detection of G. lamblia cysts is particularly alarming since these are resistant to conventional routine disinfectants, and are frequently found in sewage effluent and surface water [10]. In addition, individuals infected with G. lamblia are largely asymptomatic, and can spread the infection, contributing to high epidemic rates. Similarly, concern should also be given to the presence of B. hominis (12.7%), since its pathogenicity is still controversial [42]. In Minas Gerais (Brazil), Cabrine-Santos and colleagues [43] observed that 8% of participants with diarrhoea had only Blastocystis spp. (monoparasitism); suggesting that the parasite may have a pathogenic character.
Although soil-transmitted helminths (A. lumbricoides, T. trichiura, hookworms and S. stercoralis) are the most frequent parasites found in many countries [32, 39], they were not the predominant enteric parasites in this study. Probably these parasites cannot complete their life cycles due the absence of an adequate soil environment or the presence of road/sidewalk paving or a high construction index [16]. The low prevalence of S. mansoni infections was also observed. The transmission of S. mansoni is dependent on the presence of a water and an intermediate host snail, which may be not available in the areas of this study. According to the Brazilian Ministry of Health [44], the positive rate of S. mansoni in Rio de Janeiro State is 1.56%, making it the State with the lowest number of confirmed cases.
We noticed a positive association between intestinal parasites and the male gender. Similar results are observed in Brazil [43] and Iran [45,46], with a slightly higher prevalence of intestinal parasites in males than females. This association could be due to hygienic behaviours, specific occupations or even sexual activities, particularly among homosexuals, that may result in faecal/oral contact that subsequently leads to transmission of these agents [47,48].
Chemotherapy is one of the intervention strategies that reduce the incidence of intestinal diseases. Regular deworming with the drugs albendazole and mebendazole is the current global control strategy to reduce the prevalence of helminths, and is implemented in Brazil [44]. However, the deworming programmes are not effective against protozoa infections. In this study we clearly observed that the frequency of protozoan infections (87%) was much higher than that of helminths (13%). It is worth mentioning that nitazoxanide is an antiparasitic drug with a broad-spectrum activity against a variety of intestinal parasites (including protozoa and helminths). However, this product is not included in the list of pharmaceutical care products of the Unified Health System (SUS) in Brazil.
A number of individuals (141; 24.8%) were infected by multiple enteroparasites: 3.5% (5 of 141) of participants were infected with helminths, 80.9% (114 of 141) were infected with protozoa and 15.6% (22 of 141) by both. Polyparasitism remains persistent in the country: 18.4% of such cases were reported in São Paulo [34], 49.2% in Mato Grosso do Sul [35], 26.7% in Paraná [36], 51.2% in Bahia [17]. These works all showed the high frequency of protozoa. Polyparasitism had been observed in many countries [5,49,50]; for example, in Kenya, 7% of the study population was infected with multiple parasites [32], and Mejia Torres et al. [39] observed that 14.6% of children in Honduras were infected with more than one parasite.
This study confirms that the population has a high frequency of intestinal parasites, principally protozoa. Although the majority of parasites (62%) were non-pathogenic (B. hominis, E. coli, E. hartmani, E. nana and I. butschilii), it is important to note that these species have the same transmission path as other pathogenic protozoa, such as Complex E. histolytica/E. dispar and G. lamblia, indicating exposure to faecal contamination. The frequency of these parasites added to the high frequency of polyparasitism can be used as indicators of transmission through the faecal/oral route, thereby pointing to in the transmission of intestinal parasites via the supply of water for human consumption, or the ingestion of contaminated food.
Several authors have demonstrated the vulnerability of drinking water supply systems due contamination, which can lead to problems, such as the deterioration of water quality, which lead to the proliferation of pathogens, and, therefore, increase the risk of waterborne diseases [51,52]. Water for the citizens of the metropolitan region of Rio de Janeiro is provided by two principal supply systems, called Guandu-Piraí and Imunana-Laranjal. Both of these undergo the conventional treatment process, including coagulation, flocculation, filtration (granulated active carbon), fluoridation and chlorination [53]. Two companies carry out the operation and management of the water systems, one of which is public (State Company of Water and Sewage—CEDAE) and the other is a concession (Niterói Water). The Niterói Water Company only operates on the distribution of treated water, which is supplied by CEDAE from the water collected in the Imunana-Laranjal system. Although both systems operate satisfactorily, in agreement with Brazilian standards of technical quality and health [54], water distribution generally has problems inherent in the characteristics of the use and occupation of urban land in the metropolitan region of Rio de Janeiro, particularly in the municipalities and neighbourhoods with higher levels of social and economic inequality. In these areas, lack of access to collection services and sewage treatment leads to the contamination of the water supply network through cross connections and low pressure zones, thereby leading to the entry of sewage and rainwater into the system. This situation is exacerbated in neighbourhoods and slums located in higher areas, where the pressure in the network is insufficient to maintain a constant water flow, and, according to the Brazilian Standard, drinkability [55,56].
Although we did not directly investigate this matter, we know that in developing countries, such as Brazil, access to clean water, sanitation facilities and health infrastructure does not follow the population growth. Research conducted in two low-income communities of Campos dos Goytacazes (north of Rio de Janeiro State/Brazil) confirmed by water analysis that the entire underground water of the study area was contaminated and a high faecal contamination was detected in well water. The authors concluded that possibly inadequate sanitation, with sewage discharged directly into the soil in some points, visible leakage, along with inadequate, and negligent routine maintenance in some septic tank systems could certainly have contributed to the dissemination of diseases caused by parasites [57].
The high prevalence of intestinal parasitic infections is also closely related to the low level of education, the low household incomes family and improper hygienic practices [4,5,57]. This study evaluated the socio and economic conditions of the Rio de Janeiro population using an index of material deprivation (MDI) composed of three indicators (sanitation, income and education).
The Rio de Janeiro metropolitan area is comprised of many census tracts (CTs), very close together and with very different MDIs, resulting in the highly heterogeneous character of the Rio de Janeiro territory. For example, while the INI hospital was classified in the first deprivation quintile (q1), a large part of the resident population in its surroundings live in slums or very poor neighbourhoods and was classified in the last deprivation quintiles (q4 or q5). Such proximity of participants to slums makes them more likely to be infected with intestinal parasites. Clearly, the geospatial distribution of the detected intestinal parasitic infections was not random or homogeneous, but was influenced by the MDI and the proximity to INI. Discrepancy of the MDIs among the closest CTs reveals the need for a horizontal decision-making process, not only in the poorest areas of the municipality, but throughout their surrounding areas.
Improvements in sanitation systems, deworming and the creation of poverty reduction programmes (Bolsa Família and Favela Bairro Program) in Brazil have helped greatly to reduce the prevalence of intestinal parasites over the years, but much obviously remains to be done. Safe drinking water is a defining aspect of a developed country, and even today it is still a significant challenge to public health worldwide. Additionally, the lack of access to health services near their home forces individuals to travel great distances to demand medical treatment, and, in many cases, the lack or deficiencies in public transport prevents these people from accessing the medical units.
Access to medical care, preventative chemotherapy and improvements in water supply and sanitation are matters of urgency, and also require a massive education campaign for low and middle-income families. Water of good microbial quality must be continuously supplied to the households (avoiding storage, which is another factor for contamination), and thus preventing its theft. Diseases are not distributed occasionally or randomly, the existence of risk factors determines their distribution, so that constant and continuous monitoring is required. Efforts directed to build a health surveillance system are urgent for Rio de Janeiro, and require strategies based on: sanitary conditions, water supply, population vulnerability, socio-demographic and environmental factors such as age, gender, education, household characteristics and income. Knowing the geographical distribution of intestinal parasites in Rio de Janeiro population is an important first step that will assist in the decision-making process necessary to design effective preventive and control programs; however, more epidemiological studies are imperative. The ability to readily identify and reach individuals at highest risk of infection is an important aspect of parasitic disease control programmes.
|
10.1371/journal.pgen.1006750 | TCF21 and the environmental sensor aryl-hydrocarbon receptor cooperate to activate a pro-inflammatory gene expression program in coronary artery smooth muscle cells | Both environmental factors and genetic loci have been associated with coronary artery disease (CAD), however gene-gene and gene-environment interactions that might identify molecular mechanisms of risk are not easily studied by human genetic approaches. We have previously identified the transcription factor TCF21 as the causal CAD gene at 6q23.2 and characterized its downstream transcriptional network that is enriched for CAD GWAS genes. Here we investigate the hypothesis that TCF21 interacts with a downstream target gene, the aryl hydrocarbon receptor (AHR), a ligand-activated transcription factor that mediates the cellular response to environmental contaminants, including dioxin and polycyclic aromatic hydrocarbons (e.g., tobacco smoke). Perturbation of TCF21 expression in human coronary artery smooth muscle cells (HCASMC) revealed that TCF21 promotes expression of AHR, its heterodimerization partner ARNT, and cooperates with these factors to upregulate a number of inflammatory downstream disease related genes including IL1A, MMP1, and CYP1A1. TCF21 was shown to bind in AHR, ARNT and downstream target gene loci, and co-localization was noted for AHR-ARNT and TCF21 binding sites genome-wide in regions of HCASMC open chromatin. These regions of co-localization were found to be enriched for GWAS signals associated with cardio-metabolic as well as chronic inflammatory disease phenotypes. Finally, we show that similar to TCF21, AHR gene expression is increased in atherosclerotic lesions in mice in vivo using laser capture microdissection, and AHR protein is localized in human carotid atherosclerotic lesions where it is associated with protein kinases with a critical role in innate immune response. These data suggest that TCF21 can cooperate with AHR to activate an inflammatory gene expression program that is exacerbated by environmental stimuli, and may contribute to the overall risk for CAD.
| Coronary heart disease is the leading cause of death in the world. Both genes and the environment are important risk factors for the progression of disease, however, how genes may modulate the harmful response to the disease promoting environment is unknown and difficult to study. Here, we show that a common heritable variation in the gene TCF21 may regulate coronary heart disease risk by regulating the response of downstream gene activation by the disease environment. We find that a well-known environmental sensor, aryl-hydrocarbon receptor (AHR), is regulated by TCF21 and also interacts with TCF21, resulting in regulation of pro-inflammatory gene expression in coronary artery smooth muscle cells. We further show that oxidized LDL, a well-known driver of atherosclerosis in the plaque can activate the AHR pathway. This work describes a heritable form of gene-environment interaction identified through genome wide association studies in coronary artery disease, and presents an opportunity to define causal gene-gene and gene-environment interactions.
| Genome-wide association studies (GWAS) have identified susceptibility loci and candidate genetic variants that predispose to atherosclerotic coronary artery disease (CAD) in humans.[1–4] Despite significant advances made in mapping the genetic contribution to CAD, there has been limited progress toward understanding molecular mechanisms leading to increased atherosclerosis susceptibility that are mediated through gene-environment (GxE) interactions.[5] The difficulty in identifying the role of genetic variation in the differential response to environmental exposure stems from inaccurate quantification of the exposure, the inability to isolate the exposures of interest, and the lack of statistical power.[6]
GWA studies have identified variation at 6q23.2 to be associated with CAD in Caucasian and Han Chinese populations[1, 7], and work in this lab has identified TCF21 as the causal gene in this locus.[8, 9] Mechanistic studies employing lineage tracing in murine disease models have found that Tcf21 expression is localized to the medial and adventitial layers of the coronary vessel wall at baseline, and that Tcf21 expressing cells migrate through the lesion and contribute to the fibrous cap as disease progresses.[10] These data, in combination with in vitro studies indicating that TCF21 inhibits differentiation and promotes SMC proliferation, suggest a role for this transcription factor in the phenotypic modulation of medial SMC in the response to vascular injury.[10, 11] Further, our RNA-seq and ChIP-seq studies have shown that TCF21 binds and regulates a network of genes associated with CAD.[12] We discovered one of the central components of the TCF21 gene network to be the aryl hydrocarbon receptor (AHR), a transcription factor that mediates the response to environmental toxins and xenobiotics, and is known to regulate the inflammatory cellular response.[13–17]
AHR binds to a complex array of nuclear proteins involved in diverse processes related to signaling through hormone receptor and inflammatory pathways, chromatin remodeling, etc., and activates a number of target genes, including cytochromes P450 (CYP1A1 and CYP1B1), and AHR repressor (AHRR).[18, 19] AHR is active primarily in the liver, however it is also strongly expressed in the cardiovascular system, where it has been described to play a role in the cardiovascular development and vascular remodeling.[20–22] In the context of environmental stimuli, AHR ligands include a wide range of environmental pollutants, including 2,3,7,8-tetrachlorodibenzo-p-dioxin (dioxin), co-planar polychlorinated biphenyls (PCBs), and polycyclic aromatic hydrocarbons (PAH) which are major constituents of tobacco smoke.[15, 22–24]
The correlation of major cardiovascular risk factors with the AHR pathway relates to epidemiological evidence that dioxin exposure is linked to increased cardiovascular mortality.[25] Furthermore, murine model studies have shown that mice carrying an AHR variant with higher ligand affinity developed more severe atherosclerosis compared to wild-type mice[22], and an increase in disease burden when exposed to dioxin.[26] In humans, a common SNP associated with AHR was found to correlate with the CAD phenotype in a Chinese population.[27] In addition, the expression level of AHR in circulating peripheral mononuclear cells was associated with acute coronary syndromes (ACS), suggesting that greater AHR level might be associated with plaque instability and rupture.
Given the role of AHR in mediating inflammation and atherosclerosis, we postulated that TCF21 may alter the risk of atherosclerosis by modulating the AHR pathway. We set out to characterize the intersection of these pathways at the genomic level, identify the possible mechanisms of interaction, and to determine the role of TCF21-AHR interactions in the context of inflammation in the vessel wall. Through these studies, we define the molecular mechanisms by which these two transcriptional pathways interact to regulate the risk of atherosclerosis.
We have previously reported the analysis of an in vitro siTCF21 knockdown RNA-seq study in human coronary artery smooth muscle cells (HCASMC) and noted differential expression of a number of inflammatory genes and pathways.[10] Interestingly, this module included the gene encoding AHR which directs an inflammatory program as part of its repertoire of response to xenobiotics[13, 14, 26], and the xenobiotic pathway was identified as one of those differentially regulated with TCF21 modulation.[10] Also, ChIP-seq studies in HCASMC have shown TCF21 binding in the AHR locus, suggesting that this gene is regulated in part by TCF21, and raising our interest in possible interactions between these transcriptional networks.[12]
While AHR has not been associated with CAD risk using the statistical criteria employed for GWAS efforts, we have identified a variant within the AHR locus (rs608646) that has a nominal association (p = 0.0047) in the CARDIOGRAM+C4D GWAS data in the context of a single SNP study (S1A Fig). This SNP was also noted to regulate expression of AHR as identified with eQTL studies in GTEx tissues high in SMC content (S1B Fig). In aortic and coronary artery tissues, the AHR locus (+/- 1Mb) was generally enriched with AHR-eQTL signals with multiple peaks uniformly scattered throughout the locus (S2 Fig), indicating that AHR gene expression is genetically regulated in the vasculature and emphasizing the importance of AHR in these tissues. These data thus validate a previous candidate gene study that found association of CAD with AHR in East Asians.[26]
To further investigate overlap of these pathways, we sought to expand the repertoire of TCF21 regulated genes by performing a transcriptome analysis with lentivirus mediated TCF21 over-expression in HCASMC. Analysis of the top 500 differentially regulated genes with goseq[28] (Fig 1A, S1 Table) identified a significant number of terms related to embryonic development (lung morphogenesis, lung vasculature development), but also numerous terms related to innate immunity and inflammation (response to bacterial lipopeptide, response to lipoteichoic acid, CCR chemokine receptor binding, chemokine receptor binding, lymphocyte chemotaxis, chronic inflammatory response), most of which were upregulated with TCF21 overexpression. Employing the DAVID algorithm we found that downregulated genes were primarily associated with SMC development and phenotype (regulation of blood vessel size, contractile fiber/myofibril) while upregulated genes were primarily associated with cellular proliferation (mitosis/cell cycle, positive regulation of DNA replication) (Fig 1B, S2 Table). In addition, GO terms enrichment and PCA analysis performed using goseq yielded multiple immune system and atherosclerosis- related GO terms, implicating TCF21 in HCASMC to promote immune and pro-atherosclerotic-responses (S3 Fig).
To look for relationships between AHR and TCF21 transcriptional networks, we investigated correlations among genes that reside in the co-expression modules of both TCF21 and AHR, using publicly available microarray data sets. We created genome-wide gene expression modules using 4164 human microarray data sets, and used the GeneFriends algorithm that reports the top 5% of co-expressed genes as high order associations, as well as second order indirect, associations.[29] In this analysis, TCF21 and AHR shared a number of indirect associations that link the two modules (Fig 1C). Gene ontology analysis of associations with p<0.05 between TCF21 and AHR networks revealed strong enrichment for inflammation, extracellular matrix modification, and developmental terms (S3 Table). TCF21 and AHR appeared to be highly related when co-expression network was visualized with all other CAD GWAS implicated genes, localizing in a cluster of extracellular matrix gene COL4A1 and growth factor receptor PDGFR, and distinct from a cluster of lipid genes (LPA, APOA5, APOB, APOA1) (Fig 1D). Further, we identified the co-expression module for ARNT, the heterodimer partner of AHR, and found that it also contains genes indirectly connected to AHR and TCF21 modules (S4 Fig), suggesting functional connectivity between AHR-ARNT and TCF21 co-expressed genes (S3 Table).
Experiments were first conducted to determine whether TCF21 modulates expression of AHR and ARNT. RNA-seq and qPCR analysis in HCASMC showed that the AHR and ARNT mRNA levels were down-regulated by TCF21 siRNA knockdown (AHR 1.0±0.07 vs. 0.58±0.02, p = 0.0037; and ARNT 1.0±0.04 vs. 0.63±0.04, p = 0.0031) (Fig 2A and 2B), and up-regulated by TCF21 overexpression (S5 Fig). To further characterize the intersection of TCF21 and AHR transcriptional networks, we investigated the mechanism by which TCF21 regulates downstream genes in the AHR pathway, and chose to first study the dioxin effect on the canonical AHR target gene CYP1A1. Dioxin induction of CYP1A1 mRNA levels nearly doubled in HCASMC exposed to both dioxin and TCF21 transfection compared to dioxin alone (153.5±9.7 vs. 61.8±5.9 fold, P = 0.001), and the opposite result was seen with TCF21 knock-down in conjunction with dioxin (99.5±29.8 vs. 247.9±64.8, P<0.05) (Fig 2C and 2D). Manipulation of TCF21 expression alone did not alter CYP1A1 expression, suggesting that it does not directly affect transcription of this canonical AHR downstream gene, but does alter the response to dioxin most likely through regulation of AHR and ARNT expression levels.
To investigate the possible interaction of TCF21 and AHR in the regulation of target inflammatory pathway genes, additional studies were conducted in HCASMC. We focused on IL1A and MMP1 genes, as previous studies have found these genes to be representative targets of AHR activation through direct or indirect pathways.[30–32] mRNA levels were measured for IL1A, and MMP1 genes by RT-PCR in HCASMC with TCF21 expression perturbed by knockdown and over-expression. Knockdown of TCF21 decreased IL1A expression compared to cells treated with scrambled siRNA (0.91±0.04 vs. 0.52±0.06, p = 0.013) (Fig 2E). Dioxin treatment significantly increased expression of IL1A (0.91±0.04 vs. 1.42±0.09, p = 0.007), and co-treatment of siTCF21 blocked this effect (1.42±0.09 vs 0.65±0.04, p = 0.001). Similar results were obtained for MMP1, with siTCF21 alone decreasing gene expression (1.02±0.06 vs. 0.65±0.05, p = 0.009), dioxin increasing expression (1.02±0.06 vs. 1.62±0.28, p = 0.10), and siTCF21 knocking down the increased expression of MMP1 seen with dioxin (1.62±0.28 vs. 0.64±0.12, p = 0.034) (Fig 2F).
To begin to investigate the mechanism by which TCF21 regulates AHR expression, we correlated whole genome RNA-seq and genotype information developed in 52 HCASMC lines to evaluate expression quantitative trait locus (eQTL) effects at the AHR locus.[33] We identified SNP rs10265174 to be one of the top eQTLs for AHR (p<9e-5) (Fig 3A and 3B). Also, rs10265174 was consistently found to regulate gene expression in multiple GTEx tissues, including coronary artery, aorta and tibial artery (S4 Table). Furthermore, the SNP was located in an open chromatin region/enhancer marked by ATAC-seq, H3K27ac ChIP-Seq, JUN and JUND ChIP-Seq peaks, and within a TCF21 ChIP-Seq peak. We found the rs10265174 variant to alter the PMW scores for AP1 and TCF4 transcription factors (HaploReg) (Fig 3C). Given that TCF4 is a known bHLH binding partner for TCF21,[34] we evaluated whether TCF21 might directly regulate AHR gene expression at this site. We surveyed ChIP-seq data previously generated for TCF21 in HCASMC[12] and identified ChIP-seq peaks representing TCF21 binding sites in both the AHR and ARNT genes. We confirmed the binding of TCF21 to both genomic regions in HCASMC with ChIP-qPCR for AHR (1.02±0.29 vs. 3.58±0.75; p = 0.033) and ARNT (1.03±0.28 vs. 20.04±4.80; p = 0.017) loci (Fig 3D, S6 Fig). Taken together, these data suggest that TCF21 may directly regulate expression of both the AHR and ARNT genes at the transcriptional level.
To further investigate the overlap of TCF21 and AHR transcriptional networks at the genomic level, we determined the genome-wide relationship between binding sites for AHR-ARNT and TCF21. We scanned the human genome sequence with the position weight matrix (PWM) for AHR-ARNT, and scanned for the PWM for TCF12 as a surrogate for TCF21 (JASPAR matrices, Ahr::Arnt—MA0006.1; Tcf12—MA0521.1).[12, 35, 36] TCF12 is the primary heterodimerization partner for TCF21, binds the same primary sequence as TCF21, and this composite site is indicated here as TCF12/TCF21.[12, 36] Predicted binding sites for TCF12/TCF21 and AHR-ARNT identified co-localization within a broad region of 5kb, with 339 high stringency TCF12/TCF21 and AHR-ARNT sites that directly overlap (P<2.2e-16, Fisher exact test, using combined ENCODE open chromatin regions as background; 218 sites overlapping within the background) (Fig 4A and 4B) and 11769 lower stringency sites overlapping (P<2.2e-16, Fisher exact test, ENCODE background; 4833 sites overlapping within the background; S5 Table). Next, we tested the positional orientation of TCF12/TCF21 and AHR-ARNT sites near functional elements, such as promoters, using the collection of 100,276 human ENSEMBL transcription start sites (TSS), Hum_ENSEMBL69 from Biomart. We observed that both matrices show double peaks near the oriented ENSEMBL TSS (Fig 4C). We also noted that these double peaks are in phase with each other, suggesting conservation of spatial orientation between the two sets of predicted binding sites and possible functional interaction of the two proteins that is preserved by evolutionary constraint near the TSS. In addition, we observe that the distance between phased peaks corresponds to the position of the +1 nucleosome, with an additional peak corresponding to the +2 nucleosome in the TCF PWM profile, suggesting that functional interaction between TCF21 and AHR-ARNT would be localized at the boundaries defined by nucleosome positioning at functional regions.
To test whether TCF21 in vivo binding sites correlate with AHR-ARNT PWM predictions, we generated the precise locations of TCF21 ChIP-seq summit positions in HCASMC using the MACS ChIP-seq tool (S6 Table).[37] We found that the center of these TCF21 ChIP-seq summits co-localized with the predicted AHR-ARNT PWM sites, further suggesting that AHR-ARNT complexes co-localize genome-wide with TCF21 in vivo binding sites (Fig 4D). In contrast, control PWMs for kidney/liver specific factors HNF1A and HNF1B showed uniform background distribution near TCF21 summits (Fig 4E and S7 Fig). In addition, AHR-ARNT PWM profiles showed an increase at summits for open chromatin regions in HCASMC, defined with MACS and ATAC-seq HCASMC data sets (Fig 4F). Control HNF1A and HNF1B matrices showed a decrease in their frequency near ATAC-seq summits (Fig 4G and S7 Fig).
We further assessed the steric relationship of TCF21 and AHR binding by dividing the co-localized TCF and AHR predicted sites into two categories, rotationally phased and un-phased, as rotational phasing has been shown to be crucial for direct protein binding.[38, 39] We considered PWM sites to be phased if they occurred at distances of n(10bp), i.e. 10, 20, 30, and 40bp, in which case due to the DNA helical pitch they will be oriented on the same side of the DNA strand and capable of direct protein-protein interaction. If separated by distances of 5, 15, 25, 35, 45 bp they would be expected to be oriented on the opposite sides of the DNA molecule due to the pitch of the DNA major groove, rotationally un-phased and less capable of direct protein-protein interaction. In the case of un-phased sites, indirect interaction through a protein complex might still be possible, e.g. through intermediate protein interactions. We extracted genes that are proximal to both categories of sites and calculated GO enrichment using GREAT (Fig 5, S7 Table).[40] The un-phased sites were enriched in cellular differentiation categories such as: negative regulation of cell fate commitment, but importantly a number of terms were related to inflammatory response (regulation of cytokine production, regulation of interleukin-6 production, regulation of TNF production). GO terms for phased sites were enriched in cellular and developmental terms (skeletal system morphogenesis, response to organophosphorous, regulation of cell migration, cell-matrix adhesion, and osteoblast development) In addition, binomial fold changes for un-phased AHR-TCF sites were ~10 times higher than for phased sites, implicating the indirect interaction of AHR-TCF21 factors as predominant in the human genome.
To further evaluate whether TCF21 and AHR-ARNT complex binding co-localizes in the human genome, we compared ChIP-seq data for these three TFs. We identified TCF21 ChIP-seq peaks in HCASMC that overlapped with AHR and ARNT ChIP-seq sites identified in MCF-7 cells[41], and obtained a statistically significant co-localization of sites using the Fisher’s exact test. Overlap of TCF21 and AHR peaks produced odds ratios within confidence intervals CI: 4.34–5.4 (p = 1.88e-121), for ARNT and TCF21, CI: 4.16–5.77 (p = 1.7e-56) and for AHR, ANRT and TCF21, CI: 4.91–7.05 (p = 1.13e-56) (S8 Table). We obtained in total 322 (12.4%) genomic locations that were co-occupied by AHR and TCF21, out of which 119 sites were also occupied by ARNT. Similarly, in 143 genomic locations (10.5%) ARNT co-localized with TCF21, out of which 119 were occupied with its binding partner AHR (Fig 6A). Overlap of AHR and ARNT identified 890 sites (AHR, 34.3%; ARNT, 65.8% percent of total sites), consistent with the fact the two proteins are known binding partners. Subsequently, we selected genes that are proximal to the overlapping TCF21-AHR, TCF21-ARNT and TCF21-AHR-ARNT ChIP-seq sites, and performed GO enrichment analysis with GREAT (Fig 6B, S8 Table). TCF21 and AHR-ARNT overlapping sites classified into GO-terms related to chemokine and cytokine signaling (positive regulation of cytosolic calcium ion concentration, regulation of cytosolic calcium ion concentration), apoptosis (regulation of apoptotic process, regulation of programmed cell death), metabolic processes (cellular hormone metabolic process, isoprenoid metabolic process), and cellular signaling (cellular response to stimulus).
Next, we assessed the binding of AHR, ARNT and TCF21 near lead SNPs from the GWAS Catalog (version 2016-05-08), expanded by addition of CARDIoGRAM+C4D meta-analysis data[2, 42], using the binomial test for genomic overlap. We scanned the lead SNPs using windows of +/-2kb, +/-5kb, and +/-10kb near the ChIP-seq binding sites and calculated the significance using the gwasanalytics tool (Fig 6C–6F, S8 Fig). Using the +/- 2kb window, TCF21 binding shows general enrichment near a wide range of GWAS SNPs for cardio-metabolic phenotypes (coronary heart disease, blood pressure and type 2 diabetes) as well as chronic inflammatory diseases (Crohns disease, multiple sclerosis, and rheumatoid arthritis), as well as skeletal phenotypes (bone mineral density) and in certain neurological disorders (bipolar disorder and schizophrenia). ARNT binding was localized near GWAS SNPs for chronic inflammatory diseases (lupus erythematosus and ulcerative colitis) and prostate cancer GWAS SNPs. AHR binding co-localized with CAD variants (coronary artery disease, coronary artery calcification) as well as chronic inflammatory GWAS SNPs (e.g., Crohn’s disease). After intersection of AHR with ARNT and TCF21 the only remaining categories were coronary artery disease and coronary artery calcification, narrowing the importance of the interaction of AHR/ARNT and TCF21 factors to pathophysiological processes in cardiovascular disease. Furthermore, we surveyed the overlap of CARDIoGRAM+C4D GWAS SNPs and AHR-ARNT PWM to consider the potential role of AHR in other CAD associated genes. In total 456 ARNT-AHR sites overlapped with CARDIOGRAM+C4D SNPs (lead plus LD r2>0.8), comprising 0.27 permil (low stringency) and 0.38 permil (high stringency) of total ARNT-AHR PWMs. In comparison, there were only 7 and 5 HNF1A and HNF1B sites, comprising 0.12/0.10 permil of total HNF1A/B sites (p<0.005, comparison of AHR-ARNT and HNF using Z-score test for proportions, S9 Fig).
Given these data showing that TCF21 and AHR binding sites are co-localized in the genome (Figs 4–6), and that TCF21 expression levels directly modulate the AHR response to dioxin (Fig 2), we investigated the functional interaction of these transcription factors at target loci. First, we surveyed the genomic region of CYP1A1 for TCF21 in vivo binding in HCASMC. A TCF21 ChIP-seq binding peak was identified and localized to a region of open chromatin, as defined by ATAC-seq data in HCASMC[10], and binding was confirmed with ChIP-qPCR (IgG 1.0±0.18 vs. TCF21 4.71±0.24, p = 0.001) (Fig 7A and 7B). This peak co-localized in the same region of open chromatin with several predicted AHR-ARNT binding sites, thus suggesting coordinated regulation of CYP1A1 expression. To confirm that the regulation of CYP1A1 mRNA levels by AHR and TCF21 is mediated at the transcriptional level through the observed ChIP identified binding sites, and to look for evidence of cooperativity at this level, we conducted reporter gene transfection studies. A 50 bp sequence containing alternating binding motifs for TCF21 and AHR-ARNT binding identified in the CYP1A1 gene was cloned into a luciferase expression plasmid with a minimal promoter sequence. Dual luciferase-renilla assays revealed enhancer activity when exposed to TCF21 overexpression (1.0±0.06 vs. 3.15±0.02, p = 0.0024), and TCF21 overexpression further increased the luciferase expression induced by dioxin in these cells (3.62±0.30 vs. 6.17±1.31 fold, p = 0.0005) (Fig 7C). The combined effect was additive with no evidence of synergism that would be suggestive of cooperative binding. Further, when the TCF21 binding motifs were removed from the reporter construct, TCF21 overexpression failed to further increase the expression of luciferase, suggesting that the transcriptional effect of TCF21 is specific for protein-DNA binding (Fig 7D). These data indicate that the regulatory effect of TCF21 on AHR target genes can be mediated by direct interaction in these target loci, and requires protein-DNA binding.
We followed up our previous studies showing regulation of inflammatory mediators by AHR and TCF21 with studies investigating possible endogenous mediators of AHR activation. As shown previously, application of dioxin to HCASMC resulted in up-regulation of IL1A, and this effect was reversed when cells were treated with the AHR antagonist alpha-napthoflavone (α-NF) (Fig 8A). In the same experiments, we tested oxidized-LDL (ox-LDL) as a potential endogenous activator of AHR in HCASMC.[43, 44] The treatment with ox-LDL resulted in the activation of genes that was similarly reduced with α-NF co-treatment, suggesting that the SMC response to ox-LDL is at least partly mediated by the AHR pathway (Fig 8A). We also found activation of a dioxin response element with oxLDL in luciferase assays (S10 Fig).
Given these data suggesting that AHR targets overlap the TCF21 CAD associated transcriptional network genes, we sought to substantiate the relevance of AHR in vascular disease through expression studies in mouse and human vascular tissues[8–10] For in vivo gene expression in mice, we performed microarray analysis of carotid arteries subjected to plaque rupture induced by partial ligation in ApoE-/- animals to compare gene expression between ruptured and non-ruptured plaques.[45] We found Ahr expression to be higher in the ruptured plaques compared to non-ruptured plaques (8.60±0.20 vs. 9.58±0.16, FDR q = 0.054) (Fig 8B). Furthermore, laser capture microdissection (LCM) was performed in atherosclerotic lesions in the aortic sinus of ApoE-/- mice exposed to 12 weeks of high fat diet. We found the expression level of Ahr to be significantly higher in the intimal plaque when compared to the expression in the adventitia, localizing the expression of AHR to the pathologic intimal thickening (1.0±0.2 vs. 12.4±1.2, p = 0.0008) (Fig 8C). Next, we validated these findings in human arteries ex vivo, using microarray based expression data from normal arteries and atherosclerotic human carotid lesions from the BiKE repository.[46] Expression levels of AHR along with IL1A, and MMP1 were significantly higher in the diseased lesions (Fig 8D, AHR 8.71±0.16 in normal vs. 9.27±0.05 in plaques, p = 0.0065; S11 Fig). Furthermore, we analyzed the proteins present in human carotid plaques using liquid chromatography tandem mass-spectrometry (LC-MS/MS). Proteomic datasets were constructed from highly phenotyped patients with asymptomatic and symptomatic carotid stenoses, 10 subjects each matched for gender, statin usage and age, with plaques selected on CT and histology criteria. AHR-TCF21 unique interactors from BIOGRID protein-protein interaction database were used to display clustering patterns (Fig 8E). AHR is located in a cluster of genes that include immune related genes such as IRAK4, interleukin-1 receptor-associated kinase 4, transcription factors like SP1, and cell cycle regulated genes including XPO1. In addition, we selected ChIP-seq co-occupied genes for AHR-TCF21 and AHR-ARNT-TCF21 transcription factors and observed clustering of AHR target protein CYP1B1 with extracellular matrix factors FN1, COL18A1 and with growth factor receptor IGF1R, implicating AHR and its downstream targets in regulation of extracellular matrix component of the diseased human carotid artery plaque (S12 Fig).
We have identified TCF21 as the causal gene at 6q23.2, characterized its mechanism of association, and shown that binding of this transcription factor is enriched in other CAD associated loci.[8, 12] To investigate how TCF21 interaction with other CAD loci may regulate disease risk, we have begun to study mechanisms of association in these loci. For initial studies we have chosen the AHR gene, because it encodes a transcription factor, allowing direct study of its downstream signaling pathway, and because of the well-characterized link between this factor and environmental exposures that are relevant for cardiovascular disease. This work thus addresses two aspects of CAD that have not been directly approachable with human association studies, investigating both gene-by-gene and gene-by-environment contributions to disease genetic risk.
Although variants in the AHR locus (rs608646) have shown only nominal association with CAD risk in GWAS meta-analyses, this may be due to the technical limitation of the GWAS methodology in the AHR locus or inadequate statistical power. It remains possible if not likely that AHR functions as a hub or master regulator in CAD without harboring regulatory disease variants. We did identify a variant within the AHR locus (rs608646) that has a moderate association (p = 0.0047) in the CARDIOGRAM+C4D GWAS data (S1A Fig), and this SNP was also noted to regulate expression of AHR as identified with eQTL studies in GTEx tissues high in SMC content (S1B Fig). In aortic and coronary artery tissues, the AHR locus was enriched with AHR-eQTL signals with multiple peaks across the genomic region (Fig 3B, and S2 Fig), indicating that AHR gene expression is genetically regulated in the vasculature and emphasizing the relevance of AHR expression in these tissues. Further, we also found genome-wide enrichment of the PWM for AHR-ARNT within CARDIOGRAM+C4D GWAS loci (S9 Fig), suggesting that the effect of AHR on CAD may be partly via genetic variation in protein-DNA interaction near genes related to CAD. These data support the candidate gene study which found an association of CAD with the AHR locus in East Asians.[26]
In our studies, we have pursued numerous approaches to investigate links between these two genes and their related transcriptional networks, and to investigate mechanisms by which they may work together to modulate CAD risk. First, we have shown with targeted studies that TCF21 binds both the AHR and ARNT loci, and increases expression levels of these genes in HCASMC, confirming previously published genomic studies and RNA-seq studies reported here. Second, these studies provide evidence for overlap of the TCF21 and AHR transcriptional networks. Both TCF21 and dioxin were shown to increase expression of disease-related factors such as IL1A, MMP1 and interestingly knockdown of TCF21 was able to almost completely abolish the effect of dioxin, suggesting that the inflammatory activation by AHR is dependent on the presence of TCF21. AHR is well known to promote inflammation in a number of situations, and to work with NFkB in this regard.[17] Also, we have previously shown that TCF21 can promote expression of a number of inflammatory genes and we show that this pro-inflammatory program represents an intersection of TCF21 with AHR function, identifying a subset of TCF21 target genes that could create a highly inflammatory cellular profile that would be significantly magnified with relevant environmental exposures. We also found that oxidized LDL activated the AHR pathway in HCASMC, consistent with previous reports in other cell types.[43, 47]
Further analyses investigated additional mechanisms of interaction between these two pathways. Using PWMs for both TCF21 and AHR, we found highly significant enrichment for co-localization in regions of open chromatin in HCASMC, and characterized similar organization of these binding sites around transcription factor start sites, suggesting functional interaction between TCF21 and AHR and the basal transcriptional apparatus, as proposed previously for other TFs.[48, 49] The genomic co-localization was further refined by intersecting summit locations from TCF21 ChIP-seq data with AHR-ARNT PWM positions. Support for these observations reflecting in vivo associations was provided by co-localization of ChIP-seq peaks for TCF21, AHR, and ARNT. These data suggest a role for AHR-ARNT in the functional regulation of coronary SMC phenotype.
Co-localization of TF binding often suggests direct functional interaction, and since TCF21 and AHR may regulate transcription in the same direction, an obvious hypothesis is that they bind cooperatively either through direct protein-protein interaction or through joint recruitment of ancillary adaptor proteins.[50] In addition to the genomic co-localization data, the absence of IL1A response to dioxin in TCF21 knockdown, and our studies investigating the phasing of binding site placement also suggests some form of direct or indirect molecular interaction. The striking difference between functional annotations for the two categories of steric relationship are consistent with different functional interactions between AHR–ARNT and TCF21 in the context of DNA binding. GO terms for un-phased sites showed much stronger enrichment and significance compared to those of the phased sites, supporting indirect interaction as the likely functional mechanism. We investigated this possibility using the CYP1A1 gene as model locus where both transcription factors bind. The reporter gene transfection studies with constructs containing both TCF21 and AHR binding sites showed an additive effect suggesting that the transcriptional effects are due to each TF acting independently. However, both TCF21 and AHR are known to bind AR[51, 52], and also potentially bind to Rb1 (S13 Fig). It remains a possibility that they interact through other TFs that function as intermediaries and are either not expressed or not active in the HCASMC. In fact, we were not able to demonstrate direct protein-protein interaction using a co-immunoprecipitation assay. Future studies using chemical crosslinking with ChIP and serial ChIP may help resolve these important questions.
The functional role of TCF21 in vascular disease appears closely related to its role in embryonic coronary artery vascular development where it is expressed in SMC precursor cells, supporting proliferation and migration of these cells. AHR is also expressed in the developing coronary circulation, and the application of the AHR-ligand dioxin in zebrafish inhibited epicardial and proepicardial development.[53, 54] Both TCF21 and AHR are downstream of retinoic acid signaling pathways that are critical in coronary artery development.[55–58] The possible molecular interaction of TCF21 and AHR in this setting has not been established.
In summary, we describe a novel functional interaction between two bHLH class transcription factors and postulate association of their interaction to the development of atherosclerosis and coronary artery disease (Fig 9). The discovery of a connection between TCF21, one of the most highly replicated GWAS candidate genes for coronary artery disease, and AHR, a gene classically involved with response to environmental toxins raises an interesting hypothesis that this interaction may reflect gene-environment interactions that are contributing to CAD and presents an opportunity to define causal gene-gene and gene-environment interactions relevant to the atherosclerotic lesion. Furthermore, our findings that TCF21 and AHR are expressed in the atherosclerotic plaque and that they interact to modulate inflammatory genes and matrix modifying genes suggest that the interaction may directly promote plaque instability leading to myocardial infarction. This work serves to promote mechanistic studies as an approach to understanding gene-by-gene and gene-by-environment contributions to disease genetic risk. Also, once the exact nature of the interaction of the two proteins is fully elucidated, therapeutic targeting of the pro-inflammatory interaction of TCF21 and AHR might be possible to reduce the effects of the pro-atherogenic stimuli in the vasculature and consequently reduce disease risk.
Primary human coronary artery smooth muscle cells (HCASMC) were purchased from three different manufacturers, Lonza, PromoCell and Cell Applications and were cultured in complete smooth muscle basal media (Lonza, #CC-3182) according to the manufacturer's instructions. All experiments were performed with HCASMC between passages 5–8. HEK293 cells were maintained in DMEM containing high glucose, sodium pyruvate and L-glutamine supplemented with 10% FBS.
For the TCF21 overexpression study, HCASMC were transduced with 2nd generation lentivirus with TCF21 cDNA cloned into pWPI (Addgene #12254). Briefly, for lentiviral transduction, the cells were treated at 60% confluence with MOI of 5 for 24 hours. The virus was removed and replaced with low-serum media for 48 hours prior to collection for downstream applications. For the siRNA transfection, cells were grown to 60% confluence, then treated with 10nM siRNA or scramble control with RNAiMax (Invitrogen, Carlsbad, CA) for 12 hours. The cells were collected 48 hours after transduction and processed using RNeasy kit (Qiagen, Hilden, Germany) for RNA isolation. TCF21 knockdown was performed with siRNA oligos (Origene, Rockville, MD) using Lipofectamine RNAiMAX (Thermo Fisher Scientific, Waltham, MA) following manufacturer’s protocol.
RNA was isolated using RNeasy mini kit (Qiagen) and total cDNA was prepared using iScript cDNA synthesis kit (Biorad, Hercules, CA). Gene expression levels were measured using SYBR Green assays with custom designed probes (S9 Table) and quantified on a ViiA7 Real-Time PCR system (Applied Biosystems, Foster City, CA) and normalized to GAPDH levels. Two group comparisons were performed using student t-test, and three group comparison was performed with ANOVA.
HCASMC were cultured as described above and total RNA was purified from 5.0x105 cells using the Qiagen miRNeasy kit. RNA libraries were prepared using the Illumina TruSeq library kit as described by the manufacturer. RNA molecules were sequenced using Illumina HiSeq 2500. Reads contained in raw fastq files were mapped to hg19 using the RNA-seq aligner STAR (v2.4.0i), that processes data with short run times and yields high numbers of uniquely mapped reads (https://github.com/alexdobin/STAR). Second pass mapping with STAR was then performed using a new index that is created with splice junction information contained in the file SJ.out.tab from the first pass STAR mapping. Read that have been mapped with STAR second pass mapping algorithm were subsequently counted using the htseq-count script distributed with the HTSeq Python package (https://pypi.python.org/pypi/HTSeq). Differential expression of exons, genes, and transcripts were assayed using the DESeq2 R package from Bioconductor (http://bioconductor.org/packages/release/bioc/html/DESeq2.html), which uses negative binomial distribution to estimate dispersion and model differential expression such as to permit biological variability to be different among tested genes (transcripts). GO terms enrichment and PCA analysis was performed using GOSeq and Gene Set Enrichment and Projection Displays–GSEPD Bioconductor package.
Fifty-two human coronary artery smooth muscle cell lines are genotyped using 30X whole-genome sequencing. Genotype calling follows the GATK best practices recommendations. Briefly, after removing adapter with cutadapt, trimmed FASTQ files were aligned with BWA mem, duplicates were marked with Picard tools. After indel realignment and variant base quality recalibration, single-nucleotide variants and short insertion and deletion variants are jointly called on all samples using the GATK Haplotype caller. Called variants are recalibrated and filtered using GATK's variant quality score recalibration module. We used BEAGLE 4.1 to impute and phase recalibrated variants using 1000 Genome phase 3 version 5a as a reference panel. After imputation and phasing, we filtered variants based on MAF > 0.05, Hardy-Weinberg equilibrium p-value > 1e-6, indel length < 51 bps, dosage r2 > 0.8. Gene expression was quantified using mRNA sequencing to an average depth of 50M 75-bp paired-ended reads. Sequences were aligned using STAR two-pass mapping. To avoid allele-specific mapping bias, we removed potentially mismapped reads using WASP. Read counts and FPKM values were generated using RNAseQC. Expression eQTL were mapped with RASQUAL. To remove potential confounders, we included gender, first 3 principal components inferred on the genotypes and first 8 PEER factors inferred on 10,000 highest expressed genes. Transcription factor binding and epigenetic annotations of variants were assayed by Haploreg v4.1.
A detailed protocol was included in a previous publication [12]. HCASMC were cultured as described above. Antibodies used for ChIP-qPCR were all pre-validated according to ChIP-seq guidelines and ENCODE best practices. Purified rabbit polyclonal antibody against human TCF21 (HPA013189) was purchased from Sigma. Briefly, ChIP-qPCR confirmation was performed using primers designed for the genomic region of AHR, ARNT and CYP1A1, and compared against ChIP performed with IgG antibody. (S2 Table). qPCR values for AHR, ANRT, and CYP1A1 promoter, normalized relative to the Myogenin (MYOG) signal, used as a endogenous control, were expressed as fold change compared to IgG ChIP sample. Comparisons were performed using student t-test.
TCF21 ChIP-Seq raw data from Sazonova et al., were reanalyzed using Model-based Analysis for ChIP-Seq (MACS v1.4.2) pipeline. Parameters were set to default. Summit locations of the peaks were defined for genome wide correlations with PWM using ChIPCor module—part of ChIP-Seq Analysis Server of the Swiss Institute of Bioinformatics (ccg.vital-it.ch/chipseq/chip_cor.php). TCF21 ChIP-Seq sites were converted to bigwig files and visualized on UCSC Genome Browser.
TCF21 overexpression was achieved using cDNA expression construct driven by a CMV promoter transduced by lentivirus. TCF21 knockdown was performed with siRNA oligos (Origene) following manufacturer’s protocol. Briefly, for lentiviral transduction, the cells were treated at 60% confluence with MOI of 5 for 24 hours. The virus was removed and replaced with low-serum media for 48 hours prior to collection for downstream applications. For the siRNA transfection, cells were grown to 60% confluence, then treated with 10nM siRNA or scramble control with RNAiMax (Invitrogen) for 12 hours. The transduced cells were then treated with TCDD (Sigma Aldrich Cat#48599) at a concentration of 10nM for 24 hours. For the dual luciferase assay, double stranded DNA sequences containing the TCF21 and AHR binding motifs were subcloned into the multiple cloning site (MCS) of the pLuc-MCS vector (Promega, #E1330), located upstream of the translation stop codon and firefly luciferase reporter gene luc2, driven by the PGK minimal promoter and also carrying the renilla luciferase reporter gene hRluc, as an internal control. Culture media was changed after 6 hrs, and dual luciferase activity was measured after 24 hrs using either SpectraMax L luminometer (Molecular Devices, Sunnyvale, CA). Relative luciferase activity (firefly/Renilla luciferase ratio) is represented as the fold change of respective control condition as indicated.
Oxidized-LDL was purchased from Alfa Aesar (Haverhill, MA; Cat No. J65591). Cells were treated at concentration of 10uM for 6 hours with and without α-NF at 10nM (Sigma Aldrich, St. Louis, MO; Cat No. N5757). The changes in downstream genes were confirmed using RT-qPCR.
12 week old ApoE-/- mice on C56BL/6J background were subjected to 4 weeks of partial ligation followed by 4 days of cuff placement as described previously.[45, 59] For the aortic sinus atherosclerosis model, ApoE -/- mice were put on 12 weeks of Western high fat diet (HFD, 21% anhydrous milk fat, 19% casein and 0.15% cholesterol, Dyets no. 101511) at 4 weeks of age.
Using a Leica LMD6000, we performed LCM of atherosclerotic plaques of mouse aortic sinuses. Briefly, following sacrifice, the cardiac chamber was perfused with PBS, then the aortic sinus was dissected and embedded in Optimal Cutting Temperature (OCT) medium (Tissue-Tek). 7um cryosections were placed on to Leica membrane slides, then visualized under the microscope for LCM. Total RNA was extracted using RNeasy Plus Micro kit (Qiagen), and the quality of RNA checked with Agilient Bioanalyzer RNA 6000 Pico kit. The levels of gene expressions were compared using total RNA generated from ApoE (-/-) mouse on high fat diet for 12 weeks. Using the published SMART-Seq2 protocol [60], we amplified the ultra-low input RNA from the LCM. Briefly, reverse transcription was performed using a template switching oligonucleotide (TSO) with locked nucleic acid (LNA) and Superscript II (Invitrogen, Carlsbad, CA), followed by PCR amplification with KAPA PCR polymerase.
Human atherosclerotic carotid artery lesions were obtained from patients undergoing endarterectomy surgery for carotid stenosis, as part of the Biobank of Karolinska Endarterectomies (BiKE).[46] Details of the cohort demographics, sampling at surgery, processing and microarray analyses have been described before. Briefly, normal control samples (n = 10) were iliac arteries and one aorta from healthy organ donors without any history of cardiovascular disease. Plaques were frozen at -80°C immediately after surgery, pulverized to a powder before resuspending in Qiazol lysis reagent (Qiagen) and homogenization with a tissue homogenizer. Total RNA was extracted as described above using the miRNeasy Mini Kit (Qiagen) and RNA quality assessed using a Bioanalyzer 2100 (Agilent). Global gene expression profiles were analyzed by Affymetrix HG-U133 plus 2.0 Genechip microarrays from 127 patient derived plaque samples and 10 donor control samples. Robust multi-array average (RMA) normalization was performed and processed gene expression data presented in Log2 scale.
Atherosclerotic plaques from 18 BiKE patients (matched for male gender, age and statin medication) were analysed using LC-MS/MS as previously described.[46, 61] Briefly, protein samples were digested by trypsin and the resulting tryptic peptides were TMT-labeled and pooled. Pooled samples were cleaned by Strong Cation exchange columns (Phenomenex) and subjected to LC-MS/MS analysis. The sample pools were separated on a 4 hour gradient using an UPLC-system (Dionex UltiMate™ 3000) coupled to a Q-Exactive mass spectrometer (Thermo Fischer Scientific, San Jose, CA, USA). The fragment spectra from the mass spectrometer were matched to a database consisting of theoretical fragment spectra from all human proteins and filtered at a 1% False Discovery Rate (FDR) on the peptide level to obtain protein identities (Uniprot). Quantitative information was acquired by using the TMT reporter ion intensities. Correlation matrices were constructed by calculation of the proteomic expression correlation coefficients using the Pearson method and p-values were corrected for multiple comparisons using Bonferroni. For the clustering plots, dissimilarity index was created using the method that best discriminates all correlated pairs, given the formula: Dissimilarity = 1 –Abs (Correlation). Distance matrix was then created from the dissimilarity index. Clustering was performed with heatmap.2 in gplots.
AHR, ARNT, TCF21, HNF1A, and HNF1B co-expression modules were obtained using GeneFriends using 4164 human microarray data sets or 4133 human RNA-Seq data sets. TCF21 and AHR co-expression modules were defined using 4164 human microarray datasets through GeneFriends and visualized with Cytoscape. TCF21, AHR and ARNT co-expression modules were defined using 4133 human RNA-seq datasets through GeneFriends and visualized with Cytoscape. GeneFriends associations are defined with the threshold of 5%, meaning the gene is associated to a specific gene if it is in the top 5% co-expressed genes for a that gene. Cytoscape network file was imported for visualization from GeneFriends, containing all gene-gene associations marked as "good friends" (top 10 friends with a connection strength of 1), or "lesser friends" (genes ranking between 10 and 20 with a rank of 0.5). If a gene is an indirect connection, i.e. friend of a friend, score of 0.25 is deduced from the connection strength. Core network of direct interactions is marked on a graph with different colors to distinguish direct from indirect interactions. Connecting co-expression modules and visualization was performed using Cytoscape. Clustering was performed using Edge-weighted Spring Embedded layout. Coronary artery disease (CAD) GWAS genes were defined using Cardiogram plus C4D meta-analyses GWAS loci. In total, 77 CAD GWAS genes were used for transcription module analysis in GeneFriends:Microarray to obtain gene expression modules which were subsequently clustered in Cytoskape. Genes TCF21, AHR, COL4A1, SMAD3 and PDFGD from the main cluster were colored and indicated, node size was increased and edges to their first neighbors were colored in red. Underlying edge connections were colored purple with increased transparency. Grouping of transcription modules into three main clusters shows that CAD GWAS genes act through three main regulatory networks with TCF21 and AHR gene modules appearing in the single cluster.
TCF12 and AHR-ARNT matrices (TCF12—MA0521.1 and AHR-ARNT—MA0006.1) were obtained from the JASPAR database (http://jaspar.genereg.net) [35]. TCF12 PWM was used as it has the same binding motif as TCF21. Human genome hg19 was scanned with the two JASPAR matrices using PWMScan—Genome-wide PWM scanner (http://ccg.vital-it.ch/pwmtools/pwmscan.php). Position weight matrix sites were counted in windows of various lengths surrounding centered features using the Feature correlation tool from the ChIPCor module from ChIP-Seq Analysis Server (ccg.vital-it.ch/chipseq/chip_cor.php).
AHR, ARNT and TCF21 ChIP-Seq binding site were extended to windows +/-1000bp, +/-2000bp, and +/-5000bp using bedtools package. Overlap of extended locations of AHR, ARNT and TCF21 ChIP-Seq binding sites and GWAS Catalog SNPs was performed with bed2GwasCatalogBinomialMod1Ggplot script from gwasanalytics package. This script is a modification of the bed2GwasCatalogBinomialGgplot and calculates binomial p-value for genomics overlaps using the following criteria. The P-values were computed using binomial cumulative distribution function b(x;n,p) in R (dbinom function). We set the parameter n equal to the total number of GWAS SNPs in a particular GWAS phenotype. Parameter x was set to the number of GWAS SNPs for a given GWAS phenotype that overlap input regions and parameter p was set to the fraction of the uniquely mappable human hg19 genome (calculated with subscript) that is localized in the input regions and contains assessed GWAS phenotype SNPs. Calculated binomial p-value equals the probability of having x or more of the n test genomic regions in the open chromatin domain given that the probability of that occurring for a single GWAS genomic location is p. Plots were made using ggplot2 package and the wes anderson color palette in R (https://github.com/karthik/wesanderson).
All experiments were performed by the investigators blinded to the treatments/conditions during the data collection and analysis, using at least two independent preparations and treatments/conditions in triplicate. R/Bioconductor or GraphPad Prism 6.0 was used for statistical analysis. For enrichment analyses, we used both Fisher’s exact test and the cumulative binomial distribution test, as indicated. For comparisons between two groups of equal sample size (and assuming equal variance), an unpaired two-tailed Student’s t-test was performed or in cases of unequal sample sizes or variance a Welch’s unequal variances t-test was performed, as indicated. P values <0.05 were considered statistically significant. For multiple comparison testing, one-way analysis of variance (ANOVA) accompanied by Tukey’s post hoc test were used as appropriate. All error bars represent standard error of the mean (SE).
The BiKE study is approved by the Ethical Committee of Northern Stockholm with following ethical permits: EPN DNr 95–276/277; DNr 02–146; DNr 02–147, DNr 2005/83-31; DNR 2009/512-31/2; DNR 2009/295-31/2; 2011/950-32; 2012/619-32 and 213/2137-32. The project is performed under the Swedish biobank regulations and prospective sampling is approved with informed consent procedure (DNr 2009/512-31/2). BiKE is registered at Socialstyrelsen (The National Board of Health and Welfare) and Biobank of Karolinska and approved by the Swedish Data Inspection Agency (approval date/number 2002-09-30 DNr 916–2002). All samples are collected with oral and written informed consent from patients or organ donor guardians.
All animal procedures described in this study were approved by the Institutional Animal Care and Use Committees of Stanford University and conformed to NIH guidelines for care and use of laboratory animals. Specifically, the animal studies were approved by APLAC protocol #10022, last approved on 3-16-17 and will remain in effect until 12-12-19.
|
10.1371/journal.pbio.0050087 | Independently Evolving Species in Asexual Bdelloid Rotifers | Asexuals are an important test case for theories of why species exist. If asexual clades displayed the same pattern of discrete variation as sexual clades, this would challenge the traditional view that sex is necessary for diversification into species. However, critical evidence has been lacking: all putative examples have involved organisms with recent or ongoing histories of recombination and have relied on visual interpretation of patterns of genetic and phenotypic variation rather than on formal tests of alternative evolutionary scenarios. Here we show that a classic asexual clade, the bdelloid rotifers, has diversified into distinct evolutionary species. Intensive sampling of the genus Rotaria reveals the presence of well-separated genetic clusters indicative of independent evolution. Moreover, combined genetic and morphological analyses reveal divergent selection in feeding morphology, indicative of niche divergence. Some of the morphologically coherent groups experiencing divergent selection contain several genetic clusters, in common with findings of cryptic species in sexual organisms. Our results show that the main causes of speciation in sexual organisms, population isolation and divergent selection, have the same qualitative effects in an asexual clade. The study also demonstrates how combined molecular and morphological analyses can shed new light on the evolutionary nature of species.
| The evolution of distinct species has often been considered a property solely of sexually reproducing organisms. In fact, however, there is little evidence as to whether asexual groups do or do not diversify into species. We show that a famous group of asexual animals, the bdelloid rotifers, has diversified into distinct species broadly equivalent to those found in sexual groups. We surveyed diversity within a single clade, the genus Rotaria, from a range of habitats worldwide, using DNA sequences and measurements of jaw morphology from scanning electron microscopy. New statistical methods for the combined analysis of morphology and DNA sequence data confirmed two fundamental properties of species, namely, independent evolution and ecological divergence by natural selection. The two properties did not always coincide to define unambiguous species groups, but this finding is common in sexual groups as well. The results show that sex is not a necessary condition for speciation. The methods offer the potential for increasing our understanding of the nature of species boundaries across a wide range of organisms.
| Species are fundamental units of biology, but there remains uncertainty on both the pattern and processes of species existence. Are species real evolutionary entities or convenient figments of taxonomists' imagination [1–3]? If they exist, what are the main processes causing organisms to diversify [1,4]? Despite considerable debate, surprisingly few studies have formally tested the evolutionary status of species [1,5,6].
One central question concerning the nature of species has been whether asexual organisms diversify into species [1]. The traditional view is that species in sexual clades arise mainly because interbreeding maintains cohesion within species, whereas reproductive isolation causes divergence between species [7]. If so, asexuals might not diversify into distinct species, because there is no interbreeding to maintain cohesive units above the level of the individual. However, if other processes were more important for maintaining cohesion and causing divergence, for example, specialization into distinct niches, then asexuals should diversify in a manner similar to sexuals, although the rate and magnitude of divergence might differ [8–11].
Empirical evidence to test these ideas has been rare. Most asexual animal and plant lineages are of recent origin [9,12]. The diffuse patterns of variation typical of such taxa [13] could simply reflect their failure to survive long enough for speciation to occur or the effects of ongoing gene flow from their sexual ancestors [9,12]. Distinct genetic and phenotypic clusters have been demonstrated in bacteria [14–17] and discussed as possible evidence for clonal speciation [1]. However, all the study clades engage in rare or even frequent recombination as well as clonal reproduction [14,18,19]. Although horizontal gene transfer can occur between distantly related bacteria, homologous recombination occurs only at appreciable frequency between closely related strains [20,21]. Therefore, clusters in these bacteria could arise from similar processes to interbreeding and reproductive isolation in sexual eukaryotes [20]. Aside from issues of sexuality, previous studies looking for distinct clusters have been descriptive, relying on visual interpretation of plots of genetic or phenotypic variation rather than on formal tests of predictions under null and alternative evolutionary scenarios [1].
Here, we demonstrate that a classic asexual clade, the bdelloid rotifers, has diversified into independently evolving and distinct entities arguably equivalent to species. Bdelloids are abundant animals in aquatic or occasionally wet terrestrial habitats and represent one of the best-supported clades of ancient asexuals [22–24]. They reproduce solely via parthenogenetic eggs, and no males or traces of meiosis have ever been observed. Molecular evidence that bdelloid genomes contain only divergent copies of nuclear genes present as two similar copies (alleles) in diploid sexual organisms rules out anything but extremely rare recombination [25–27]. Yet, bdelloids have survived for more than 100 million y and comprise more than 380 morphologically recognizable species and 20 genera [28]. The diversity of the strictly asexual bdelloids poses a challenge to the idea that sex is essential for long-term survival and diversification [29]. However, taxonomy does not constitute strong evidence for evolutionary species: the species could simply be arbitrary labels summarizing morphological variation among a swarm of clones [7]. We adopt a general evolutionary species concept, namely, that species are independently evolving and distinct entities, and then break the species problem into a series of testable hypotheses derived from population genetic predictions [3]. We use the word “entity” to refer to a set of individuals comprising a unit of diversity according to a given criterion or test: the question of whether to call those entities “species” will be returned to below.
Focusing on the genus Rotaria (Figure 1), one of the best-characterized genera of bdelloids, we use combined molecular and morphological analyses to distinguish alternative scenarios for bdelloid diversification (Figure 2). First, the entire clade might represent a single species, that is, a swarm of clones with no diversification into independently evolving subsets of individuals. Second, the clade may have diversified into a series of independently evolving entities. By “independently evolving,” we mean that the evolutionary processes of selection and drift operate separately in different entities [8,9], such that genotypes can only spread within a single entity. Possible causes of independence include geographical isolation or adaptation to different ecological niches [10,17]. The expected outcome is cohesion within entities but genetic and phenotypic divergence between them [9–11].
We first test for the presence of independently evolving entities. Under the null scenario of no diversification, genetic relationships should conform to those expected for a sample of individuals from a single asexual population (H0, Figure 2A). Under the alternative scenario that independently evolving entities are present, we expect to observe distinct clusters of closely related individuals separated by long branches from other such clusters (H1, Figure 2A; and [9]). Coalescent models can be used to distinguish the two scenarios [30]. Failure to reject the null model would indicate a lack of evidence for the existence of independently evolving entities.
Next, to investigate the role of adaptation to different niches in generating and maintaining diversity within the clade, we extend classic methods from population genetics to test directly for adaptive divergence of ecomorphological traits. If trait diversity evolves solely by neutral divergence in geographic isolation, we expect morphological variation within and between entities to be proportional to levels of neutral genetic variation (H0, Figure 2B, Materials and Methods). If, instead, different entities experience divergent selection on their morphology, we expect greater morphological variation between clusters than within them, relative to neutral expectations (H1, Figure 2B; and [31]). Past work has often discussed sympatry of clusters as evidence for niche divergence [1], but, in theory, coexistence can occur without niche differences [32]; hence, we introduce an alternative, more direct approach.
Our results demonstrate that bdelloids have diversified not only into distinct genetic clusters, indicative of independent evolution, but also into entities experiencing divergent selection on feeding morphology, indicative of niche divergence. In common with findings of cryptic species in sexual organisms [33,34], the morphologically coherent groups experiencing divergent selection often include several genetic clusters: this introduces difficulties in deciding which units to call species, but this problem is shared with sexual organisms [3,33]. In short, bdelloids have diversified into entities equivalent to sexual species in all respects except that individuals do not interbreed. The results demonstrate the benefits of statistical analyses of combined molecular and morphological data for exploring the evolutionary nature of species.
We collected all individuals of Rotaria encountered during 3 y searching rivers, standing water, dry mosses, and lichens, centered on Italy and the United Kingdom but also globally [35]. Individuals were identified to belong to nine taxonomic species (Tables S1 and S2). Most of the described species of Rotaria missing from our sample are known from only one record or are very rarely encountered (Protocol S1). Bayesian and maximum parsimony analyses of mitochondrial cytochrome oxidase I (cox1) and nuclear 28S ribosomal DNA sequences provide strong support for the monophyly of taxonomic species (Figures 3, S1, S2, and S3 and Text S1), with the sole exception of R. rotatoria, which was already suspected to comprise a species complex based on disagreements among authors [36,37].
Morphometric analyses further support the distinctness of taxonomic species. Bdelloid morphology is hard to measure because of their shape-changing abilities; hence, we used geometric morphometrics [38] to measure the only suitable trait, their hard jaws, called trophi [39] (Figures 1 and S4). Trophi size and shape are not characters that have been used in the traditional taxonomy of the genus (Table S2). Trophi scale weakly with rough measures of body size of each species (mean trophi size against log body length from [37]: r = 0.55, p = 0.2, Spearman's rank test), and both the size and shape of trophi likely reflect different types or sizes of particulate food consumed, although the details of how food is processed remain unclear [28]. Discriminant analysis of the first five principal components (PCs) describing trophi shape (cumulative explained variance, 97.1%; Materials and Methods) produced a correct classification with respect to traditional taxonomy of most specimens of R. macrura, R. neptunia, R. sordida, and R. tardigrada (Table S3). The remaining species overlapped in shape but could be discriminated by size (Figures 4 and S5). Related species on the DNA trees tend to have similar morphology: for example, R. magnacalcarata, R. socialis, and R. rotatoria FR.2.1 and IT.5 overlap in shape, but are more distant from R. rotatoria UK.2.2. Only two of the traditional species found to be monophyletic in the DNA tree displayed significant variation in size or shape among populations: R. sordida and R. tardigrada. In both cases, the populations that differed were deeply divergent in the DNA tree as well.
Congruence between molecules and morphology confirms that most traditional Rotaria species are monophyletic clades but does not rule out the possibility that taxa reflect variation within a single asexual species or swarm of evolutionarily interacting clones. Under the alternative scenario that independently evolving entities are present, we expect to observe clusters of closely related individuals separated from other such clusters by longer internal branches on a DNA tree [9,30,40]. We therefore tested for significant clustering by comparing two models describing the likelihood of the branching pattern of the DNA trees: first, a null model that the entire sample derives from a single population following a neutral coalescent [41], and, second, a model assuming a set of independently evolving populations joined by branching that reflects the timing of divergence events between them, that is, cladogenesis [9,30,42]. The models allow departures from strict assumptions of constant population size and rates of cladogenesis (see Materials and Methods).
The results indicate significant clustering within Rotaria, as expected if several independently evolving entities are present and consistent with patterns of mtDNA diversity from a broad sample of bdelloids [43]. The maximum likelihood solution for the independent evolution model on the combined tree infers 13 isolated clusters, with the remaining individuals inferred to be singletons (Figure 3; Table S4). Two monophyletic taxonomic species contained two separate clusters: R. magnacalcarata has two clusters corresponding to the U.K. and Italian samples, whereas R. macrura has two clusters not matching sampling locality. Uncorrected pairwise distances of cox1 within clusters ranged from 0% to 3.3% (mean, 1.5%), and those between clusters ranged from 4.1% to 23.1% (mean, 16.0%). The null model that the entire lineage represents a single cluster can be rejected (log likelihood ratio test, 2 × ratio = 30.8, χ2 test, three degrees of freedom, p < 0.0001).
Our results indicate that independently evolving entities are present in bdelloids but at a lower level than taxonomic species, that is, cryptic taxa within the taxonomic species. However, the nature of independent evolution remains unclear. Clusters might simply represent geographically isolated, or even partially geographically isolated, populations evolving neutrally [32,44]. Alternatively, the clade might have diversified into ecologically distinct species experiencing divergent selection pressures. To resolve these alternatives, we test directly for divergent selection between different lineages, adapting classic methods from molecular population genetics [31,45]. If rotifers have experienced divergent selection on trophi morphology between species, for example, adapting to changes in habitat or resource use, we expect low variation within species and high variation between species, relative to the same ratio for neutral changes.
To explore the level at which divergent selection acts on morphology, we compared rates of morphological change within clusters, between clusters within taxonomic species, and between taxonomic species, in each case relative to silent substitution rates in cox1, assumed to reflect neutral changes (see Materials and Methods). The test is robust to sampling issues and differences in mutational mechanism between morphology and cox1 (see Materials and Methods). The results reveal significant evidence for divergent selection on trophi size and PC2 (Figure 5; Table S5). However, divergent selection occurs between taxonomic species, not between clusters; both traits are conserved within taxonomic species but diverge rapidly between species, relative to neutral expectations. Changes in PC1 are more complex, being lower between clusters either than within clusters or between taxonomic species. However, overall the results demonstrate divergent selection on the size and some aspects of shape of the trophi.
Our results show that Rotaria has undergone adaptive diversification in feeding morphology, presumably associated with specialization to different habitats. The finding is supported by observations of ecological differences among the traditional species. For example, R. socialis and R. magnacalcarata live externally on the body of the water louse Asellus aquaticus but partition their use of the host, with the former living around the leg bases and the latter on the anterior, ventral surface. Our analyses show that these traditional species, which are found living together on single louse individuals, are evolutionarily independent and distinct entities. Another traditional species, R. sordida, is found in more terrestrial habitats than the other species, although it sometimes co-occurs with R. tardigrada, which is generally more aquatic (Table S2). Therefore, informal observations of habitat partitioning and coexistence at local scales add further support to the role of niche partitioning.
Not all of the entities identified as genetic clusters display evidence of divergent selection on feeding morphology: the signature of divergent selection was detected at a broader level than that of independently evolving clusters. One possible explanation is that some clusters arose solely from neutral divergence in complete or partial geographical isolation [32,44]. Some of the clusters do comprise geographically localized sets of samples, but at least one traditional species, R. macrura, contains two clusters without obvious geographical separation. Alternatively, divergent selection might act at different hierarchical levels on different traits [17]: clusters might have diverged in unmeasured traits such as behavior, gross body morphology, or life history. Future work sampling additional genetic markers and phenotypic traits for the identified clusters might distinguish these alternatives.
So which level should we call “species”? As increasingly recognized in reviews of species concepts, the answer will depend on which aspect of diversity is of most interest and on the intended use of the delimitation [3,46]. For evolutionary studies, for example, into how bdelloids might adapt to changing environments, the genetic clusters provide statistical evidence of independent evolution within the traditionally recognized species that needs to be taken into account. For ecological studies, the traditional species conform closely to units that are ecologically distinct in terms of feeding morphology. Perhaps surprisingly for a poorly studied group of microscopic animals, traditional species limits appear to be robust for many purposes with the exception of the paraphyletic R. rotatoria. However, the important point here is that the same issues apply to studies of sexual organisms. Genetic surveys often reveal cryptic species within morphologically coherent sexual species and elicit the same arguments over their interpretation [3,33,34].
We conclude that bdelloids display the same qualitative pattern of genetic and morphological clusters, indicative of diversification into independently evolving and distinct entities, as found in sexual clades. This refutes the idea that sex is necessary for diversification into evolutionary species. Similar approaches could be used to explore the nature of species in sexual clades—for example, how often is speciation accompanied by ecological divergence compared to a null model of reproductive isolation and neutral divergence [32,47,48]? In addition, clades differing in levels of recombination could be compared to determine how sexual reproduction affects the strength and rate of diversification. Does the requirement for reproductive isolation limit opportunities for speciation in sexuals, or do their faster adaptive rates promote stronger patterns of diversification than in asexuals [9,49]? Microbial eukaryotes, prokaryotes, and fungi could provide additional study clades for such studies [12], linked to genetic studies verifying the presumed lack of recombination [50].
Our study highlights the advantages of statistical analyses of combined morphological and molecular data. Recent work delimiting or identifying species from DNA barcode data [34,51] has been criticized for relying on organelle genome markers, which may not reveal recent divergences or reflect the history of nuclear genes [52,53]. Morphology provides a ready window on adaptive differences between populations, often the first sign of divergence and at the present easier to sample than the genes underlying important traits [54,55], but has lacked the theoretical framework of DNA. Combined analyses, sampling at the population level across entire clades, offer new potential to uncover the nature of species and biological diversity. Our methods could be readily applied to sexual clades and to other cases presenting challenges to current theories, such as groups in which barriers to interbreeding appear to be weak or nonexistent [1].
DNA was isolated either from clonal samples of five to 25 individuals grown in the laboratory from a single wild-caught individual or from single wild-caught individuals using a chelex preparation (InstaGene Matrix; Bio-Rad, http://www.bio-rad.com). The 28S rDNA and cox1 mtDNA were amplified and sequenced by PCR as described in Protocol S1. Trees were reconstructed from the cox1 and 28S rDNA matrices separately and from a combined matrix for all individuals with at least one gene sequenced. Bayesian analyses were run in Mr Bayes (http://mrbayes.csit.fsu.edu) 3.1.1 for 5 million generations with two parallel searches, using a general transition rate (GTR) + invgamma model [56]. The combined analysis implemented a partition model with a separate GTR + invgamma model and rate parameter for the two partitions. Maximum parsimony support was assessed using 100 bootstrap replicates, searching each heuristically with 100 random addition replicates and TBR branch swapping in Paup*4.10. Eight individuals from the related genus Dissotrocha were included as outgroups. Comparisons of the two genes are described in Protocol S1 and Text S1.
Trophi were prepared for scanning electron microscopy (SEM) by dissolving soft tissues on a cover slide with sodium hypochloride (NaOCl 4%), rinsing with deionized water, dehydrating at room temperature, and sputter-coating a thin layer of gold. Shape was measured by Generalized Procrustes Analysis (GPA) [57] of six landmarks on digitized pictures of the cephalic (ventral) view (Figure 1). GPA coordinates were used for PC analysis after projection onto an Euclidean space tangent to the shape space (see Protocol S1). Size was expressed as centroid size of the landmark configuration. We attempted to culture all individuals, to allow morphometrics and sequencing on individuals from the same clone. However, not all clones survived in the laboratory; for these, we used replicate individuals from the same wild population where possible. In total, we measured 326 SEM pictures of trophi from 23 populations belonging to eight species (see Table S1b). For species with both laboratory-cultured and wild-caught measures, we found no evidence that sample type influenced either the mean or variance of size and shape measures (Table S6), indicating respectively that species differences are genetically based (not environmental) and that there appears to be little genetic variation for morphology within populations. Statistical analyses were performed using the R statistical programming language [58] and routines in the Tps series of programs [59].
Under the null model that the entire sample derives from a single population obeying a single coalescent process, we calculated the likelihood of waiting times, xi, between successive branching events on the DNA tree as
with
where ni is the number of lineages in waiting interval i, λ is the branching rate for the coalescent (the inverse of twice the effective population size in a neutral coalescent), and p is a scaling parameter that allows the apparent rate of branching to increase or decrease through time, fitting a range of qualitative departures from the strict assumptions of a neutral coalescent, for example, growing (p < 1) or declining (p > 1) population size [30]. Under the alternative model that the sample derives from a set of independently evolving populations, each one evolving similarly to the null case, we calculated the likelihood of waiting times as Equation 6 from Pons et al. [30]. The alternative model optimizes a threshold age, T, such that nodes before the threshold are considered to be diversification events with branching rate λD and scaling parameter pD. Branches crossing the threshold define k clusters each obeying a separate coalescent process but with branching rate, λC, and scaling parameter, pC, assumed to be constant across clusters. The alternative model thus has three additional parameters. Models were fitted using an R script available from T.G.B. to an ultrametric tree obtained by rate smoothing the combined analysis DNA tree using penalized likelihood in r8s (http://ginger.ucdavis.edu/r8s) and cross-validation to choose the optimal smoothing parameter for each tree [60].
In an asexual clade, all genes have the same underlying genealogy: the entire genome is inherited as a single unit. Assuming that silent substitutions are neutral, the expected number of silent mtDNA substitutions on a branch of the genealogy is μt, where t is the branch length in units of time and μ is the mutation rate of the gene. Assuming a neutral morphological trait evolving by Brownian motion, the expected squared change (variance) along a branch is
, where
is the mutational rate of increase of variance [61]. The expectations are the same for branches within populations or between them. Therefore, the average rate of change of a neutral trait expressed as variance per silent substitution should be the same within populations as between them, that is,
This prediction holds even if mutation rates vary across the tree, providing they do so without a systematic bias between the branch classes being compared, a reasonable assumption shared with widely used molecular versions of the test [31].
We reconstructed evolutionary changes in trophi size and shape (PC1 and PC2) onto the DNA tree using the Brownian motion model by Schluter et al. [62] implemented in the Ape library for R [63]. Branch lengths were optimized as the proportion of silent substitutions per codon using PAML software [64]. The null model assumes a constant rate of morphological change across the entire tree. The alternative model labels branches as between taxonomic species, within species and within clusters, and estimates different rates for each class. Under a three rate-class model, the likelihood of the reconstruction, Equation 3 of [62] becomes the product of the equivalent likelihood for each class of branches.
where k indicates the branch classes from 1 to 3, βk is the rate parameter for each class of branches, Nk is the number of nodes ancestral to each class of branch, and Q(ũk) is the sum of the scaled variance of changes across branches [62] of class k. Optimization was implemented in a modified version of the “ace” function of Ape, available from T. G. B. Divergent selection between taxonomic species, for example, would be indicated by a significantly lower rate within cluster and within species branches (classes 1 and 2) than between species branches (class 3). Assumptions and robustness of the test are discussed further in Protocol S2.
DNA sequences have been deposited at GenBank (http://www.ncbi.nlm.nih.gov/Genbank) under accession numbers DQ656756 to DQ656882.
|
10.1371/journal.pbio.2004486 | Calmodulin fishing with a structurally disordered bait triggers CyaA catalysis | Once translocated into the cytosol of target cells, the catalytic domain (AC) of the adenylate cyclase toxin (CyaA), a major virulence factor of Bordetella pertussis, is potently activated by binding calmodulin (CaM) to produce supraphysiological levels of cAMP, inducing cell death. Using a combination of small-angle X-ray scattering (SAXS), hydrogen/deuterium exchange mass spectrometry (HDX-MS), and synchrotron radiation circular dichroism (SR-CD), we show that, in the absence of CaM, AC exhibits significant structural disorder, and a 75-residue-long stretch within AC undergoes a disorder-to-order transition upon CaM binding. Beyond this local folding, CaM binding induces long-range allosteric effects that stabilize the distant catalytic site, whilst preserving catalytic loop flexibility. We propose that the high enzymatic activity of AC is due to a tight balance between the CaM-induced decrease of structural flexibility around the catalytic site and the preservation of catalytic loop flexibility, allowing for fast substrate binding and product release. The CaM-induced dampening of AC conformational disorder is likely relevant to other CaM-activated enzymes.
| Calmodulin is a widespread and highly conserved protein that interacts with a wide variety of eukaryotic proteins and enzymes, controlling their activities in response to calcium. The adenylate cyclase toxin (CyaA) of Bordetella pertussis, the causative agent of whooping cough, is one such calmodulin target. Once transported across the plasma membrane of eukaryotic cells, the catalytic domain (AC) of CyaA is activated by calmodulin, producing high levels of cAMP, which can induce cell death. We use an integrative structural biology approach combining several biophysical techniques to characterize the structural rearrangements in AC upon calmodulin binding and to elucidate their relationship to CyaA activation. We show that a disordered stretch of 75 amino acid residues in AC serves as a bait for calmodulin capture. Binding induces significant folding within this region, a prerequisite for CyaA activation. Calmodulin binding promotes the stabilization of the distant catalytic site, whilst maintaining its catalytic loop in a flexible and exposed state. Both phenomena contribute to the high enzymatic activity of AC, allowing for fast substrate binding and cAMP release. The calmodulin-induced reduction of AC conformational disorder is likely relevant to other calmodulin-activated enzymes.
| Calmodulin (CaM) is a major mediator of calcium signaling in eukaryotic cells, ubiquitously distributed and highly conserved in the whole eukaryotic kingdom. This small acidic protein of 148 amino acids interacts with a wide variety of target proteins or enzymes to control their activities in response to calcium. CaM is made of two pairs of calcium-binding EF-hand motifs that are connected by a long flexible helix, adopting a dumbbell shape in solution. The binding of calcium ions to each EF hand triggers conformational changes that result in the exposure of hydrophobic patches, altering the association of CaM with its target proteins.
Typically, CaM interacts with its targets by binding a short segment, or CaM-binding site (CBS), of about 20–30 residues that are positively charged and can adopt an amphipathic helical structure. CaM binding to the CBS is associated with a large conformational change as the protein bends from its extended dumbbell shape into a more globular structure in which its two N- and C-lobes wrap around the target helical sequence. In many instances (e.g., myosin light chain kinase [MLCK], CaM kinases, calcineurin, CaM-activated Ca-ATPase/pumps), the CBS sequence is proximal in the primary structure to an auto-inhibitory domain (AID), which occupies the catalytic site of the enzyme in the resting state, thus inhibiting enzymatic activity. Binding of CaM to the CBS induces a conformational change that displaces the AID from the enzymatic site to release the full catalytic activity.
An original mechanism of CaM regulation has been uncovered in two toxins produced by two pathogenic bacteria, Bacillus anthracis, the causative agent of anthrax, and B. pertussis, responsible for whooping cough. These toxins, the edema factor and the adenylate cyclase toxin (CyaA), respectively, are adenylate cyclases secreted by the bacteria and are able to invade eukaryotic target cells, in which they are potently activated by CaM to produce cytotoxic, supraphysiological levels of cAMP [1, 2]. CyaA is essential for the early steps of colonization of the respiratory tract by B. pertussis, selectively targeting the innate immune cells to favor bacterial survival. CyaA is a 1,706-residue-long bifunctional protein; the CaM-activated, catalytic domain (AC) is located in the 364 amino-proximal residues, while the carboxy-terminal 1,342 residues are involved in toxin binding to target cells and delivery of the AC domain across the plasma membrane to the cell cytosol [3–5].
How CaM binding regulates the activity of the edema factor toxin has been illuminated by a series of structural studies carried out by W.J. Tang and colleagues, who solved the structures of both the CaM-free and -bound forms [6]. Edema factor consists of an N-terminal protective antigen (PA)-binding domain; a catalytic core made up of two globular subdomains, CA and CB; and a C-terminal helical domain. The latter makes substantial contacts with the catalytic core in the absence of CaM and locks the enzyme in an inactive state. Most of these contacts are disrupted upon CaM binding, which inserts in an extended conformation between the catalytic core and the helical domain. Through a series of conformational changes, CaM indirectly stabilizes the conformation of a substrate-binding loop that is fully disordered in the CaM-free form.
In contrast to edema factor, the exact mechanism of CyaA activation by CaM is unclear. To date, the structures of AC in the absence of CaM and of AC bound to full-length CaM remain unsolved; Guo and colleagues [7] have determined the structure of AC in complex with the C-terminal domain of CaM (C-CaM) only. Although AC and the edema factor display limited sequence identity, their catalytic cores share substantial structural similarity: like edema factor, AC is organized into two globular subdomains, CA and CB. The catalytic site, located at the interface between these two domains, is essentially identical to that of the edema factor: it is made of three highly conserved regions (catalytic region 1 [CR1], residues 54–77; catalytic region 2 [CR2], residues 184–198; and catalytic region 3 [CR3], residues 295–315) that are directly involved in substrate binding and catalysis. Nevertheless, the edema factor and AC significantly differ in CaM binding. AC has no equivalent of the edema factor C-terminal helical domain that is involved in CaM binding and activation. Instead, C-CaM binds mainly to the H-helix (spanning residues 234–254 [7]), with typical characteristics of classical CBS, i.e., a basic and amphipathic sequence, which protrudes from the CA subdomain. C-CaM also interacts with a loop-helix-loop (LHL) motif (residues 341–364) located at the very C-terminal end of AC. As this segment directly contacts the CR3 catalytic loop, C-CaM could thus indirectly stabilize it in a configuration favorable for catalysis [7, 8].
Herein, we have characterized the conformational changes of AC upon binding to its CaM activator through a combination of structural studies in solution, including synchrotron radiation circular dichroism (SR-CD), hydrogen/deuterium exchange mass spectrometry (HDX-MS), and small-angle X-ray scattering (SAXS). Our results indicate that, in the absence of CaM, AC exhibits a large intrinsically disordered region (IDR) of about 75 residues (circa 20% of the AC domain) encompassing the main CBS. CaM binding triggers the folding of the CBS and a strong overall reduction in disorder that is likely crucial for AC activation. The presence of a large IDR in AC may facilitate intoxication (translocation across the membrane) of target cells and subsequently favor its association with the CaM activator via a “fly-casting” mechanism [9].
We performed parallel studies of five different samples using both HDX-MS and SAXS: CaM and AC alone, CaM in complex with the H-helix peptide from AC, and a peptide from myosin light-chain kinase (MLCK) as a reference, and finally, the AC-CaM complex. SAXS measurements were recorded using the size exclusion chromatography followed by SAXS (SEC-SAXS) setup available on the SWING beamline at the SOLEIL synchrotron (a succinct description of instrumental conditions is given in Materials and methods and [10]). All SAXS experimental and analytical details are given in S1 Table, together with the values of global structural parameters (radius of gyration [Rg], maximal dimension [Dmax], and molecular mass) derived from the scattering data. Of note, all molecular mass values indicate that the investigated particles are present as monomers of AC or CaM in solution. HDX-MS experimental details are to be found in the relevant section of Materials and methods.
SAXS data from the free AC protein were initially modeled, starting from the atomic coordinates of AC extracted from the Protein Data Bank (pdb) dataset 1YRU of the crystal structure of AC in complex with the C-CaM [7]. Using AllosModFoxs [11], we adjusted the calculated scattering profile against experimental SAXS data by refining the proposed position of the missing residues in 1YRU (residues 226–232, defined as the Hom-loop, and the first six N-terminal residues), while keeping the AC structure unchanged. An example of one of the fits (χ2 = 2.57) obtained in this way is shown in Fig 1A. The corresponding model, shown in blue in Fig 1D, suggests that AC adopts a conformation roughly similar to that of the crystal, albeit with significant differences. Using the same degrees of freedom and a description in terms of ensemble of conformations did not allow us to further improve the agreement with our experimental data.
We therefore undertook a second cycle of SAXS modeling by incorporating the structural data deduced from HDX-MS and far-ultraviolet (far-UV) SR-CD data. Regarding local-level HDX-MS analysis, pepsin digestion of AC yielded 181 unique peptides identified from their accurate masses and production spectra. A total of 36 peptides covering 100% of the AC sequence were included in the final data analysis (S1 Fig; the nomenclature of secondary structures from W.J. Tang [7] is used throughout this work and reported in the legend of S1 Fig). HDX-MS experiments performed on the isolated AC protein show a significant difference in structural content between the N-terminal (T25 trypsin-cleavage fragment of CyaA, from residues 1–224) and C-terminal (T18 trypsin-cleavage fragment of CyaA, from residues 225–364) regions of the protein (Fig 2A, top panel; S1 Data). Dynamic HDX-MS behavior, indicative of the presence of secondary structural elements, was observed throughout the N-terminal moiety of AC (specifically, in A–E-helices and the beta-sheets of T25), confirming that T25 domain is well folded. Two peptides, namely 51–68 and 163–172, give no dynamic HDX-MS pattern and correspond to loops between structural elements. The C-terminal moiety, on the other hand, appears poorly structured: the few regions exhibiting dynamic HDX-MS behavior cover the I- and J/J′-helices and the second beta-sheet of the T18 fragment (T18b2) region, respectively. Strikingly, the amide hydrogens from the region covering residues 201–275 are fully exchanged from the first time point of the HDX experiment to the last, indicating that this region of circa 75 amino acids is essentially disordered (Fig 2A and 2C). This result provides direct evidence that the F-, G-, H-, and H′-helices, the Hom-loop, and the first beta-sheet of the T18 fragment (T18b1) are disordered in the isolated AC protein. This is somewhat unexpected, as in the X-ray structure of AC:C-CaM, this region is predominantly helical [7]. Finally, the catalytic loop (residues 300–315) is fully exposed to the solvent.
In view of the major differences observed between our HDX-MS results and the crystal structure in a helix-rich part of the crystal structure, we decided to complement our experimental approach using the spectroscopic far-UV SR-CD method, which constitutes a very sensitive probe of a protein helical content in solution. The helical content of AC was estimated from the SR-CD spectrum (S2 Fig) using the BestSel software [12] and compared to that of the crystal structure using the Dictionary of Secondary Structure of Proteins (DSSP) algorithm [13, 14] (S2–S4 Tables and S2 Data). The analysis of SR-CD data indicate that 27% of all residues are part of helical secondary structures, a value significantly lower than the 36% value derived from the X-ray structure using DSSP. However, the SR-CD value favorably compares with HDX-MS results, in which the region corresponding to F-, G-, H- and H′-helices and the T18b1 appear to be disordered in solution (Fig 2A).
Taking into account these results from HDX-MS and far-UV SR-CD, the residues from 200–270 were released and allowed to move in SAXS modeling to obtain a new set of models and associated fits, an example of which is shown in Fig 1B (χ2 = 1.36). The distributions of reduced residuals corresponding to the two fits are shown in Fig 1C, while the conformation with relaxed F–H′-helices is shown in red in Fig 1D. This model exhibits a much better fit to the SAXS data and also displays a helical content close to that experimentally determined by SR-CD (S2 Fig and S2 Table). Overall, our model of AC in solution accounts for all three sets of experimental data.
CaM has been the object of many SAXS studies over the last few decades [15–19]. Ab initio modeling of our CaM data using the DAMMIF/DAMMIN programs [20] yields the well-known dumbbell shape shown in Fig 3A. Yet, the experimental scattering pattern differed significantly from that calculated from the crystal structure of CaM (1CLL [21]) (Fig 3D, χ2 = 37, green curve)—not surprisingly, given the flexibility of the inter-domain helix, as revealed previously by nuclear magnetic resonance (NMR) studies and corroborated by our present HDX-MS data (see Fig 4, described later). This flexibility makes more difficult the ab initio modeling of the protein on the basis of the SAXS data. Hence, CaM was first described as an ensemble of conformations using the Ensemble Optimization Method (EOM) suite of programs [22]. Ten thousand conformations were generated by the Ranch program, starting from the structures of each Ca2+-bound domain determined in solution from NMR data by Chou and colleagues (pdb files 1J7O and 1J7P for N- and C-terminal domains, respectively [23]). Residues in the center of the inter-domain helix were described as dummy residues with variable positions. These dummy residues were substituted by a full-atom description using the program PD2 [24]. At this point, the Gajoe program was used to select ensembles of conformations, the average scattering pattern of them being fitted against our experimental data using a genetic algorithm. Fig 3D shows a typical ensemble of conformations that yields a good fit to the experimental data (χ2 = 2.15, red curve). This ensemble of conformations illustrates that CaM is made of two folded lobes connected by a highly flexible hinge.
Pepsin digestion of CaM yielded 64 unique peptides, covering approximately 88% of the protein sequence. From these, 20 peptides were selected for HDX-MS analysis (S3 Fig). The dynamic HDX-MS pattern of the isolated CaM protein is typical of a folded protein (Fig 4A, S4 Fig and S3 Data). Notably, however, the central part of the inter-lobe helix (residues 66–92, represented here by peptide 74–85) is fully exchanged, indicating that these residues do not adopt a stable helical conformation but rather behave as a flexible link between the N-terminal domain of CaM (N-CaM) and C-CaM. The α-helical structural content of the isolated CaM protein inferred from SR-CD experiments (S2 Fig, S2 Table, and S2 Data) is 51 ± 3%. This is, however, significantly lower than that observed in the CaM X-ray structure (62%) and most likely due to the disordered segment within the central part of the inter-lobe helix observed in our HDX-MS results.
The H-helix region of AC represents the primary site of interaction with CaM, as seen in the crystal structure of AC with C-CaM [7]. Accordingly, the complex of CaM with the H-helix peptide was investigated in parallel. This allowed us to decipher those additional regions of AC that are involved in AC:CaM complex formation and stabilization. Although the joint presence in solution of both 1:1 and 1:2 CaM:H-helix peptide complexes severely limits modeling attempts, ab initio modeling indicated that the global dumbbell shape of CaM was not altered in the complex, as shown in Fig 3B, which presents a typical envelope obtained with DAMMIF/DAMMIN. This is clearly apparent when examining the two distance distribution functions P(r) that exhibit similar maximal extensions with or without the H-helix peptide bound (Fig 3E). These data suggest that H-helix peptide binding does not cause any significant modification of the CaM dumbbell shape and its inter-domain helix.
Far-UV SR-CD indicates that more than 20 amino acids undergo a disorder-to-helix transition upon CaM:H-helix complex formation (S2 Fig, S4 Table and S2 Data). The binding of the H-helix peptide does not create any additional dynamic HDX-MS activity within CaM (S4 Fig and S4 Data). Consequently, the observed increase in helical content can be attributed to the disorder-to-order transition within the H-helix peptide itself. The H-helix peptide appears to stabilize the C-CaM helices of CaM (Fig 4C and S4 Fig) and to acquire a helical conformation upon CaM binding (S2B Fig, S2 and S4 Tables).
Ab initio modeling of SAXS data indicated that the global dumbbell shape of CaM was not altered upon binding of the H-helix peptide. We therefore undertook equivalent experiments on the CaM protein in complex with the myosin light-chain kinase peptide (PMLCK), which is known to reshape CaM from its extended dumbbell conformation into a compact globular structure. In marked contrast to the H-helix peptide, the shape of the CaM:PMLCK complex determined ab initio using the programs DAMMIF/DAMMIN was clearly globular and isometric (Fig 3C). Noticeably, the experimental scattering curve of CaM:PMLCK is very similar to the scattering pattern calculated from the known crystal structure of this complex (2K0F.pdb) (Fig 3F), in which the two domains of CaM are wrapped around the PMLCK peptide. Altogether, these SAXS studies confirm the distinct modes of interaction of CaM with the H-helix peptide of AC and the PMLCK peptide from MLCK. This difference is also clearly evident when comparing the distance distribution functions P(r) (Fig 3E): CaM:PMLCK exhibits a bell-shaped curve with a single maximum, while the CaM:H-helix profile presents an extra shoulder characteristic of a two-domain structure.
Finally, PMLCK binding exerts a more pronounced reduction in solvent accessibility than the H-helix throughout the full-length CaM protein, with both N- and C-CaM lobes affected (S5 Fig and S5 Data). However, the linker connecting both domains remains solvent exposed throughout, as previously observed with both the H-helix peptide and the full-length AC domain.
The HDX-MS data indicate that the binding of CaM induces several statistically significant (p ≤ 0.01) changes in AC (Fig 2A and 2D and S1 Data). Globally, the N-terminal moiety is stabilized, while the C-terminal domain now exhibits several regions characterized by HDX dynamic patterns, indicative of secondary structure acquisition within T18. The main effects in the N-terminal part of AC are the stabilization of both B- and C-helices and of the embedded T25 antiparallel beta-sheet (peptides 25–39, 46–61, and 192–201), as highlighted in Fig 2B. These structural elements become completely resistant to deuteration over the experimental timescale (Fig 2A and 2C). The internal loop (51–68), which was fully accessible in the isolated AC, exhibits a dynamic HDX-MS behavior in AC:CaM. In contrast, the regions corresponding to the G-helix as well as the Hom-loop are unaffected by CaM binding, suggesting that these solvent-exposed regions in the isolated AC protein remain accessible throughout (Fig 2A–2C). Upon complex formation, the majority of the C-terminal half of AC acquires secondary structure elements, except the catalytic loop (residues 304–315 from CR3), which remains completely solvent exposed. Most notably, the regions encompassing the H/H′-helices and T18b1, which were fully exposed in AC alone, are structured in the presence of CaM (see dynamic HDX behavior in Fig 2A–2C). The I- and J/J′-helices, which were already structured in the absence of CaM, are further stabilized, as evidenced by the observed reduction in deuteration level.
Reciprocally, we examined the consequences of AC binding for CaM structural dynamics. HDX-MS data show that AC binding induces a strong stabilization of C-CaM, as evidenced by the dramatic reduction in accessibility imposed by the wrapping of C-CaM helices around the H-helix region of AC (Fig 4B and S6 Fig). The loop between the first two helices, H5 and H6, of C-CaM is also significantly stabilized, likely via interactions with the J/J′-helices and the C-terminal loop of AC. The involvement, beyond the CBS-containing H-helix, of various regions of AC in CaM stabilization is best illustrated by the “difference of uptake differences” plot (Fig 4C). In contrast, the majority of N-CaM is unaffected, besides peptide 38–48, which covers the loop connecting helices H2 and H3. Notably, the central part of the inter-lobe helix of CaM is also unaffected by AC binding and remains solvent exposed upon complex formation.
Finally, we recorded the far-UV SR-CD spectrum of the AC-CaM complex (S2 Fig) to estimate the CaM-induced helical content increase of AC. Assuming a constant helical content of CaM, we derive for AC an increase in its helical content from 27% to 42 ± 3% upon CaM binding (S2–S4 Tables). This increase, corresponding to about 40 amino acids (S3 Table), is proposed to be primarily due to the CaM-induced folding of the H/H′ regions, as reported by the HDX-MS results (Fig 2).
The first modeling attempt of the AC-CaM complex from our SEC-SAXS data made use of the BUNCH program, which considers that the particle adopts a unique conformation in solution and searches for one conformation that yields a good fit to the experimental data. More specifically, AC and each CaM domain were handled as rigid bodies while the program searched an optimal conformation of the inter-domain helix of CaM, the only variable part of the complex. We started from the crystal structure 1YRU of the complex of AC with the C-CaM and the solution structure 1J7O of the N-CaM. Furthermore, the central residues of the inter-domain helix were described as dummy residues, the positions of which were varied by the program in order to adjust the calculated scattering pattern to experimental data. We made one change to the 1YRU structure based on our experimental HDX-MS data: residues 200–234, corresponding to helices F and G together with the disordered Hom-loop, remain unfolded in the AC:CaM complex. Indeed, examination of the crystal structure and of its contacts with neighboring molecules shows that these two helices are interacting with their counterpart from an adjacent molecule within the crystal. Moreover, these residues are not predicted to adopt a helical conformation by secondary structure predictions. Accordingly, we assumed that the two helices are strongly stabilized within the crystal but may only transiently exist in solution. Therefore, we released the position of residues 200–234 of AC so that this stretch was also variable during the refinement process. Thirty runs of the program BUNCH yielded models similar to that shown in S7 Fig. The corresponding fit to the experimental data is also shown in S7 Fig (χ2 = 2.0). Noticeably, the N-CaM is immersed in the solvent at a distance from the rest of the complex, without any contact with AC (S7 Fig).
The high flexibility of the central CaM linker suggests that a better description of the complex would make use of an ensemble of conformations (as opposed to a unique structural model). Accordingly, we undertook a second round of complex modeling using the program EOM. The 10,000 conformations of the initial pool differ in the relative position of the two CaM domains as well as in the conformation of the region 200–234 of AC. An example of the resulting ensemble of conformations is displayed in Fig 5A, with the corresponding fit to the SAXS data shown in Fig 5B (χ2 = 1.41). Interestingly, the SAXS data are compatible with a description of the complex in which the flexibility of the CaM inter-domain helix makes possible transient interactions of the CaM N-terminal domain with AC, as suggested by the results of HDX-MS experiments (CaM peptide 38–48, see Fig 4B).
Exploiting an integrative structural biology approach, the present study proposes detailed and novel structural descriptions of the catalytic AC domain of the CyaA toxin, both in isolation and in the presence of its activator, CaM, and describes how specific conformational rearrangements within each molecule may be requisite to trigger the enzymatic activity. The isolated AC protein appears significantly less folded in solution than in the crystal structure of the AC:C-CaM complex [7]. Although the N-terminal T25 domain is folded, the C-terminal T18 moiety contains few structural elements: most noticeably, HDX-MS identifies a large region of 75 amino acids (residues 201–275) in T18, which is predominantly structurally disordered. This IDR includes the H-helix region, which is the main CBS of AC, as well as its flanking regions, namely the Hom-loop and the F-, G-, and H′-helices (Fig 6A). This is in agreement with prior studies that showed that the isolated AC domain is partially folded [25–27].
AC:CaM complex formation primarily takes place between the H/H′-helix region of T18 and C-CaM, inducing the folding of T18 as well as establishing long-distance stabilization of structural elements within the T25 moiety (Fig 6B and 6C, S8 and S9 Figs; [7, 28–30]). The SR-CD and HDX-MS data of both AC:CaM and H-helix:CaM complexes strongly suggest that CaM binding triggers the folding of the H- and H′-helices (S1–S4 Tables). However, the area upstream of the H-helix region clearly remains highly flexible in the AC:CaM complex. Several regions of T18, including the I- and J/J′-helices (residues 273–290 and 331–344, respectively) as well as the C-terminal LHL motif (residues 345–358, which interact with the AC catalytic loop [8]) are significantly stabilized upon CaM binding. These different interactions may contribute to the shaping of the catalytic site through the correct positioning of the CR3 loop, which directly interacts with the purine moiety of the ATP substrate. CaM binding also induces significant modifications in the T25 region of AC, including the B- and C-helices, and the conserved catalytic regions CR1 and CR2, which contact the ribose and triphosphate moieties of ATP [7] (Figs 2 and 6). In contrast, the G-helix region remains highly flexible in the AC:CaM complex in solution, and it is likely that the helix is stabilized in the crystal structure of AC:CaM through interactions with its counterpart in symmetry mates. Our experimental SAXS data are indeed fully compatible with AC:CaM models in which all AC residues from 200–234 remain largely disordered.
The significant level of structural disorder observed in AC alone, with a large IDR of about 75 residues, has major implications for the biological functions of the molecule. In the first instance, the IDR and global structural disorder of the T18 moiety may contribute via an “entropic chain effect” to destabilizing the catalytic site and thus preventing cAMP synthesis in the absence of the activator—most importantly, at the time of CyaA production in B. pertussis, when any significant basal activity would be deleterious to the bacterial host, as illustrated by the high toxicity observed when CaM and AC are co-expressed in recombinant Escherichia coli [31]. Secondly, the weak stability of significant parts of the AC structure is favorable for toxin delivery from the bacterial host to eukaryotic target cells. CyaA is secreted outside the bacteria by a type I secretion system (T1SS), and then, upon binding to target cells, AC is delivered across the plasma membrane to reach the eukaryotic cell cytosol [32]. There is evidence to suggest that both the secretion through the T1SS and the translocation across the eukaryotic cell membrane are critically dependent upon the ability of the polypeptide chain to adopt partially unfolded states [5, 33–36]. The disordered region and overall low structural content of AC in solution should clearly facilitate these key steps of the intoxication process. Thirdly, the intrinsic disorder of the central region of AC, which harbors the main CBS, namely the H-helix, potentially plays a crucial role in molecular recognition of the activator [37, 38]. Indeed, the CBS located within the IDR of AC (see the long loop in Fig 6A) exemplifies the canonical properties of so-called molecular recognition features (MoRFs) or short linear motifs (SLIMs), which are short and conserved motifs involved in recognition (in silico analysis in S8 Fig). MoRFs often comprise a few hydrophobic residues amidst charged residues and are frequently embedded within IDRs that serve as flexible linkers to efficiently expose the MoRFs to the solvent, thereby favoring their potential interactions with partner biomolecules [37]. The MoRF in AC is typical of eukaryotic CaM-binding targets: it is 10–20 residues long and is flanked by disordered regions, with a high net charge and a propensity for structural disorder in the CaM-free state (S8 Fig). Moreover, it undergoes a disorder-to-helix transition upon complex formation [39]. Such coupled binding and folding processes appear to be a common mechanism by which CaM activates its target enzymes, including, for example, the CaM-dependent protein kinases and calcineurin [40]. The AID of these latter proteins occludes access to the catalytic site in the resting state. A CaM-induced conformational change physically shifts the AID from the enzymatic site, allowing catalysis. In the case of AC, the structural disorder of the catalytic site induced by the IDR replaces the steric hindrance of the AID.
Once AC has been delivered across the target cell plasma membrane, the long IDR may facilitate the efficient capture of the intracellular CaM through a “fly-casting” mechanism [9]. CaM binding promotes folding of the H/H′ region and stabilization of several structural elements of T18 (Figs 2 and 6). The reduction of disorder is then propagated to the core of the T25 domain to allosterically stabilize other key regions shaping the catalytic site (Fig 6B and 6C). This suggests that the overall dampening of disorder triggered by CaM-binding is crucial for AC activation. Interestingly, our HDX-MS data indicate that, although CaM binding imposes a strong entropy reduction and steric restrictions on regions flanking the CR3 catalytic loop, the loop itself (residues 300–314) remains largely dynamic and solvent exposed, as indicated by the rapid and high deuterium uptake (S10 Fig). The high flexibility and structural dynamics of this catalytic loop that binds the purine ring of ATP are likely crucial for the efficient catalysis of AC, which has a turnover number higher than 1,000 s-1, i.e., one ATP converted to cAMP in less than a ms.
A striking property of AC is that it is fully activated by the C-terminal lobe of CaM, with an affinity that is only 10–20-fold lower than that of the full-length CaM protein [7]. As only the structure of AC in complex with C-CaM has been reported thus far, the structure and mode of interaction of AC with the full-length CaM molecule has remained elusive. Our HDX-MS analysis indicates that the C-CaM lobe is strongly stabilized upon complex formation (Fig 6D and S11 Fig), while the N-CaM domain remains mostly dynamic and essentially unaffected by the presence of AC (Figs 4 and 5). This is in good agreement with the Springer and colleagues NMR study, which also showed that AC binding primarily affects C-CaM, with minimal structural changes to N-CaM [41]. Notably, when bound to either AC or the H-helix peptide, CaM remains in an extended conformation (Figs 3 and 5), with high flexibility of the central CaM inter-lobe helix (Fig 4, S4 and S6 Figs). Our SAXS data are in excellent agreement with structural models in which CaM interacts with AC essentially via its C-terminal lobe, while its N-terminal lobe, connected through the flexible central linker region (residues 77–83), remains highly dynamic, without establishing stable contacts with AC (Fig 5), although transient interactions are likely, as suggested by both our EOM analysis of SAXS data and our HDX-MS results. Interestingly, these models are reminiscent of those proposed by Guo and colleagues [30] on the basis of molecular dynamics simulations, although these authors further suggested that the β-hairpin region (residues 259–273) of AC directly contacts the second calcium-binding motif of the extended CaM. To probe this model, the group performed cross-linking experiments between AC and CaM variants harboring a unique cysteine at positions 260 and 50, respectively (AC-Q260C and CaM-D50C). However, only a very faint disulfide bond cross-link was detected, indicating that the proposed conformations with the two cysteines in proximity are weakly populated in solution. From our point of view, these data support the view of a fairly dynamic N-CaM lobe that samples a variety of conformations, restricted only by the flexibility of the central hinge (Fig 4). How the N-CaM lobe may contribute to the 10–20-fold higher affinity of native CaM as compared to that of C-CaM [42, 43] remains an intriguing question. One attractive hypothesis could be that the mere presence of the N-CaM module, acting as a “local crowder,” may stabilize the complex by restricting the conformational dynamic of the AC polypeptide chain (mainly the T18 region). Additional stabilization may also be provided by global electrostatic interactions between the overall negatively charged N-CaM and the positively charged cluster of residues around 221–224 of AC [29].
Interestingly, our combined SAXS, HDX-MS, and SR-CD analyses also show that the CaM dumbbell shape [44, 45] is not significantly modified upon binding to the AC H-helix peptide. This is in marked contrast to what is observed with the classical CaM-binding peptide from MLCK, PMLCK (Fig 3), around which both N- and C-CaM are seen to wrap in a conformation that requires a sharp bending of the CaM central helix (Fig 3C, S5 and S12 Figs) [46]. The ability of individual N- and C-terminal lobes of CaM to independently associate to either distinct sites on the same protein or to the same site on two distinct polypeptides, thus facilitating their dimerization, have been described for numerous CaM targets, although in many instances, the affinity of individual lobes with their target sequence is lower than that with the AC H-helix. Our data further highlight the remarkable conformational diversity of CaM upon binding to its targets [44–49].
In the bacterial cytosol, there is no chaperone, inhibitor, or antitoxin available to prevent cAMP production, and thus we propose that structural disorder acts as the inhibitor of CyaA catalytic activity. Once translocated into eukaryotic cells, bacterial toxins such as CyaA hijack CaM, which is used as a disorder dampener to activate the enzyme. The disorder-to-order transition observed in the MoRF region of AC is not directly involved in catalysis but rather transduces the stabilization toward the enzymatic region. This CaM-induced folding and stabilization, from local to distal regions, allosterically tunes the enzymatic activity of bacterial toxins.
Buffer A composition is 20 mM HEPES, 150 mM NaCl, pH 7.4, and 2 mM (HDX-MS and SR-CD experiments) or 4 mM (SEC-SAXS experiments) CaCl2 (final calcium concentration), unless otherwise stated.
The H-helix peptide corresponds to residue numbers 234–254 of AC (LDRERIDLLWKIARAGARSAVG). H-helix peptide molecular mass deduced from mass spectrometry (MS) was 2,509 g/mol, its pI was 11 and it contained +2 charges at neutral pH, and its molar epsilon at 280 nm was 5,600 M−1 cm−1. The CaM-binding peptide (PMLCK) from the smooth muscle MLCK corresponds to residues RRKWQKTGHAVRAIGRLSS (Em: 5,600 M−1 cm−1). All peptides contain an N-terminal acetyl and C-terminal amide as capping groups. All peptides were purchased from Genosphere Biotechnologies (Paris, France).
Human CaM protein was produced in E. coli, as previously described [27]. MS confirmed the integrity and identity of all proteins in the study. MS analysis of CaM gave a single peak of 16,706.25 Da (expected mass = 16,706.30 Da). A molar epsilon at 280 nm of 2,980 M−1 cm−1 and a pI of 4.1 were derived from the CaM sequence using ProtParam.
The adenylate cyclase catalytic domain (AC) toxin under investigation corresponds to residues 1−364 of B. pertussis CyaA. The protein was produced in E. coli and purified as follows: protein-containing inclusion bodies from the bacterial cell pellet were resuspended in 6 M GdnHCl, 20 mM HEPES, pH 7.4, overnight at 4°C, and then buffer exchanged from 6 M GdnHCl to 6 M urea on prepacked G25SF desalting columns. Inclusion body solubilization was performed in GdnHCl instead of urea to prevent AC364 carbamylation. AC364 was purified by two successive chromatographic separations on DEAE-Sepharose [4]. The protein was initially eluted in 8 M urea, 20 mM HEPES, and 500 mM NaCl, pH 7.4, followed by a second elution in 20 mM HEPES and 100 mM NaCl, pH 7.4. AC364 was further purified on a Sephacryl S200 column equilibrated with Buffer A and concentrated on an Amicon 10 kDa MWCO filter. The identity and purity of protein batches was monitored by SDS-PAGE and MS. AC364 gave a single MS peak of 39,380.90 Da (expected mass = 39,381.50 Da). A molar epsilon at 280 nm of 28,880 M−1 cm−1 and a pI of 6.2 were derived from the AC364 sequence. All proteins and peptides, whether in solution or lyophilized, were stored at −20°C.
SR-CD experiments were carried out at the synchrotron facility SOLEIL (DISCO beamline, Saint-Aubin, France). SR-CD spectra were recorded in the far-UV region from 180–260 nm with an integration time of 1.2 s, a bandwidth of 1 nm with a 1 nm resolution-step at 20°C in QS cells (Hellma), with a path length of 50 μm. Each far-UV spectrum represents the average of 3 individual scans. Experimental details on the procedure are described elsewhere [33, 35]. Protein and peptide concentrations ranged from 50–200 μM. The far-UV CD spectra were processed using BestSel to estimate the alpha-helical content of proteins, peptides, and complexes [12].
|
10.1371/journal.pntd.0002031 | Polymorphism in the HASPB Repeat Region of East African Leishmania donovani Strains | Visceral leishmaniasis (VL) caused by Leishmania donovani is a major health problem in Ethiopia. Parasites in disparate regions are transmitted by different vectors, and cluster in distinctive genotypes. Recently isolated strains from VL and HIV-VL co-infected patients in north and south Ethiopia were characterized as part of a longitudinal study on VL transmission.
Sixty-three L. donovani strains were examined by polymerase chain reaction (PCR) targeting three regions: internal transcribed spacer 1 (ITS1), cysteine protease B (cpb), and HASPB (k26). ITS1- and cpb - PCR identified these strains as L. donovani. Interestingly, the k26 - PCR amplicon size varied depending on the patient's geographic origin. Most strains from northwestern Ethiopia (36/40) produced a 290 bp product with a minority (4/40) giving a 410 bp amplicon. All of the latter strains were isolated from patients with HIV-VL co-infections, while the former group contained both VL and HIV-VL co-infected patients. Almost all the strains (20/23) from southwestern Ethiopia produced a 450 bp amplicon with smaller products (290 or 360 bp) only observed for three strains. Sudanese strains produced amplicons identical (290 bp) to those found in northwestern Ethiopia; while Kenyan strains gave larger PCR products (500 and 650 bp). High-resolution melt (HRM) analysis distinguished the different PCR products. Sequence analysis showed that the k26 repeat region in L. donovani is comprised of polymorphic 13 and 14 amino acid motifs. The 13 amino acid peptide motifs, prevalent in L. donovani, are rare in L. infantum. The number and order of the repeats in L. donovani varies between geographic regions.
HASPB repeat region (k26) shows considerable polymorphism among L. donovani strains from different regions in East Africa. This should be taken into account when designing diagnostic assays and vaccines based on this antigen.
| HASPB belongs to a hydrophilic repeat-containing surface antigen family found in Leishmania. The L. infantum/L. donovani protein has been used for diagnosis of visceral leishmaniasis, and is a putative vaccine candidate for this disease. Visceral leishmaniasis is a fatal disease, and approximately one third of the cases are found in East Africa. The k26 – PCR, which amplifies the repeat region of HASPB, produced different amplicon sizes for recent Ethiopian L. donovani depending on the strain's geographic origin. Further analysis showed that the number and order of the peptide motifs, either 13 or 14 amino acids long, comprising the L. donovani repeats varies between endemic regions of East Africa. Polymorphism in the amino acid sequence of the peptides was also observed. In addition, the 13 amino acid peptide motifs prevalent in L. donovani are rare in L. infantum. The observed polymorphisms in the HASPB repeat region suggests that custom antigens may be needed for diagnosis or vaccination in distinct endemic foci.
| Parasites belonging to the Leishmania donovani complex, L. donovani and L. infantum (synonym = L. chagasi), are the main causative agents of visceral leishmaniasis (VL), also known as kala-azar. This disease is invariably fatal if not properly diagnosed and treated. The World Health Organization (WHO) estimates that the yearly incidence of VL is between 2–400,000 cases, resulting in 20–40,000 deaths annually with the majority of cases, >90%, occurring in Brazil, the Indian subcontinent and east Africa [1]. VL in the latter region is found primarily in Sudan, South Sudan and Ethiopia where an estimated 30,000–57,000 cases occur each year [1], [2], [3]. In East Africa and India, VL is primarily caused by L. donovani, and believed to be an anthroponosis, while in other regions, where VL is caused by L. infantum, this disease is a zoonosis with dogs and wild canids acting as reservoir hosts [4].
In Ethiopia, VL is distributed throughout the lowlands with the most important foci found in northwestern and southwestern parts of the country. However, the ecology, vectors responsible for parasite transmission, and epidemiology of VL differ between these regions. Northwestern Ethiopia (NW) accounts for ∼60% of the VL cases [3], and a majority of the HIV - VL co-infections, with the disease focused in the Metema - Humera region near the Sudanese border. This is a semi-arid region, with extensive commercial monoculture, and scattered Acacia - Balanite forests. Phlebotomus orientalis is the suspected vector responsible for transmission [3], [5]. The recent large increase in VL in NW Ethiopia has been correlated with agricultural development, and the large influx of seasonal workers ([6], [7]. Migrant workers returning from this area to the non-endemic highlands appear to be responsible for introducing the VL into the latter regions, as typified by the recent outbreak that occurred in Libo-Kemkem, South of Gondar [6]. In southwestern Ethiopia (SW), VL foci are mainly located in the Omo River plains, Segen and Woito Valleys, and near the border with Kenya [3], [8]. These regions include savannah and forest, and P. martini and P. celiae have been implicated as vectors [3], [9], [10]. Disease in Southern Ethiopia appears to be sporadic and stable occurring most frequently among children or young adults [8].
Analysis of parasites belonging to the L. donovani complex using multiple molecular markers that included DNA sequences of protein coding, non-coding and intergenic regions, microsatellites (MLMT) and other techniques, resulted in a revised taxonomy [11]. East African strains, previously split into L. donovani, L. archibaldi or L. infantum by multilocus enzyme electrophoresis (MLEE) are now classified in one group as L. donovani s.s. This large study confirmed several earlier publications using individual molecular techniques [12], [13], [14]. Several of these studies identify genetically distinct populations among the L. donovani complex associated with different geographic regions [13], [14]. Recently, analysis using 14 unlinked microsatellite markers of 90 East African strains, including 63 new isolates from Ethiopia, showed that L. donovani can be divided into two genetically distinct populations, Sudan plus NW, and Kenya plus SW. These major groups could also be further divided into several subpopulations [15]. Although MLMT easily distinguishes between two main L. donovani genotypes in Ethiopia, NW and SW type, and can produce individual parasite genetic pedigrees, it is relatively expensive, requires more sophisticated analysis, and not available in most laboratories working on Leishmania.
HASPB (hydrophilic acylated surface protein B) belongs to a family of orthologous genes, originally called the LmcDNA16 locus, found in Old and New World Leishmania species [16]. The protein is expressed only by metacyclic promastigotes and amastigotes, and is characterized by amino acid repetitive domains that show both inter- and intra-species polymorphism [17], [18], [19]. A recent study using L. major LmcDNA16 locus null mutants, and parasites complemented for either HASPB or the whole locus showed that this protein is involved in metacyclogenesis and promastigote localization in the sand fly vector [20]. The repeat region of the L. donovani and L. infantum HASPB protein, also known as k26, is recognized by human and canine VL sera, and has been used with varying success for serodiagnosis [19], [21], [22], [23], [24], [25], [26]. In addition, HASPB has been shown to be a potential vaccine candidate [27], [28], [29].
A specific PCR targeting the L. donovani complex HASPB repeat region (k26 – PCR) was shown to distinguish between L. donovani and L. infantum strains grouping them according to the size of the amplicon [30]. However, only a few East African strains from Sudan (n = 6) and Ethiopia (n = 2) isolated between 1954 and 2000 were examined. More recently, Gadisa et al. [31] characterize five clinical isolates from VL patients in Ethiopia by k26 - PCR. Only a single PCR fragment was observed, all the same size as the WHO reference strain LV9 (MHOM/ET/67/HU3).
In this study, we characterized 63 recent L. donovani strains from Ethiopia using k26 - PCR, and high resolution melt (HRM) analysis. Several strains from Kenya, Sudan and India were also included for comparison. Analysis by these techniques split the Ethiopian strains into groups that are correlated with the geographic origin of the parasite strain. DNA sequencing of the amplicons showed that the number and organization of the peptide motifs comprising the L. donovani HASPB repeat domain varies with the geographic origin of the strain. Potential effect of k26 polymorphism on use of HASPB for serodiagnosis and vaccination is discussed.
This study was conducted according to the Helsinki declaration, and was reviewed and approved by the Institutional Review Board (IRB), Medical Faculty, Addis Ababa University. Written informed consent was obtained from each study participant.
Leishmania strains (n = 63) recently isolated from patients with VL or HIV - VL co-infections in northwestern (n = 40) and southern Ethiopia (n = 23), see Figure 1, were cultured in M199/Hepes pH 6.8 medium supplemented with 10% fetal calf serum and antibiotics [32]. DNA extraction was carried out using the Gentra DNA extraction kit (Gentra system, Minneapolis, MN). In addition, DNA from L. donovani strains, Ethiopian (n = 24) and Kenyan (n = 7) previously examined by MLMT [15], and from Sudan (n = 2) and India (n = 2) was also analyzed. The strains used in this study are described in Table S1.
Internal transcribed spacer 1 (ITS1) - PCR followed by restriction fragment length polymorphism (RFLP) analysis was carried out as described [33]. A modified, “short” cpbE/F - PCR was used to distinguish between L. infantum and L. donovani, and was carried out using the primers 5-GTTATGGCTGCGTGGCTTG-3 (this study) and 5-CGTGCACTCGGCCGTCTT-3 [34]. DNA (50–100 ng) was added to a PCR - Ready Supreme reaction mix (Syntezza Bioscience, Jerusalem, Israel) in 25 µL total reaction, and performed as follows: Initial denaturation 4 min at 95°C; followed by 35 cycles with each cycle consisting of denaturation 30 s at 94°C, annealing 15 s at 50°C, and extension 60 s at 72°C. Final extension step was carried out for 10 min at 72°C. PCR products were separated by 2% agarose gel electrophoresis, stained with ethidium bromide and visualized using UV light. L. infantum gives a 361 bp product, while L. donovani give a 400 bp product in the short cpbE/F PCR.
K26 - PCR was carried out as described [30], and analyzed by agarose gel electrophoresis as above. HRM analysis of the k26 amplicons was carried out as follows: DNA (20 ng) or no DNA control was added to Type-it HRM PCR Kit reaction mix (12.5 µl, QIAGEN GmbH, Germany) containing the k26 primers (1 µM each final concentration), and ultra-pure PCR-grade water (final volume 25 µl/PCR). Amplification conditions were as follows: 10 min denaturation at 95°C, followed by 40 cycles of denaturation 5 s at 95°C; annealing 10 s at 55°C; and extension 20 s at 72°C. HRM ramping was carried out at 0.2°C/s from 70 to 95°C. HRM PCR and analysis were performed using a Rotor-Gene 6000 real-time thermal analyzer (Corbett Life Science, Australia). Positive-control (reference strain DNA, 20 ng/reaction) and negative-control reactions were included in each experiment. A normalized melt window, ∼85 to 90°C, was used in analyzing the HRM curves.
For direct sequencing, the PCR products were purified using Wizard SV gel and PCR clean-up system purification kit (Promega, WI, USA). The eluted DNA was sequenced at the Center for Genomic Technologies, The Hebrew University of Jerusalem, and the sequences submitted to GeneBank at NCBI. Peptide sequences were obtained using the ExPASy Translate Tool (http://web.expasy.org/translate/). DNA and peptide sequences were aligned using CLUSTAL 2.1 (http://www.ebi.ac.uk/Tools/msa/clustalw2), and linear B-cell epitopes predicted using BepiPred and ABCpred (http://www.cbs.dtu.dk/services/BepiPred and http://www.imtech.res.in/raghava/abcpred/index.html, respectively) [35] [36].
DNA was purified from 63 Leishmania strains isolated from Ethiopian patients presenting with either VL or HIV-VL co-infections. As an initial step the DNA's were first examined by ITS1 - PCR RFLP, and shown to belong to the L. donovani complex (data not shown). Since it can be difficult to distinguish between L. infantum and L. donovani using the ITS1 - PCR RFLP [33], we also analyzed these strains using a modified cpbE/F – PCR based on the procedure described by Hide and Banuls [34]. The L. infantum cpbE and L. donovani cpbF genes are similar except for a 39 bp insert only present in the latter species. This difference is more easily observed by gel electrophoresis using the short cpbE/F – PCR where the amplicon size is 361 bp for L. infantum and 400 bp for L. donovani, rather than 702 and 741 bp, respectively, in the original procedure [34], since the relative size difference between the two short PCR products is larger. This alleviates the need for additional treatments, such as digestion with restriction enzymes [31], [37], which can facilitate species identification. Using the short cpbE/F – PCR all 39 new Ethiopian VL patient strains gave 400 bp PCR products typical of L. donovani (Figure 2, and data not shown), and are identical to the Sudanese reference strain (MHOM/SD/1962/1S cl2, lane Ld). As expected, the L. infantum reference strains (MCAN/IL/2000/LRC-L792 – lanes Li1 and MHOM/TN/1980/IPT1 - Li2) gave a shorter 361 bp product.
The k26 - PCR, a L. donovani complex specific assay, targets the repeat region of the HASPB gene, and was shown to differentiate among L. donovani strains based on the size of the PCR product. L. donovani strains from East Africa gave products <430 bp, and Indian isolates showed significantly larger products (∼660 bp) [30], [31]. Strains previously examined from Sudan (n = 6) and Ethiopia (n = 2) gave two main products, ∼284 and ∼430 bp, with one Ethiopian isolate in each group. These strains were isolated between 11 to 49 years ago, and mutations in the HASPB gene may have occurred over time, or due to repeated passage in culture. In a recent report where five clinical isolates from Ethiopia were examined only one product, ∼290 bp, was observed [31]. Therefore, we decided to examine a large number, n = 63, of recent L. donovani strains isolated from VL and HIV – VL co-infected patients in different geographic regions of Ethiopia. Interestingly, four different amplicon sizes were observed: ∼290, ∼360, ∼410 and ∼450 bp (Figure 3). The PCR product sizes for all the strains examined are summarized in Table S1. Surprisingly, there was a good correlation between geographic origin and amplicon size with strains isolated from patients in northwestern Ethiopia giving either ∼290 or ∼410 bp products, and all the strains isolated in southern Ethiopia, except for three, giving ∼450 bp products. Interestingly, the four strains in the k26-410 cluster were isolated from 3 HIV – VL co – infected patients. Two of the strains were obtained from the same patient, one before drug treatment (LDS 373), and one following relapse (DM376). Prior to drug treatment, the parasites cultured from the spleen or bone marrow of the same patient (LDS 373) gave different k26 amplicon sizes, k26-290 or k26-410 respectively, when examined by PCR. The remaining 11 NE strains isolated from HIV – VL patients all grouped in the k26-290 cluster together with all of the strains isolated from HIV negative VL patients.
Endemic regions for VL in northwestern and southern Ethiopia extend into neighboring Sudan and Kenya, respectively. For this reason, it was interesting to see whether AM553 (k26-360), which gave a unique amplicon different from the other southern strains, represented a second group. This strain is from Negele-Borena close to the border with northwest Somalia and northeast Kenya. Seven L. donovani strains from Kenya were screened by k26 – PCR. All of the Kenyan strains produced amplicons larger than the Ethiopian L. donovani strains examined here. Of these, 6/7 Kenyan strains gave products ∼500 bp and 1/7 strains gave a product of ∼650 bp. Both Sudanese reference strains examined in this study gave a 290 bp PCR fragment, similar to that previously reported [30], and belong to the k26-290 cluster (data not shown).
HRM analysis is a rapid and inexpensive method for detecting polymorphisms in double stranded DNA that can potentially distinguish between single base differences. This technique was used in conjunction with k26 - PCR to examine the Ethiopian strains. Typical results are shown in Figure 4. These results show that this technique can be used to rapidly and easily distinguish between the groups found in Ethiopia (k26-290, -360, -410 and -450).
The k26 - PCR and HRM results suggested that there is little size and DNA sequence variation within each Ethiopian L. donovani geographic cluster. This was confirmed by DNA sequencing of 15 amplicons (Genbank accession Nos.: JX088380 - JX088392, JX294866, JX294867) from samples belonging to the four Ethiopian clusters. Analysis of the amino acid sequences (Figure 5) showed that the HASPB repeat region for each L. donovani group in Ethiopia is comprised of two motifs, A and B, 14 and 13 amino acids long respectively. These motifs are further distinguished by the amino acids GHTQK and DHAH present in the central region of each peptide (shown in italics). Two peptides, A3 (PKEDGHTQKNDGDG) and B2 (PKEDDHAHNDGGG), comprise 81% of the peptides found in the repeat region, and represent 62.5 and 92.3%, respectively, of the A (Figure 5, yellow) and B (Figure 5, blue) motifs observed in the Ethiopian strains. Several amino acid substitutions, primarily at positions 5, 12–14 of peptide A3 or positions 3 & 12 of peptide B2, also occur in each of the motifs (Figure 5A). As expected, the number of repeats correlates with the size of the PCR amplicon (Figure 5B), however the organization of the peptide repeats is different for each cluster, and doesn't appear to be due to simple DNA duplication or deletion. The order of the peptide motifs observed for each of the Ethiopian cluster can be thought of as a bar code specific for that region.
Kenyan and Indian L. donovani strains produce larger k26 -PCR amplicons than the Ethiopian strains (this study and [30], [31]). As such it was interesting to sequence these products and determined the peptide composition and organization of the HASPB repeat region (Genbank accession No.: JX294868–JX294870). This region in the Kenyan and Indian L. donovani strains is also comprised of the same peptide motifs, A and B, found in the Ethiopian strains. Several amino acid substitutions (A0, a10, A23, A24 and B22), not observed in HASPB of the Ethiopian strains, are found in these parasites (Figure 5), but A3 and B2 still comprise a majority of the sequences observed. Together, these two peptides comprise 75 and 66.6% of the sequences found in the Kenyan and Indian strains, respectively. The combined percentage of peptides A3 and B2 for the Indian L. donovani strain described here is similar to that reported for other Indian isolates, 59.7% [29], even though additional peptide sequences, not observed in our study, were found in the latter isolates (Table S2). However, if the motif A (yellow) or B (blue), rather than the specific peptide sequence, is examined, then a similarity in organization of the repeats, ABBABBB, in the Kenyan and Ethiopian-450 k26 clusters is readily apparent.
The repeat region of the L. chagasi (syn = L. infantum) HASPB gene was previously characterized and cloned; and has been used in serological assays for VL with mixed results [19], [21], [22], [23], [24], [25], [26]. Sequences for L. infantum strains from Brazil, France, Greece, Iran, and Spain (Genbank accession Nos.: AF131228.1, EF504256.1, EF504255.1, EF504258.1, EF504257.1, DQ192034.1, and FR796455.1) show that the HASPB repeat region is only comprised of 14 amino acid peptide repeats. Two peptides A3 (PKEDGHTQKNDGDG) and a10 (PKEDGRTQKNDGDG) comprise a majority of the L. infantum k26 repeats. Peptide A3 is identical to the peptide found in the L. donovani repeat region, while peptide a10 only differs from peptide A3 by substitution of arginine for histidine at position 6 (underlined), and should be considered a member of the peptide A archetype family. However, the latter peptide, a10, does not appear to be very common in East African L. donovani, appearing only once among all the parasites examined to date. Conversely, the L. donovani 13 amino acid peptide B archetype family, exemplified by PKEDDHAHNDGGG (peptide B2) and other B peptides (Figure 5 and Table S2), was not present in any of the seven L. infantum sequences examined above, as well as six additional strains from Israel (data not shown). However, peptide B8 (Table S2) belonging to the B family archetype appears once in a L. infantum strain previously analyzed [29]. HASPB repeat region in fifteen L. infantum/L. chagasi strains contained almost exclusively peptides belonging to the A family archetype. The organization of peptide motifs was very similar for all the L. infantum strains where sequence data was available (Figure 5). However, most of the isolates analyzed belong to clusters 1a and 1b [30] which both give 626 bp amplicons by k26 - PCR.
The HASPB repeat region of L. donovani and L. infantum strains is predicted to contain multiple linear B-cell epitopes using two different programs (Figure 6, and data not shown [35], [36]). Most of the predicted epitopes (16 amino acids long, threshold ≥ 0.8 out of 1.0) in the L. donovani k26 clusters (East Africa and India) span motif junctions (A|A, A|B, B|A or B|B, 84%) with a unique L. donovani sequence, K/HNDGD/GG | PKEDDHAHND, accounting for 32/50 (64%) of these epitopes (Figure 6). This sequence is even more predominant, 80–100% of the predicted epitopes, in the southern Ethiopian, Kenyan and Indian L. donovani k26 clusters which contain multiple B motifs. This epitope is not seen in the L. infantum k26 repeat region, as the B motif is rarely observed in this species. Instead most of the predicted B-cell epitopes, 75%, contain the complete 14 amino acid A motifs, with only a few centered at the A | A motif junctions. Several of the predicted L. infantum B-cell epitopes are also found in L. donovani.
In this study we examined 63 recent strains isolated from Ethiopian VL patients in different regions of the country. All the parasites were shown to be L. donovani by three techniques, confirming previous findings that this species, not L. infantum, is responsible for VL in Ethiopia. Interestingly, we found that parasites from northwestern and southern Ethiopia could be easily distinguished based on the size of the k26 – PCR amplicons or their corresponding HRM curves. A similar clustering into two major populations by geographic origin was first reported using multiple microsatellite markers that grouped Sudanese and northwestern (Metema, Humera and Belessa) strains separately from Kenyan and southern strains (Negele-Borena and Konso) [15]. Clustering into genetically separate populations is perhaps, expected, since the primary sand fly vectors, P. orientalis and P. martini, and habitats are different for the two regions. Other differences between parasites isolated from patients in these two regions, such as sensitivity to paromomycin, have been reported [38]. Interestingly, parasites from northwestern Ethiopia could be divided into additional groups based on the k26 amplicon size, 290 bp and 410 bp. All the Sudanese parasites examined so far gave PCR products similar in size to parasites from northwestern Ethiopia (this study, [13], [30], and data not shown). While the k26-290 group contained isolates from both VL (n = 25) and HIV-VL co-infected (n = 11) patients, the k26-410 group only contained strains from HIV-VL co-infected patients (n = 3). Of the latter isolates, 3/4, were previously analyzed using microsatellite markers [15], and belong to subpopulation B2. Interestingly, this subpopulation was postulated to represent one parent strain of putative hybrid/mixed genotypes.
Different k26 – PCR products were also found when parasite strains from southern Ethiopia were analyzed, k26 −290, −360 and −450. All of the strains examined except three (AM422, AM452, and AM553) produced a 450 bp amplicon. Since microsatellite analysis grouped southern Ethiopian and some Kenyan parasites together [15], and P. martini is the primary vector involved in the L. donovani transmission in these regions [9], we decided to examine several Kenyan strains by k26 – PCR. Surprisingly, the k26 amplicons for all the Kenyan parasites tested were larger (∼500 and ∼650 bp) than those found for the south Ethiopian isolates, and similar in size to Indian L. donovani parasites (this study and [30]). Thus, there doesn't seem to be a direct correlation between the size of the k26 amplicon, and the microsatellite cluster to which the strain belongs. It is not clear whether the two southern Ethiopian strains that gave the 290 bp PCR product represent a third group present in this region, are a result of human migration or are due to culture contamination. The k26 DNA sequence for these strains is identical to the other 290 bp Sudanese and northern Ethiopian strains examined (Table S1, and data not shown). Interestingly, one of the strains, AM422, originates from the Omo Valley where transmission by both vectors may occur, and is close to Sudan. More work is needed to determine whether there is a direct correlation between the parasite vector and k26 genotype, as HASPB plays a role in parasite differentiation and localization in the sand fly [20]. At this time it is not clear why L. donovani strains from different regions in East Africa show variations in the k26 – PCR fragment size, or the factors responsible for the size polymorphism, however this technique appears to be useful for rapid mapping of strain origin on a large scale.
The HASPB1 protein is a potential vaccine candidate, as well as a diagnostic antigen [19], [21], [22], [23], [24], [25], [26], [27], [28], [29], [39], [40]. However, serodiagnostic assays using the HASPB1 protein or k26 repeat region as antigen have produced conflicting results. While assays using sera from canine or human VL caused by L. infantum give consistently high sensitivity (94–100%) and specificity (100%) [23], [24], [39], similar assays using VL sera from patients in India and Sudan showed variable sensitivity (India −21.3 and 38%; Sudan −92 and 93.5%) [21], [22], [25], [26]. Assay specificity in latter studies was consistently high (80–100%). Interestingly, the assays showing low sensitivity in Indian VL patients used the L. infantum k26 antigen [21], [22], while assays demonstrating high sensitivity in Sudanese VL patients used the L. donovani antigen [25], [26]. The B-cell epitopes recognized by serum antibodies in the HASPB1 repeat region have not been extensively analyzed, though one study reported that the 17 amino acid peptide, GDGPKEDGRTQKNDGDG from L. infantum reacted strongest with canine VL sera [41]. Interestingly, when putative linear B-cell epitopes in the L. infantum k26 repeat region were predicted (Figure 6) using a recurrent artificial neural network (ABCpred server [36]) a peptide, DGPKEDGRTQKNDGDG, 16 amino acids in length, and identical in 16/17 amino acid residues to the peptide recognized by canine sera above, ranked first with a score of 0.88 out of 1.0. This peptide includes the 14 amino acid motif (a10 – PKEDGRTQKNDGDG) frequently found in L. infantum (Figures 5 and 6), but rarely in L. donovani strains (this study and [29]). The a10 motif was predicted to be a B-cell epitope (score = 0.81). On the other hand, none of the peptide motifs (B2, B4 and B22; PKE/DDDHAHNDGG/DG) unique to L. donovani rk26 are found in L. infantum, and combinations of these motifs generated L. donovani B-cell epitopes giving the highest scores (e.g., KNDGDGPKEDDHAHND, 0.88; HNDGGGPKEDDHAHND, 0.87; HNDGDGPKEDDHAHND, 0.87; and data not shown). It will be interesting to see if better sensitivity and specificity can be obtained using either single antigen or mixtures of recombinant k26 antigens produced from the L. donovani strains responsible for local disease in Ethiopia and Sudan. This work is in progress.
HASPB1 is differentially expressed by metacyclic promastigotes and intracellular amastigotes [42]. Immunization of BALB/c mice with L. donovani HASPB1, even in the absence of adjuvant, generates a protective CD8+ T-cell response via an immune complex-mediated complement activation involving natural antibodies against a challenge with this parasite [27], [28]. The CD8+ T-cell epitopes were shown to reside in both the conserved and repeat regions of the protein [29]. While a role for HASPB in the development of metacyclic promastigotes was demonstrated [20], the function of these proteins in amastigotes is not yet clear. Interestingly, an orthologous protein, O-HASP, from L. (Viannia) braziliensis showed considerable genetic polymorphism in the repeat region among clones isolated from individual patients [43], and it was postulated that genetic variation may play a role in immune recognition. A similar phenomenon appears to occur in Old World Leishmania causing VL, as one report suggests that clonal variation is present in HASPB of Indian L. donovani strains [29]. However, DNA sequencing of 21 clones from four Ethiopian strains (k26-290 bp) did not identify any polymorphism in the repeat region of this protein (data not shown).
In summary, we show that the number, order and arrangement of the L. donovani k26 repeat region of the HASPB protein varies among strains from different geographic regions, and that the repeat motifs are different from those observed for L. infantum. The role that this genetic variation plays in the interaction with the host and vector is not clear and should be investigated further.
|
10.1371/journal.pntd.0007017 | Naja annulifera Snake: New insights into the venom components and pathogenesis of envenomation | Naja annulifera is a medically important venomous snake occurring in some of the countries in Sub-Saharan Africa. Accidental bites result in severe coagulation disturbances, systemic inflammation and heart damage, as reported in dogs, and death, by respiratory arrest, in humans. Despite the medical importance of N. annulifera, little is known about its venom composition and the pathogenesis of envenomation. In this paper, the toxic, inflammatory and immunogenic properties of N. annulifera venom were analyzed.
Venom proteomic analysis identified 79 different proteins, including Three Finger Toxins, Cysteine Rich Secretory Proteins, Metalloproteinases, Phospholipases A2 (PLA2), Hyaluronidase, L-amino-acid oxidase, Cobra Venom Factor and Serine Proteinase. The presence of PLA2, hyaluronidase, fibrinogenolytic and anticoagulant activities was detected using functional assays. The venom was cytotoxic to human keratinocytes. In an experimental murine model of envenomation, it was found that the venom induced local changes, such as swelling, which was controlled by anti-inflammatory drugs. Moreover, the venom caused death, which was preceded by systemic inflammation and pulmonary hemorrhage. The venom was shown to be immunogenic, inducing a strong humoral immune response, with the production of antibodies able to recognize venom components with high molecular weight and to neutralize its lethal activity.
The results obtained in this study demonstrate that N. annulifera venom contains toxins able to induce local and systemic inflammation, which can contribute to lung damage and death. Moreover, the venom is immunogenic, an important feature that must be considered during the production of a therapeutic anti-N. annulifera antivenom.
| N. annulifera is a dangerous snake that belongs to the Elapidae family. It is found in some of the countries in Sub-Saharan Africa and has caused accidents in humans and dogs. In this study, we characterized some of the biochemical, toxic and immunogenic properties of N. annulifera venom. We showed that the venom is composed of several proteins, some of which display enzymatic activities, such as phospholipase A2, hyaluronidase, metalloproteinases and serine proteinases. The venom promoted disturbances in the human coagulation system and was cytotoxic to human epidermal cells. Using a mouse model, we showed that the venom promotes local reactions that were reduced with anti-inflammatory drugs. The venom caused systemic inflammation, lung hemorrhage and death. Further, the venom stimulated production of high antibody titers when injected into mice and the antiserum produced was able to inhibit venom-induced death. This study demonstrated that N. annulifera venom contains toxins that trigger inflammatory process, which may contribute to the envenomation pathology. Moreover, the venom is immunogenic, an important aspect for the production of an efficient N. annulifera antivenom.
| Envenoming from snakebites is a public health problem in rural areas of the tropical and subtropical countries in Africa, Latin America, Asia and Oceania [1, 2]. This medical condition kills more than 95,000 people per year and leads over 300,000 victims to live with permanent sequelae [3].
It is also estimated that there are approximately 3,700 species of snakes worldwide [4]. Of these, approximately 15% are venomous and have caused serious accidents involving humans and other animals [5]. Venomous snakes belong to the Colubroidea superfamily, which is composed of several families, such as Colubridae, Viperidae, Lamprophiidae and Elapidae [6, 7]. The Elapidae family consists of 61 genera and includes 365 species [8], which are distributed in the tropical and temperate regions of Africa, America, Asia and Australia. These snakes can live in terrestrial or aquatic environments and present variable diet, including small vertebrates, such as birds, rodents, reptiles and fishes or invertebrates [9, 10].
The venom of Elapidae is well known to contain powerful neurotoxins that play a role in the snake defense against predators and prey capture. These neurotoxins may also be responsible for some of the clinical manifestations observed in human envenomation, such as respiratory arrest [11, 12]. However, several studies have noted the presence of different components in these venoms, including phospholipases A2 (PLA2), hyaluronidases (HYA) [13, 14, 15], metallo- (SVMP) and serine proteinases (SVSP) [16, 17, 18], inhibitors [19], peptides [20] and cytotoxins [21, 22]. All of these components may have cytotoxic [14, 23, 24], hemorrhagic, anticoagulant [25], pro-inflammatory [26, 27] or immunogenic [28] properties.
The genus Naja presents specimens that cause a large and serious number of accidents [3, 8, 23, 29]. One dangerous representative of this genus is Naja annulifera, which is popularly called the Banded or Snouted Cobra [30, 31]. It is found in Sub-Saharan Africa countries, including Zambia, Malawi, Mozambique, Swaziland, Zimbabwe, Botswana and South Africa, in savannah grasslands, deserts, and rocky areas and near human habitations. This ranging of habits is associated to its broad diet, once it feeds on frogs, lizards, birds and it eggs, snakes and rats.
Accidents involving N. annulifera are considered severe. The envenomed individuals experience swelling, pain and local burning at the site of the bite, followed by pain throughout their entire body. Beside these clinical findings, affected individuals can present with dizziness and palpebral ptosis. Some can progress to respiratory arrest and, without a specific treatment, death. The treatment for envenomated individuals is serum therapy and in respiratory arrest cases, mechanical ventilation. Some studies have also reported that envenomed individuals in South Africa may develop necrosis at the site of the bite as well as hematologic disturbances [23, 31, 32, 33].
Veterinary epidemiologic data demonstrated that approximately 60% of dogs poisoned by snakebites in South Africa were bitten by N. annulifera. These dogs presented with various clinical findings, including hematologic alterations, such as leukocytosis and thrombocytopenia [34], increased plasma levels of Cardiac Troponin I and C Reactive Protein (CRP) [35] and disturbances in the coagulation system [36].
Information about the components of N. annulifera venom is scarce. Some authors have noted the presence of several cyto/cardiotoxins and neurotoxins [37, 38, 39, 40, 41, 42]. Despite its medical importance, epidemiologic, clinical and experimental studies of N. annulifera venom are limited and the mechanisms by which it causes toxicity remains poorly understood. The goal of the present study was therefore to describe the components of N. annulifera venom and its toxic activities, utilizing in vitro and in vivo models to better understand the pathology of envenomation by this snake.
Bovine serum albumin (BSA), Concanavalin A from Canavalia ensiforms (ConA), Wheat Germ Agglutinin from Triticutun vulgaris (WGA), Tween 20, Triton X-100, 3’3’-Diaminobenzidine (DAB), Hyaluronic Acid, anti-mouse IgG horseradish labeled with Peroxidase (IgG-HRPO), Ortho-phenylenediamine (OPD), cetyltrimethylammonium bromide (CTAB), Gelatin, Comassie Brilliant Blue R-250, Human Fibrinogen (Fb), Human Thrombin, 1,10 Phenantroline (1,10 Phe), Phenylmethasulfonyl Fluoride (PMSF), Dimethyl Sulfoxide (DMSO), Cromolyn, Dexamethasone, Trypsin, Solid Phase Extraction Disks (SDB-XC membranes), Hepes, urea, iodoacetamide and trifluoroacetic acid (TFA) were purchased from Sigma-Aldrich (Missouri, USA). The Bicinchoninic Acid (BCA) Protein Assay Kit was purchased from Pierce Biotechnology, Inc. (Wisconsin, USA). SDS-PAGE (Amersham ECL Gel 8–16%, 10 wells) gels were obtained from GE Healthcare Life Sciences (Uppsala, Sweden). Dithiothreitol (DTT) was from Calbiochem (Darmstadt, Germany). The nitrocellulose membrane and EnzChek Phospholipase A2 Assay Kit were purchased from Thermo Fisher Scientific (Massachusetts, USA). Aluminum Hydroxide (Al(OH)3) was purchased from Prati Donaduzzi (Sao Paulo, Brazil). Anti-mouse IgG labeled with alkaline phosphatase (IgG-AP), Nitroblue Tetrazolium chloride (NBT), 5-bromo-4-chloro-3-indolyl-phosphatase (BCIP) and CytoTox 96 Non-Radioactive Cytotoxicity Assay Kit were obtained from Promega Corp. (Madison, Wisconsin, USA). ELISA plates were purchased from Costar Corning, Inc. (New York, USA). Calcium chloride, Cephalin and Thromboplastin were obtained from Stago (Saint-Quen-I’Aumône, France). Indomethacin, MK-886 (Sodium Salt) and WEB-2086 were purchased from Cayman Chemical (Michigan, USA). The BD Cytometric Bead Array (CBA) Mouse Inflammation and Mouse IL-1β ELISA kits were obtained from BD Biosciences (New Jersey, USA). The mouse IL-17 DuoSet3.2.4.3 ELISA kit was obtained from R&D Systems (Minnesota, USA). The Panótico Rápido kit was purchased from Laborclin (Parana, Brazil). The HaCat human keratinocyte lineage was obtained from the Rio de Janeiro Cell Bank (Rio de Janeiro, Brazil). Dulbecco's Modified Medium Eagle—Gibco (DMEM) medium, penicillin and streptomycin were purchased from Invitrogen Corp. (California, USA). Fetal Bovine Serum was obtained from Cultilab (São Paulo, Brazil). 3-(4,5-dimethylthiazol-2-yl)2,5-di-phenyltetrazolium bromide (MTT) was obtained from Merk (Darmstadt, Germany). Frits for SPE cartridges were obtained from Agilent (California, USA) 10 μm Jupiter C-18 beads were purchased from Phenomenex (Torrance, USA). 3 μm ReproSil-Pur C-18 beads were obtained from Dr. Maisch (Ammerbuch, Germany). 75 μm I.D. or 100 μm I.D. x 360 μm O.D. polyimide coated capillary tubing were purchased from Molex (Lisle, USA).
N. annulifera venom (South Africa specimens) was purchased from Latoxan Natural Active Ingredients (Valence, France). The lyophilized venom was reconstituted in sterile saline solution at 5 mg/mL. The protein content was assessed with the BCA Protein Assay Kit according to the manufacturer’s recommendations. The samples were aliquoted and stored at -80°C until use. Bothrops jararaca, Crotalus durissus terrificus and Tityus serrulatus venoms were supplied by Butantan Institute, SP, Brazil, and their use was approved by the Brazilian Institute of Environment and Renewable Resources (IBAMA), an enforcement agency of the Brazilian Ministry of the Environment (protocol number: 010035/2015-0), and by SisGen (Sistema Nacional do Patrimônio Genético e do Conhecimento Tradicional Associado (protocol numbers AD50761 and AEE9AEA).
HighIII (HIII) female mice weighing 18–22 g were obtained from the Immunogenetics laboratory, while Balb/c male mice weighing 18–22 g were obtained from the Center for Animal Breeding, both from Butantan Institute. All procedures involving animals were in accordance with the ethical principles for animal research adopted by the Brazilian Society of Animal Science and the National Brazilian Legislation n°.11.794/08. The protocols used in the present study were approved by the Institutional Animal Care and Use Committee of the Butantan Institute (protocols approved n° 01092/13 and 01262/14).
Experiments using samples obtained from humans were previously approved by the Human Research Ethics Committee of the Municipal Health Secretary of São Paulo. Human blood samples were obtained from healthy donors who knew of the purposes of this study and signed the corresponding informed consent form (protocol approved n° 974.312).
Blood samples from healthy donors were placed into tubes containing sodium citrate (3.2%) and centrifuged at 260× g at room temperature to obtain platelet-poor plasma (PPP). The PPP samples were aliquoted and stored at -20°C until their use.
Statistical analysis was performed using one-way ANOVA followed by Tukey’s Multiple comparison test, two-way ANOVA followed by a Bonferroni multiple comparison or t-tests. All statistical analyses were performed using Graphpad Prism 5 software (La Jolla, California, USA). Differences were considered significant when p ≤ 0.05.
The analysis of the electrophoretic profile of N. annulifera venom using a 8–16% SDS-polyacrylamide gel, under non-reducing conditions, revealed protein bands with relative molecular masses between ~10 kDa and ~190 kDa. Under reducing conditions, the venom displayed a more simple profile, with main protein bands ranging from ~6 kDa to ~100 kDa (Fig 1A), indicating that the venom contains protein in oligomeric state. To better characterize venom proteins, nitrocellulose membranes to which venom samples had been transferred were incubated with the ConA and WGA lectins, which recognize mannose [64] and N-acetylglucosamine [65] residues, respectively. Fig 1B shows three protein bands (~70, ~86 and ~103 kDa) that interacted with WGA indicating the presence of N-acetylglucosamine. Moreover, three bands that interacted with ConA (~63, ~81 and ~170 kDa) indicated that these proteins contain mannose residues in their carbohydrate moieties (Fig 1C).
The venom proteomic analysis using in-solution trypsin digestion and LC-MS/MS resulted in the identification of 79 proteins, including 24 types of venom components with or without enzymatic function, and six proteins of unknown function. Despite the variety of proteins identified in the venom, it is dominated by the presence of 3FTx (27 proteins) (Table 1; S1 and S2 Tables; S1 Spectra). Other protein classes identified in the venom include cysteine-rich secretory protein (CRISP), Kunitz-type protease inhibitor, snake venom metalloproteinase (SVMP), snake venom serine proteinase (SVSP), phospholipase A2 (PLA2), hyaluronidase (glycosyl hydrolase 56 family) and Cobra Venom Factor (venom complement C3 homolog).
N. annulifera venom showed no proteolytic activity on zymography using gelatin as substrate. Under the same experimental conditions, proteins from the positive control of B. jararaca venom showed gelatinolytic activity, as demonstrated by clear regions in the gel (bands with molecular mass above 60 kDa) (Fig 2A). However, when proteolytic activity on fibrinogen was assessed, it was observed that the venom contained proteinases that were able to cleave this protein at the alpha chain, generating a fragment with a molecular mass of ~ 40 kDa (Fig 2B). In addition, when inhibitors were added to the reactions, cleavage was inhibited by 1, 10 Phe and PMSF, demonstrating the contributions of SVMP and SVSP to this hydrolysis (Fig 2B).
The hyaluronidase activity of N. annulifera snake venom was assessed via a turbidimetric assay, and the results are expressed in UTR. It was demonstrated that N. annulifera venom has a low hyaluronidase activity (UTR = 14.4) compared to the positive control T. serrulatus scorpion venom (UTR = 52.8) (Fig 2C).
PLA2 activity was measured with a fluorimetric assay, and the results obtained are expressed as specific activity (SA: UF/min/μg venom). Fig 2D shows that N. annulifera snake venom has significantly lower phospholipase activity (SA: 566.2) than the C. d. terrificus venom positive controls (SA: 3061.3).
Cytotoxic properties of the venom were evaluated by cell viability and LHD release assays. N. annulifera venom was able to reduce human keratinocyte viability in a dose-dependent manner but was not able to promote LDH release by the cells (Fig 2E).
N. annulifera venom induced disturbances in the Activated Partial Thromboplastin Time in a dose dependent manner (Table 2). Platelet-poor plasma (PPP) samples were incoagulable when the highest venom concentrations were used (12.5 to 50 μg), while the coagulation time (seconds) was significantly prolonged at lower concentrations (2.5 to 6.25 μg) (p≤ 0.05). Moreover, the R-time shows that the disturbances promoted by N. annulifera venom were severe, as they were much higher than normal. Prothrombin Time assays demonstrated that the venom led to a significant prolongation in the coagulation time of PPP at the highest venom concentrations, i.e., 25 and 50 μg. Besides that, both venom concentrations were able to cause a significant alteration in the R-time (Table 2).
LD50 was determined 72 hours after venom administration. Prior to their death, animals presented with several clinical findings of envenomation, including apathy, bending at the column, a rough hair coat, dyspnea and paralysis of their hind limbs. LD50 was calculated with a probit analysis with a 95% confidence interval and was found to be 94.14 μg (68.78–115.74).
The systemic reactions promoted by the venom were evaluated using two different experimental protocols. By injecting 60% (56.48 μg) (Sublethal dose) of LD50 of N. annulifera venom, it was observed that animals presented several clinical findings of envenomation, including apathy, bending at the column, a rough hair coat, dyspnea and difficulty walking. This dose was also able to cause alterations in some systemic parameters, such as a decrease of circulating lymphocytes [Fig 4A] and an increased number of circulating neutrophils [Fig 4B]. Nonetheless, this dose was not able to lead to changes in the total number of leukocytes and histological organs damage. However, a sublethal dose promoted an increase in the plasma levels of MCP-1 [Fig 4C] and IL-6 [Fig 4D]. These alterations persisted for several hours. However, 24 hours after experimental envenomation, all of these parameters returned to their normal values.
Injection of 2LD50 led to several clinical findings of envenomation, such as apathy, bending at the column, a rough hair coat, dyspnea, hind limb paralysis and death. In addition to these clinical findings, histopathologic alterations were observed in the lungs, among these, moderate vascular congestion and multiple hemorrhagic foci [Fig 4I]. Furthermore, the animals demonstrated leukocytosis [Fig 4E] characterized by neutrophilia and monocytosis [Fig 4F], as well as increased plasma levels of MCP-1 [Fig 4G] and IL-6 [Fig 4H].
The immunogenicity of N. annulifera snake venom was evaluated in HIII mice immunized with 10 μg of venom. The antibody titers were determined by ELISA. Fig 5A presents the antibody response over time. N. annulifera venom is highly immunogenic, promoting the production of high antibody titers, already after the second immunization dose. The recognition profile of the venom proteins by the experimental N. annulifera mouse antiserum was determined by Western Blot. Fig 5B shows that the experimental serum was able to recognize venom components, mainly proteins with Mr above 50 kDa. Moreover, as determined by Probit analysis, it was shown that this antivenom was able to neutralize the N. annulifera venom lethal activity with a high potency, i.e., 1 ml of antivenom was able to neutralize 4.5 mg of venom.
In this study, we characterized some of the biochemical, toxic, immunogenic and physiopathologic properties of the venom from N. annulifera, which is a medically important snake related to accidental bites in the countries of Sub-Saharan Africa. The results show that N. annulifera venom contains several toxic components able to induce systemic inflammation, which may contribute to the pathology observed in envenomed individuals. Moreover, the venom is immunogenic, an important feature that must be considered during the production of a therapeutic anti-N. annulifera antivenom.
Electrophoretic analysis of N. annulifera venom showed that it contains several components, including low molecular mass proteins, which suggested the presence of neurotoxins [18], among these, 3FTx [14, 15] and PLA2 [16]. The presence of these components in the venom was confirmed by trypsin digestion of proteins and LC-MS/MS analysis. These components may be responsible for some of the clinical findings observed during envenomation, such as heart damage and systemic inflammation in dogs [35], as well as respiratory arrest in humans [23]. The data from this study corroborate the results of other studies, which showed the presence of these components in N. annulifera venom [37, 38, 39, 40, 41, 42]. Envenomed individuals from South Africa showed local dermonecrotic injury after a bite, which could be caused by the high content of cytotoxins, as shown in our LC-MS/MS analysis. In accordance with data from Panagides and colleagues [66], here we also showed that N. annulifera venom could promote decrease in human epidermal cells viability, as evaluated by the MTT method. However, it was not possible to detect release of LDH by these cells, which possibly indicate that the keratinocyte membranes were not damaged and that the cell death promoted by N. annulifera could be due to apoptosis [67] and not necrosis. Alternatively, it is possible to consider that N. annulifera venom contains components able to reduce mitochondrial activity, since MTT method evaluates cell viability as enzymatic conversion of the tetrazolium compound to water insoluble formazan crystals by dehydrogenases occurring in the mitochondria of living cells.
Western Blot lectin analysis demonstrated the presence of mannose and N-acetylglucosamine residues in N. annulifera venom proteins. These carbohydrate residues were found in the venom proteins of different genera of snakes [46, 68, 69, 70], including in important toxic components, such as SVSP and SVMP [69, 70]. Although we have not identified the families of these glycosylated components, it is possible that some of them are linked to SVMPs or SVSPs, as they have predicted molecular masses similar to some of the previously described glycosylated proteolytic enzymes from viperid venoms [69, 70]. Moreover, SVMPs and SVSPs were detected in N. annulifera venom by LC-MS/MS. Proteins containing these carbohydrate residues are also present in many pathogens, such as bacteria and fungi, and can be recognized by different immune cells and molecules, triggering inflammatory and immune responses [71, 72]. This recognition could contribute to the clinical manifestations observed during envenomation, such as the systemic inflammation observed in dogs [35].
Functional biochemical assays were performed to confirm the presence of some of the toxic-enzymatic components found in the LC-MS/MS analysis. The presence of proteinases in animal venoms can contribute to different clinical manifestations during envenomation, such as inflammation, tissue damage, disturbances in coagulation and bleeding [73]. In elapid venoms, proteolytic activity is usually low or nonexistent [26, 74, 75]. In contrast to the data shown by Phillips et al. [76], under the experimental conditions used in this paper, N. annulifera venom did not show any proteolytic activity on either gelatin. However, although detected at low abundance in the proteomic analysis, the venom contains SVMP and SVSP able to cleave the fibrinogen alpha chain, which suggests that these proteinases can contribute to the hemostatic alterations as observed in dogs envenomated by N. annulifera [36].
Hyaluronic acid is cellular cement that, together with other components of the extracellular matrix, forms a protective gel that prevents the entry of foreign agents. A variety of animal venoms contain hyaluronidases, which cleave hyaluronic acid molecules to facilitate access of the venom from the tissue to the bloodstream. This enzyme is also called “spreading factor” [77], and its action can promote local and systemic inflammation by increasing tissue permeability [78]. Moreover, the products derived from hyaluronic acid cleavage, which can be recognized by immune receptors, such as TLR2 and 4, trigger the production of inflammatory mediators, such as cytokines and chemokines [79]. N. annulifera venom showed evidence of hyaluronidase activity, but it was relatively low compared with other elapid venoms from the Naja [74] and Micrurus [16] genera, and accordingly, only one hyaluronidase was identified by mass spectrometry in the venom.
PLA2 belongs to a superfamily of lipolytic enzymes that catalyze specific hydrolysis of the ester linkage at the sn-2 position of glycerophospholipids, generating arachidonic acid and lysophospholipids. PLA2-like proteins found in snake venoms may be devoid of catalytic activity, although it may exhibit myotoxic or neurotoxic activities [80, 81]. Moreover, these enzymes may present with other toxic properties during envenomation, such as the ability to cause cytotoxicity and inflammation [82, 83]. As predicted by LC-MS/MS, the presence of PLA2 activity in N. annulifera venom was also observed in enzymatic assays. Nonetheless, N. annulifera venom exhibits a very lower content of PLA2 /enzymatic activity when compared to other Naja venoms [84, 85, 86, 87, 88], suggesting a strong interspecific variation associated to this particular toxin.
As observed in dogs, we showed here that N. annulifera venom promoted hemostatic disturbances in human plasma, making it incoagulable. These disturbances in the hemostatic system can be attributed to the fibrinogenolytic proteinases detected in the venom, or to PLA2 found in the LC-MS/MS analysis, since these components can promote plasma incoagulability via direct binding to FXa, thereby preventing thrombin generation [89, 90, 91, 92]. This phenomenon, promoted by several species of Naja venom, has been observed in different clinical and experimental studies [93, 94]. It is therefore very important to evaluate the mechanisms involved in these alterations and their consequence, since the current literature on the topic is scarce. In addition, these alterations may be a therapeutic target for Naja envenomation.
In our in vivo experimental model, the venom was able to induce swelling and several histopathologic changes in the hind paws of mice, that decreased only after 24 hours. Among these tissue alterations, myonecrosis associated with inflammation was observed, an event that is commonly found in experimental models of Elapidae envenomation, which is attributed to cytotoxins and PLA2 [95, 96].
The inflammatory events promoted by the venom may be attributed to hyaluronidase, PLA2, glycosylated proteins, SVSP and SVMP since they can promote different inflammatory events, which include complement activation, mast cell degranulation, release of eicosanoids and cytokines and leukocyte homing [16, 26, 27, 58, 59, 60, 61, 83]. Knowing that N. annulifera venom promotes inflammation and pain in humans and dogs [31, 32, 33, 34, 35, 36], which were also observed in our experimental model, pharmacologic studies were performed to analyze the role of some inflammatory mediators in the edema process. Pre-treating mice with different compounds that were able to modulate different steps of the inflammatory process, including, mast cell degranulation (Cromolyn), lipid mediator production (Dexamethasone, Indomethacin and MK-886) and action (WEB-2086) significantly decreased the edema promoted by N. annulifera venom. All of the compounds showed a similar pattern of inhibiting peak edema. However, cPLA2, COX isoforms and FLAP inhibitors controlled the edema for a long time, suggesting a stronger contribution of eicosanoids, such as prostaglandins, thromboxanes and leukotrienes, to this process. Moreover, it is possible that these lipid mediators may contribute to the pain observed in humans after a bite, making them good therapeutic targets for the local reactions promoted by N. annulifera venom.
The LD50 of N. annulifera, established here via i.p. route in Balb/c mice, was 94.14 μg. Ramos-Cerrillo and collaborators [56] showed that the N. annulifera venom LD50 when administered intravenously was 53.9 μg. To evaluate whether N. annulifera venom was able to promote systemic changes, a sublethal dose of the venom was established, and different inflammatory parameters were evaluated. As in dogs envenomed by N. annulifera [34], it was observed that a sublethal dose of the venom promotes acute systemic inflammation, which was characterized by neutrophilia and increased levels of IL-6 and MCP-1 in the plasma. However, unlike dogs envenomed by N. annulifera the sublethal dose was not able to cause organ injury, as observed by Langhorn et al. [35] in dogs.
By administering a superdose (2LD50) of venom to mice, it was possible to observe systemic inflammation, which was characterized by an increase in the plasma levels of IL-6 and MCP-1. However, in contrast with the sublethal dose, the 2DL50 dose caused leukocytosis, which was characterized by neutrophilia and monocytosis. In addition, dead animals showed multifocal hemorrhaging in their lungs.
These data suggest that the systemic inflammatory process induced by high doses of venom may be associated with the lung alterations observed in humans. In fact, in different models of hemorrhagic shock, plasma, pulmonary and hepatic increases in IL-6 and MCP-1 were observed along with inflammation and lung injury, which may culminate in acute respiratory distress syndrome [97, 98, 99]. It is important to emphasize that in addition to cytokines, some other factors may be associated with the pulmonary hemorrhaging and death caused by a venom overdose. Further, the hemostatic alterations promoted by SVMP, SVSP and PLA2 can also contribute to lung hemorrhage.
The treatment indicated for envenomation by N. annulifera is serum therapy. N. annulifera venom is part of the antigenic mixture used for the production of polyvalent serum by the South African Vaccine Producers (SAVP) (Pty) Ltd [28], although its immunogenicity and neutralizing potential have been poorly investigated. Here, we show that N. annulifera venom is highly immunogenic in murine model. Although with different intensities, this mouse antivenom was able to recognize venom components by Western Blot, mainly the ones with high molecular weight, which includes components as HYA, LAAO, CVF and SVMP. Moreover, this monovalent antivenom was able to protect the animals from death induced by venom, with high potency. In contrast, other studies have shown that horse antivenoms produced against venom mixtures, in which N. annulifera was included, were not able to neutralize the lethal effects of this venom [56, 100, 101]. This may be due to differences related to the animals used (mouse versus horse) or to a low level of neutralizing antibodies generated by other venoms present in the immunization pool.
In conclusion, here, we show that N. annulifera snake venom contains several components with toxic and pro-inflammatory properties. Some of these toxins promote coagulation disturbances, local and systemic inflammatory reactions, which may contribute to the pathologic events, observed in our murine model and possibly in dogs and humans envenomated by N. annulifera. Moreover, the venom promoted lung haemorrhage, an event that may also occur in cases of human envenomation, since death by respiratory arrest can be the result of the sum of neurotoxin activity and lung haemorrhage. High levels of IL-6 and MCP-1, as detected in the plasma of the envenomated animals, may be associated with pulmonary damage, since systemic inflammatory conditions can be deleterious and affect several organs, including the lungs. Thus, inflammation may be considered as target for the development of new therapeutic strategies in cases of N. annulifera human envenomation. Moreover, we showed that the venom is highly immunogenic and that the experimental serum was able to neutralize its lethal activity in the murine model. These data encourage further studies to characterize and produce monospecific therapeutic antivenom against N. annulifera.
|
10.1371/journal.pcbi.1004484 | Sensorimotor Model of Obstacle Avoidance in Echolocating Bats | Bat echolocation is an ability consisting of many subtasks such as navigation, prey detection and object recognition. Understanding the echolocation capabilities of bats comes down to isolating the minimal set of acoustic cues needed to complete each task. For some tasks, the minimal cues have already been identified. However, while a number of possible cues have been suggested, little is known about the minimal cues supporting obstacle avoidance in echolocating bats. In this paper, we propose that the Interaural Intensity Difference (IID) and travel time of the first millisecond of the echo train are sufficient cues for obstacle avoidance. We describe a simple control algorithm based on the use of these cues in combination with alternating ear positions modeled after the constant frequency bat Rhinolophus rouxii. Using spatial simulations (2D and 3D), we show that simple phonotaxis can steer a bat clear from obstacles without performing a reconstruction of the 3D layout of the scene. As such, this paper presents the first computationally explicit explanation for obstacle avoidance validated in complex simulated environments. Based on additional simulations modelling the FM bat Phyllostomus discolor, we conjecture that the proposed cues can be exploited by constant frequency (CF) bats and frequency modulated (FM) bats alike. We hypothesize that using a low level yet robust cue for obstacle avoidance allows bats to comply with the hard real-time constraints of this basic behaviour.
| Echolocating bats can fly through complex environments in complete darkness. Swift and apparently effortless obstacle avoidance is the most fundamental function supported by biosonar. Despite this, we still do not know which acoustic cues, from among the many possible cues, bats actually exploit while avoiding obstacles. In this paper, we show using spatial simulations (2D and 3D) that the Interaural Intensity Difference (IID) and travel time of the first millisecond of the echo train in combination with alternating ear positions provide robust and reliable cues for obstacle avoidance. Simulating the echoes received by a flying bat, we show that simple phonotaxis can steer a bat clear from obstacles without performing 3D reconstruction of the layout of the scene. As such, this paper presents the first computationally explicit explanation for obstacle avoidance in realistic and complex 3D environments. We hypothesize that using low level yet robust cues for obstacle avoidance allows bats to comply with the hard real-time constraints of this basic behaviour.
| Rhinolophidae are echolocating bats specialized in hunting for airborne prey among vegetation using echolocation. To cope with clutter echoes returning from vegetation they employ a unique sensorial strategy for detecting prey. They emit long narrow-band pulses and listen for frequency and amplitude shifts, so called glints, in the echoes caused by fluttering prey [1]. Echoes from stationary obstacles do not contain these glints and do not interfere with the detection and localization of prey [2].
While the sensorial adaptations of Rhinolophidae for prey detection have been extensively researched (see [1] for a review), the cues supporting the ability of these bats to navigate and orient in cluttered environments have received much less attention. Nevertheless, their ability to navigate small spaces [3–6] and their well-studied echolocation apparatus [1, 7] makes them an interesting taxon to study how echolocating bats avoid obstacles in natural environments. Indeed, as argued in the discussion, understanding the cues Rhinolophidae use to negotiate space is potentially informative about how other bats using frequency modulated pulses could avoid obstacles as well.
It would seem that Rhinolophidae, using long narrowband signals, lack both the bandwidth and the temporal resolution available to bats using short broadband signals. Indeed, bats using broadband signals typically shorten their calls (typically 1–3 ms [8]) and increase the bandwidth when moving into cluttered spaces [8]. Rhinolophidae, in contrast, negotiate cluttered space using much longer (about 10–50 ms) and narrowband signals that seem not particularly well suited for obstacle avoidance. Indeed, while Rhinolophidae also shorten their calls and increase the bandwidth when moving into cluttered space [9, 10], their calls remain longer and more bandwidth limited than those of FM bats under the same conditions.
The characteristic cyclical pinna movements shown by Rhinolophidae [11, 12] have been suggested to compensate for the lack of spatial cues available to bats relying on broadband calls. Mogdans et al. [3] performed behavioural experiments to test specifically the role of these ear movements for obstacle avoidance based on Interaural Intensity Differences (IIDs). The hypothesis [3, 10] that the moving ears generate changing IIDs encoding the reflector position in both the horizontal and the vertical plane was found by these authors to be in agreement with the results from their wire-avoidance experiments and put forward as a possible explanation for the bats’ obstacle avoidance ability. Since then, simulation studies and robotic experiments have corroborated that these ear movements do indeed provide various localization cues that would allow localizing individual reflectors, such as prey items [13–15]. However, natural environments encountered by bats are typically made up of objects that consist of many stochastic reflectors returning many overlapping echoes [16]. Therefore, for 3D localization of reflectors, e.g. based on typical ear movement induced IID patterns, to be considered a plausible mechanism underlying the obstacle avoidance abilities of bats, it has to be proven first that such a localization capability is robust in the presence of multiple overlapping echoes. Hence, while it has been shown that pinnae movements play a significant role in obstacle avoidance [3], it is still not clear what information Rhinolophidae extract from such pinna movements to allow them to avoid natural (and complex) obstacles.
To complement behavioural experiments, we use the synthetic methodology, i.e. understanding natural systems by building artefacts [17–19], computer simulations, in this case, to study bat obstacle avoidance behaviour. In particular, we propose a sensorimotor system that does not rely on the bat reconstructing the 3D spatial layout of reflectors from the echoes, but instead relies on the dynamics of the bat-obstacle interaction to result in obstacle avoidance behaviour. A similar approach is taken in ref. [20] for prey-catching behaviour in echolocating bats assuming that only a single reflecting target is present giving rise to a unique isolated echo. This assumption is warranted in the case of prey-catching behaviour as the bat can choose to hunt away from clutter [8] or take active measures to separate the echoes from the foreground prey item from the clutter background ones (e.g. [21, 22]). In contrast, realistic obstacles, e.g. foliage and/or man-made structures, will always give rise to multiple overlapping echoes [16].
The sensorimotor system we propose is intentionally kept as simple as possible. It uses IID and time delay of the first echo onset in combination with alternating pinna movements to guide the bat. In particular, it processes only the first millisecond of the echo train. Furthermore, it does not need the right and left ear echo signals to be segmented into contributions from individual reflectors, as would be required by any approach that reconstructs the spatial layout of the bat’s surroundings. While approaches that attempt to reconstruct the spatial layout of the environment first as a prerequisite for obstacle avoidance [23, 24], when successful, are clearly sufficient to explain such behaviour, we aim to show with the proposed sensorimotor system that such a reconstruction capability is not a necessary condition. The main advantage of the proposed obstacle avoidance mechanism is that because of its simplicity as well as its reliance on the first millisecond of the echo train only it can react very rapidly to the relevant information contained in an otherwise very complex echo signal consisting of many overlapping echoes. This allows the system to respond appropriately under hard real-time conditions independent of the complexity of the environment.
In this paper, we first present the environments used to simulate the echoes received by a bat moving through realistic, cluttered spaces. Next, we propose a sensorimotor system that results in obstacle avoidance behaviour by extracting echo delay and IID information from the onset of the first echo in combination with alternating pinna movements. Finally, we test the performance of the sensorimotor system in simulated 2D and 3D environments showing that despite its simplicity the system can avoid obstacles in a complex environment without the need to reconstruct the 3D spatial layout of the reflectors present.
We tested the proposed sensorimotor system both in environments that were artificially generated and in environments derived from 3D laser scans of real bat habitats. Below we discuss the construction of both types of test environments.
The intensity of the echo returning from each point reflector i was calculated for each call. The intensity gi (in dB) of the echo received from reflector i is given by the sonar equation [31],
g i = g b a t + 40 · log 10 0 . 1 r i + 2 · ( r i - 0 . 1 ) · a f + d ϕ i , p + s i + c ϕ i (1)
In Eq (1), gbat is the intensity of the call at 10 cm from the mouth, in this paper taken to be 120 dBspl [9]. The parameters ri, af, dϕi,p, si give the range to reflector i, the atmospheric absorption at frequency f [35], the directional sensitivity dϕi,p of the sonar apparatus of the bat for angle ϕi and pinnae position p (see below), and the echo strength si of the reflector respectively. Simon et al. [36] ensonified leaves for a range of aspect angles and found reflector strength to vary from −30 dB to −6 dB. Therefore, variations in aspect dependent reflector strength si were modelled by choosing the reflector strength randomly from a uniform distribution over this interval for each call.
As stated above, for the regularly spaced artificial environments mimicking the wire avoidance tests of Mogdans et al, [3] the reflector strength si was fixed at −66 dB corresponding to the target strength of a wire with a diameter of 0.16 mm [29]. In the torus environment, the reflector strength was chosen randomly from the interval −46 to −34 dB. This corresponds to −40 dB, the approximate target strength of a sphere with diameter 5 cm [31], plus and minus 6 dB.
In Eq (1), cϕi denotes an additional attenuation reflecting changes in cochlear sensitivity for different frequencies. The cochlea of Rhinolophidae is highly tuned to the species-specific constant frequency component of the call (Reviewed in [1]). While flying, these bats compensate the Doppler shift of the returning echoes by lowering the emission frequency. In doing this, they effectively ensure that echoes return with a frequency very close to the frequency their cochlea is tuned to, i.e. the reference frequency. However, the Doppler shift Δfϕi of an echo depends on the heading direction ϕi of reflector i as follows,
Δ f ϕ i = f e m i s s i o n · 2 · v b a t v s o u n d · cos ϕ i (2)
We were unable to find flight speed data for R. rouxii. However, bats weighing about 10 grams were reported to commute with a speed of 6 ms−1 [37, 38]. Therefore, we modelled the maximum speed of R. rouxii as vbat = 6 ms−1. R. ferrumequinum is capable of drastically reducing its flight speed when near an obstacle. Aldridge [4] reports a flight speed of about 0.3 ms−1 at the maximum turning rate for R. ferrumequinum. Moreover, this bat starts reducing its speed from about 5 meters before landing [9]. Hence, we model the flight speed of R. rouxii as 0.3 ms−1 and 6 ms−1 at 0 and 5 meter (and more) from the nearest obstacle respectively (See Fig 1a). We interpolate linearly between these points. Notice that this implies that the simulated flight speed in the regularly spaced artificial environments (see below) where obstacles are spaced 15 cm apart is maximally about 0.47 ms−1.
The details of how Rhinolophidae lower their emission frequency when faced with multiple reflectors with different Doppler shifts remain unknown. Experiments using masking tones [39] suggest the bats lower their emission frequency such that the frequency of the maximally Doppler shifted echo is close to the reference frequency (i.e. the frequency they are maximally sensitive to). However, the compensation exhibited depends also on the intensity and delay of the echoes as well as the time constant of the feedback loop [39, 40]. As a first order approximation, we assumed that the synthetic bat lowers its emission frequency by about 2.6 kHz to compensate the Doppler shift for reflectors with heading ϕ = 0 (at vbat = 6 m/s and femission = 75 kHz). Lower flight speeds result in reduced Doppler shifts. This implies that we assume that reflectors i with ϕi > 0 return echoes with frequencies between 0 and about 2600 Hz below the reference frequency. Hence, in our simulations, we attenuate echoes for ϕi > 0 as bats are less sensitive to frequencies below the preferred frequency. The attenuation cϕi for each echo as a function of the heading angle ϕi was determined based on data reported by Neuweiler [7] (See Fig 1b). It should be noted that this simple implementation of the Doppler compensation mechanism overestimates the loss in sensitivity due to Doppler shifts. Indeed, we assume the maximum Doppler shift experienced (and, hence the decrease in emission frequency) is always equal to the hypothetical Doppler shift for an object with heading zero degrees—even if these echoes have large delays or low amplitudes. In reality, bats lower their frequency to a lesser extent when echoes have low intensity and/or long delays [39, 40].
In the current simulations, we modeled the bat Rhinolophus rouxii which uses constant frequency calls in the range 73–79 kHz [5]. We choose to approximate the call frequency using 75 kHz. The atmospheric absorption af at 75 kHz was set to 2.4 dB/m [35]. The directional sensitivity dϕi,p of the synthetic bat’s hearing and emission for 75 kHz was taken from previous simulation studies [13, 14, 41]. The maximum gain of the head related transfer function was set to 4.5 dB at 75 kHz [42].
As pointed out above, experimental results confirm that the typical ear movements of Rhinolophidae support obstacle avoidance [3]. The continuous movement of the pinnae is approximated by modeling the directional sensitivity of the two extreme positions p of the ears. This is warranted by the fact that the controller proposed in this paper (detailed in the next section) only processes the onset of the echoes, i.e. the first millisecond. The available evidence [11, 12, 43] suggest that the pinnae are in the most extreme position at the onset of the echo and sweep to the inverse orientation while receiving the echo(es).
Pinna movements are simulated by rigidly rotating the hearing spatial sensitivity pattern before combining it with the emission directivity to obtain the complete directional sensitivity (see [13, 14] for details). Measurements have shown that the ears of Rhinolophidae do not undergo rigid rotations but instead deform while rotating [44]. However, current evidence leaves open the question whether the effects of this deformation on the hearing spatial sensitivity pattern is functionally relevant or not. Ref. [45] discusses the validity of modeling the ear movements as rigid rotations. The modeled head related transfer functions for the two pinna positions p are depicted in Fig 2.
At the heart of the sensorimotor system responsible for obstacle avoidance behaviour we propose a biologically feasible controller that does not rely on explicit reconstruction of the 3D layout of individual reflectors to explore the possibility that Rhinolophidae can avoid obstacles without making use of a 3D model of the world. The controller is illustrated in Fig 3. We assume that the flight parameters are updated after every call based only on the echoes of the last call. Hence, the proposed controller constructs no internal model of the world and does not explicitly exploit changes in echo characteristics across calls. Assuming otherwise would require us to specify a segmentation and grouping mechanism by which individual echoes from subsequent calls are assigned to so-called echo-streams corresponding one-to-one with particular objects. The use of such echo-streams has been hypothesized [46] as a means for a bat’s perceptual system to organize acoustic information from complex environments. However, no explicit computational mechanism capable of the required segmentation and grouping of complex echo signals has been put forward so far. Also, while neurophysiological evidence [47, 48] for an echo stream based representation for single reflector stimuli has been found, no multiple reflector stimuli have been experimented with yet. Hence, until the possible use of an echo stream based representation in obstacle avoidance behaviour is further clarified we propose our reactive controller as a simpler and computationally explicit hypothesis. The main advantage of a reactive approach is that it considers the world as its own best model [49, 50] which is always exactly up to date and always contains every detail there is to be known [51]. By avoiding the delay due to the reconstruction of a 3D model of the environment and/or planning a path, a reactive approach results in a highly responsive and robust controller [50]. However, it should be noted that relying only on the echoes from the last call to determine the controller’s response does not make the proposed sensorimotor system memoryless. Indeed, the dynamics of the interaction between the controller and its environment introduce an implicit memory of information extracted from previous call-echo pairs. Put differently, the state of the controller, i.e. position and velocity, and latest call-echo pair jointly determine the bat’s next move, thereby ensuring that the perceptual history, i.e. previous call-echo pairs, and not just the last call-echo pair determine the controller’s response.
In the simulations, we assume the speed of the synthetic bat vbat to be a function of the time of flight of the first echo, i.e. the distance to the nearest object. The range of speeds goes from 6 ms−1 to 0.3 ms−1 (see above). In addition to the speed, the flight direction also needs to be updated based on the echoes from each call. For an obstacle avoidance algorithm based on sonar, desirable flight directions are characterized by low amplitude echoes. Indeed, for the same reflector strength, weaker echoes imply obstacles that are further away or located more to the periphery. A heuristic leading to weaker echoes is to turn towards the direction of the ear which receives the weakest echoes, e.g. turning right if the right ear receives the weakest echoes.
With stationary ears, moving in the direction of the ear receiving the weakest echo would only allow for updating the horizontal flight direction. However, the ear movements of Rhinolophidae result in the main sensitivity axis of each ear to alternately point up and down. The available evidence suggests that Rhinolophidae move one ear up and the other ear down while receiving echoes [11, 12, 43]. The ears move in the other direction while receiving the next echo. In this paper, we simplified the continuous movement of the pinnae by modeling only the two extreme positions of the ears (see below and Fig 2). Considering the extreme positions of each ear results in the sonar system sampling four directions during each pair of successive calls. Therefore, we propose our controller to turn left or right depending on which ear receives the weakest echo. In addition, the controller steers up or down depending on whether the ear receiving the weakest echo is currently pointing up or down.
Echoes arriving earlier are reflected by more proximate obstacles. Hence, the initial part of the echo signal is of greater importance to an obstacle avoidance sensorimotor system. Therefore, we chose to take only the first millisecond of the echo into account (i.e. the controller only uses the onset of the echo train). We do not claim that the remainder of the echo has no function in obstacle avoidance, but we propose, as indicated by the results, that the onset of the echoes already contains sufficient information. Rhinolophidae have their ears at extreme positions in between calls and move them into the opposing configuration while receiving echoes [12]. Hence, by focussing on the onset of the echoes, we can further simplify the model and use only the extreme ear positions for each call. Apart from the resulting simplifications to our model we argue that focussing on the onset of the echoes has advantages for bats as well. Any mechanism that makes use of specific characteristics of the modulation pattern of the echo introduced by the complete pinna movement instead (e.g. [13, 14]), needs to control and/or to measure the ear movement in greater detail requiring a more complex and less robust system.
In our simulations, the echoes received at each ear t during the first millisecond after the arrival of the first echo are summed with randomized phase shifts. The intensity gt, in decibels, of the summed echoes i received at ear t is given by,
g t = 20 · log ( | ∑ i 10 g i 20 e j · φ i , t | ) (3)
In Eq (3), ϕi,t is a random phase angle (between −π and π) modeling the interference between narrowband echoes. Note that this phase angle is randomized independently for each reflector i and ear t. The hearing threshold was assumed to be 0 dBspl. Therefore, echo amplitudes gi lower than 0 dBspl were set to 0 and did not contribute to intensity gt.
We propose the bat rotates in the direction of the ear receiving the weakest echoes, given by gt (Fig 3, box 5). If gl < gr, the bat turns left. Conversely, if the right ear receives the weakest echoes (gl > gr), the bat turns to the right. Moreover, if gl < gr and the left ear is pointing up (down) the bat turns up (down).
In addition to the direction of the turn, the controller also needs to specify the magnitude of the turn (Fig 3, box 6). In the proposed controller, the magnitude of the turn depends on the flight speed (which in turn depends on the distance to the closest obstacle, Fig 3, box 7). Jones and Rayner [6] report on the speed and angular rotation of Myotis daubentonii (See Fig 1c). We fitted a linear function to this data to obtain the following expression for angular rotation R in degrees per second as a function of flight speed, R = 665 − 116 × Vbat. Values of R smaller than zero were set to zero resulting in the curve depicted in Fig 1c. Incidentally, the turning rates thus obtained correspond largely to those reported by Holderied [38].
Note that for low flight speeds the turning rate could be greatly increased. For example, Aldridge [4] reports that R ferrumequinum is capable of turning with a curvature of up to 115 m−1 (turning radius < 1 cm, angular rotation speed ∼ 1900 deg/s) when suddenly faced with a barrier. Nevertheless, as we did not aim at modeling such last minute avoidance manoeuvres, we opted for fixing the maximum turning rate to the conservative value of 665 degrees per second at Vbat = 0.
In summary, the controller turns left or right depending on whether the left or the right ear received the loudest echoes. In addition, it turns up or down depending whether the ear receiving the loudest echoes is currently pointed up or down. The speed of the bat is determined by the closest (detected) obstacle (Fig 1a). In turn, the rotation speed is determined by the speed of the bat (Fig 1c). See Algorithm 1 for a listing of the computations and Fig 3 for a graphical depiction of the complete controller.
We give the synthetic bat the same aerodynamic freedom in the horizontal (left and right) and vertical plane (up and down). This is; it can turn at the same rate without taking gravity into account (but see below for a version of the controller taking into account the gravity vector). Indeed, if the synthetic bat turns upwards/downwards for long enough, it might eventually fly upside down with respect to its initial orientation. There are two reasons for modeling the vertical rotation in this way. First, while it is well known that bats are very agile, to the best of our knowledge very little information is available about the aerodynamic constraints on climbing and ascending flight of the bat. Second, and more importantly, by introducing the same constraints on both horizontal and vertical rotations, we can compare the sensorial performance of the algorithm in both the horizontal and vertical plane in the absence of differences in motor constraints. Nevertheless, we are aware that having the same constraints for both turning rates is artificial. Hence, we also test a variant of the controller that introduces a constraint on the maximum vertical rotation (see ‘constrained’ controller below).
Both R. rouxii and R. ferrumequinum emit a pulse every 80 to 90 ms on average [5, 52]. For computational ease and to simulate a lower bound update rate, the synthetic bat was simulated to emit a pulse every 100 ms. On approaching a landing site, the pulse rate of R. ferrumequinum was found to increase to about 80 Hz (i.e. about 12 ms interval) [9, 53]. However, the informational update of 80 Hz might not translate into an ability of the bat to update its direction 80 times a second. Rhinolophidae flap their wings at about 12 Hz (i.e. about 80 ms interval) irrespective of their air speed [54]. Considering a wing beat as the minimal unit that allows changing the direction of the flight, would allow for an update rate of at most 12 Hz. In the proposed controller, each pulse corresponds to a single update in the flight direction. Hence, as 12 Hz is very close to the modeled pulse rate of 10 Hz, the interpulse interval was fixed at 100 ms.
In the simulations, we account not only for the time needed for the echoes to arrive but also for the time required by a bat to process the echoes and produce a motor response. Übernickel et al. [55] found a reaction time of about 50 ms to transient targets in the trawling bat Noctilio leporinus in accordance with a similar range of reaction times 47–63 ms found in ref. [56]. Hence, we allowed for 50 ms to process the echoes. In the interval between the emission of the call and the start of the turn, the current direction and speed of flight is maintained. The interval between call and start of turn is given by (1) the time for the first echo to arrive, (2) 1 ms over which the echoes are summed and (3) 50 ms of processing time. Note that as the duration of the turn is given by the fixed call period (100 ms) minus the interval between call and start of turn, both the rotation speed R and the duration of the bat’s turn depend on the distance to the closest object. As the time for the first echo to arrive gets shorter, the turn duration gets longer. This increases the rotational gain of the controller even more for nearby obstacles.
The controller was tested in 2D and 3D environments. The performance of the controller described above, referred to as the default controller from now on, was compared to that of five related controllers:
Fixed ears: This controller models a bat with static ears. A single directional sensitivity is used for each ear. As such, the directional sensitivity does not change from call to call. The directionality used is depicted in the top row of Fig 2b. The azimuthal rotation is updated as in the default controller. However, the sign (up/down) of the elevation rotation was selected at random. Off axis pinnae: This controller is identical to the default controller. However, this controller consistently has the left ear pointing downwards and the right ear pointing upwards. If the left (right) ear receives the weakest echoes, the bat turns downwards (upwards). The azimuthal rotation is updated as before. Random A: The controller differs from the default controller by randomly turning left or right at each call. While the direction of rotation is chosen randomly, the magnitude is calculated as in the default controller. Random B: The same controller as the Random A variant with the addition that the rotation speed of the bat is also chosen randomly from the interval 0 to 350 degrees per second. Constrained: This controller is identical to the default controller, but it constrains the angle of the bat in the vertical plane. The bat can maximally attain a climbing or descending angle of ±60 degrees.
The controller with fixed ears allows us to test the contribution of ear movement to obstacle avoidance. The controller with the pinnae fixated in an off-axis position allows us to test whether the cues necessary for obstacle avoidance are still present if the ears are fixed but are not aligned with the horizontal plane. Random A and Random B are included in the tests as baseline conditions against which to compare the other controllers. Similarly, in behavioural obstacle avoidance experiments, the performance of the bat is typically compared to the number of collisions expected from following a random path through space, e.g. [3, 25, 26]. Finally, the constrained controller adds more realistic constraints to the vertical rotation of the bats.
We tested the controller and its four variants in environments populated with reflectors on hexagonal grids spaced 15 cm apart (See Fig 4c and 4f for examples). In these environments, collisions are counted as the number of time steps (calls) the controller was closer than 2.5 cm to any obstacle. Hence, we modeled the synthetic bat as having a body width of 5 cm, in agreement with Mogdans et al. [3]. In most results reported below, the various controller variants differed in the resulting average distance kept from reflectors and, therefore, in their average speed and distance travelled. To compensate for this, we normalized the number of collisions for all controllers to the number of collisions per 100 m travelled.
The results show that the default controller successfully avoided both the vertical wires and the horizontal wires (Fig 4). Indeed, in these 2D tasks the number of registered collisions was much lower than in both random A and B. The controller with the fixed ears performed equally well in avoiding the vertical wires. However, avoidance of the horizontal wires was reduced to chance level by fixing the ears in the horizontal plane. In contrast, fixating the pinnae off axis, restored the obstacle avoidance performance for the horizontal wires (and did not reduce performance for the vertical wires).
The controller that was constrained in its vertical rotation performed much worse than the default controller in avoiding horizontal wires (Fig 4d). This indicates that while the moving ears supplied the necessary information to avoid obstacles, the imposed aerodynamic constraint is too restrictive to allow for successful obstacle avoidance in our grid of simulated wires.
Overall the performance results in the regularly spaced grids match the finding of Mogdans et al. [3] that obstructing the pinnae movements only interferes with the avoidance of horizontal wires, i.e. only obstacle avoidance in the vertical dimension is affected. Fixating the ears did not have an effect on the avoidance of the vertical wires. In addition, our simulations suggest that pinnae fixated in an off-axis position provide sufficient cues for obstacle avoidance in both azimuth and elevation.
Fig 5a and 5d show the number of collisions registered for the bat in 100 replications with reflectors scattered in either the horizontal or the vertical plane for the four variants of the controller. In these runs, collisions are defined as the number of time steps (per 100 m travelled) the controller was closer than 15 cm to the nearest obstacle, i.e. approximately half the wingspan of R. rouxii. The results indicate that the default controller is capable of avoiding obstacles in both the vertical and the horizontal plane. Fixing the pinnae has no effect on obstacle avoidance in the horizontal dimension. However, obstacle avoidance in the vertical dimension is reduced to chance level (i.e. similar number of collisions than controller Random A).
In the horizontal plane, the constrained controller has the same degrees of freedom as the default controller and, therefore, has the same performance. Constraining the elevation angle of the bat clearly limits its freedom. Hence, the number of collisions does increase compared to the horizontal plane. However, the number of collisions is still less than in both random baselines. The reduction in performance for the constrained controller is less dramatic than for the regularly spaced obstacles discussed above as can be seen from comparing the performance of the constrained controller in Figs 4d and 5d. We provide two movies illustrating the behaviour of the controllers in the 2D environments of Fig 5 as supplementary material.
The default controller, as well as the four derived controllers, were also tested for obstacle avoidance in 3D point clouds (Fig 6, also provided as MATLAB figure in the supplementary material (S3 Fig). The default algorithm performs best. Fixing the ears does not result in an increase in the number of collisions. However, it results in flying somewhat closer to obstacles. The number of collisions does not increase by fixing the ears as the controller is still able to avoid obstacles in the horizontal plane. This implies the controller with the fixed ears solves the 3D obstacle avoidance problem as a sequence of 2D problems. Indeed, the 3D point clouds do not require the controller to perform obstacle avoidance in both horizontal and vertical plane simultaneously, it can avoid collisions by avoiding obstacles in a single plane.
The two random controllers performed worse than the default controller with a drastic increase in the number of collisions. The constrained controller performed at the same level as the default controller with respect to the number of collisions. Hence, the reduced freedom in elevation rotation does not seem to hamper this controller in this environment.
The tilted torus environment explicitly tests whether the controller(s) can follow a corridor in both azimuth and elevation. The results depicted in Fig 7 show that the random controllers result in more collisions (Fig 7a) and flying closer to reflectors (Fig 7b, (also supplied as a MATLAB figure in the supplementary material (S4 Fig) than the other controllers. The number of collisions follows a similar pattern as the number of collisions in the 3D environment depicted in Fig 6. However, more importantly, only the controllers with moving ears (i.e. the Default and Constrained controllers) succeed in following the torus. The random controllers often exit the torus quickly, explaining the low number of collisions for the controller Random B. The controller with fixed ears stays in the torus without colliding but is unable to complete a circular path inside the torus. It is confined to a subsection of the torus.
Fig 8 shows the results of 50 replications of the experiment using the 3D scanning data from the fir forest. Likewise Fig 9 shows the results of 50 experimental runs using the 3D scan of the forest corridor. As real bats show nearly 2D flight behaviour in similar real environments (as found e.g. in Holderied [57]), we ignored the elevation commands of the controller resulting in 2D flight paths in these simulations. In both environments, Random A and B performed substantially worse than any other variant. Note that, in these experiments, while the bat’s flight path is restricted to a plane the echo signals the controller derives its decisions from are calculated based on the full 3D environment.
Echolocation supports the execution of many tasks varying widely in computational complexity such as object recognition [16, 58], prey localization [22, 59], finding water [60] and navigation [61]. Understanding the echolocation ability of bats can be thought of as isolating the (minimal set of) cues needed to perform each of these different tasks and confirming the sufficiency of those cues in behavioural experiments [36, 60], simulations [13, 14, 20, 62] or robotic studies [15, 20, 63, 64]. For some tasks, a minimal set of sufficient cues has been determined. For example, water bodies can be identified as horizontal reflective surfaces. Indeed, any horizontal surface with the correct reflective properties is readily mistaken by bats as a water surface [60]. Other tasks for which a minimal set of cues has been determined include the recognition of flower size [36] or prey size [65, 66]. However, while the ability of bats to avoid obstacles was the first to be studied (e.g. [26] and reviewed in [25]) following the groundbreaking experiments by Lazzaro Spallanzani (1729-1799, described in [67]), relatively little is known about the minimal set of cues sufficient to support obstacle avoidance, one of the most basic echolocation supported tasks [68].
One explanation for this hiatus seems to be the assumption that the bat’s ability to avoid complex obstacles relies on the more basic competence of reconstructing the 3D layout of its surroundings first. It has been suggested [3, 69] and verified [15] that Rhinolophidae, using long narrowband calls, could localize single targets using the changing IID cues generated by their moving ears. However, complex reflectors in bat habitats, such as plants and trees [16], return many overlapping echoes. The multitude of echoes returned by natural obstacles is problematic for such a localization strategy because none of the proposed cues has been demonstrated to be robust in the face of many overlapping echoes. Similarly, bats using frequency modulated calls can locate single targets based on binaural spectral cues, e.g. [59, 62, 70, 71]. However, due to the temporal integration in the bat’s auditory system [72, 73], spectral cues will also degrade when faced with many overlapping echoes [74]. In addition, spectral cues are unreliable for low amplitude echoes [62]. As a growing body of research on bat echolocation shows that bats can cope with extremely challenging situations (e.g., [27, 75]), we conclude that the cues used for obstacle avoidance must be robust and available even (or, especially) in situations where a multitude of complex objects generate many overlapping echoes. Hence, we argue that it is unlikely that a 3D reconstruction capability would be a precondition for successful obstacle avoidance.
In this paper, we propose an alternative obstacle avoidance strategy that does not rely on explicit 3D reconstruction. This strategy is capable of obstacle avoidance, even when faced with complex obstacles returning overlapping echoes. Indeed, even though the controller only uses the first millisecond of the returning echo train, this short interval typically contained echoes from multiple reflectors (see histograms in Fig 10). The median number of reflectors returning a detectable echo within the first millisecond varied across simulations. For example, the median number of echoes returned by the 3D artificial environment was 3. In contrast, the median number of detectable echoes returned by the torus was 36. In addition to variation across conditions, the number of detectable echoes also varies from call to call within conditions. This is demonstrated by the long tails of the distributions.
Our results from both 2D and 3D simulations in artificial and natural environments show that the IID and delay cues derived from the onset of the first echo when combined with characteristic ear movements are sufficient to support obstacle avoidance using the bat R. rouxii as a model. It is true that, while this minimal set of cues seems sufficient to avoid obstacles in most cases, the controller did fly into obstacles on a number of occasions, e.g. Fig 4. However, real bats are not perfect at avoiding obstacles either and they sometimes have to fall back upon some last moment collision avoidance behaviours, presumably if echoes become too loud or too close [25]. For instance, Aldridge [4] found that R. ferrumequinum was capable of turning with an angular velocity of up to 1900 degrees per second when suddenly faced with a barrier. Therefore, bats are capable of performing very agile last minute evasive manoeuvres. Such behaviour was not programmed into the controller but could have avoided collisions in the limited number of instances where the synthetic bat ventured too close to obstacles.
An interesting further result pertaining particularly to constant frequency bats is the matching of the experimental results of Mogdans et al. [3]. In spite of their limited magnitude (about 30 degrees, see Fig 2), the success of our default controller confirms that ear movements can explain obstacle avoidance in the vertical plane (Fig 5). Furthermore, as in the experiments of Mogdans et al. [3], fixing the ears reduced the synthetic bat’s ability to avoid horizontal wires while leaving the ability to avoid vertical wires intact. Therefore, this obstacle avoidance study, by suggesting a specific mechanism, adds further evidence in favor of the functional relevance of these small ear movements for constant frequency bats.
A prediction following from our proposed controller is that cyclic ear movements are not strictly necessary for obstacle avoidance. The controller only uses a snapshot at the onset of the cycle. Indeed, for obstacle avoidance in the horizontal plane the controller relies on the different azimuthal directions in which both ears point. Similarly, for obstacle avoidance in the vertical plane, the controller only requires the ears to be pointing in different elevation directions. We tested this prediction using ears that were fixed in an off-axis position (Fig 4, controller with off-axis ears). The results confirmed the prediction. Hence, we suggest that Rhinolophidae with their ears fixed in an off-axis position have access to sufficient information to avoid obstacles in both planes. In this respect, it is interesting to note that Mogdans et al. [3] report: “Single photographic flight records of intact bats revealed that bats sometimes passed vertical wires with the head tilted off the horizontal plane”. Such head tilting would have a similar effect than fixing the ears off-axis as we did in our experiments. Please note that we do not want to imply that the cyclic ear movements do not provide additional essential information that can be used to control other behaviours, e.g. the localization of individual reflectors such as prey [13–15].
The controller proposed in this paper is dependent upon the sign of the IID only: it turns left/right and up/down based on which ear receives the lower echo amplitude. Therefore, the algorithm supposes robustness against any alterations of the head related transfer function that preserve the tendency for ipsilateral reflectors to be louder than contralateral ones. Early experiments have shown that FM bats of which both pinna and tragus were removed avoided obstacles just as well as bats with intact ears [28]. Furthermore, disrupting the IID cues by plugging one ear reduces the obstacle avoidance performance of R. ferrumequinum. Plugging both ears lightly (attenuation 15–25 dB) does not deteriorate obstacle avoidance [76]. Plugging both ears more tightly (echo attenuation 55–60 dB, [76]) or completely [28] reduces obstacle avoidance performance presumably by preventing echoes from being detected. These results indicate that crude binaural intensity cues, those that are unaffected by the removal or the plugging of both ears but are affected by changing the sensitivity of a single ear, are sufficient to avoid obstacles. These results are in agreement with the predictions from the proposed controller, as it relies only on the sign of the IID. In contrast, these early findings [28, 76] can not be explained by assuming that obstacle avoidance depends on a 3D reconstruction of the bat’s surroundings. Indeed, deforming the outer ears by gluing the tragus forward to the side of the head has been shown to increase sound localization errors in FM bats [59, 70, 77]. A 3D reconstruction would imply the bat can simultaneously localize multiple reflectors in both azimuth and elevation. This seems to require the presence of intact pinnae [59, 70, 77], which was not the case in the study of Hahn [28]. Since he found that the bats could still avoid obstacles without pinnae a 3D reconstruction of the obstacle layout does not seem necessary for obstacle avoidance.
It should be noted that the experiments reported in ref. [28] were not conducted in the dark. Therefore, bats might have relied both on vision and on echolocation. Nevertheless, depriving the bats of their hearing by filling the meatus with plaster (but not the removal of the tragi and pinnae) resulted in increased collisions. Hence, while the bats could be relying partly on vision in ref. [28], they were clearly echolocating as well and could not solve the obstacle avoidance task by relying on vision alone.
To validate the plausibility of the flight speeds generated by the proposed controller we evaluated the simulated flight speeds for a number of the simulations discussed above by plotting their distribution (Fig 11). From these plots, we conclude that the flight speeds are realistic. The default algorithm resulted in an average speed of about 2.2–2.3 ms−1. Fawcett and Ratcliffe [78] reported on the flight speed of untrained M. daubentonii in a small and a large flight room with a ground surface area of 3 m × 3 m and 7 m × 4.8 m respectively. The weight of this species ranges from 5 to 10 grams [79]. Commuting speeds between 3 and 8 ms−1 have been reported [80]. In the experiments of Fawcett and Ratcliffe [78], single bats adopted an average flight speed of about 2.2 ms−1 in the larger flight room and 1.3 in the smaller flight room. Hence, the average flight speed used by our default algorithm is very close to the flight speeds reported for the larger flight room. Hence, although we were unable to find flight speeds for R. rouxii, evidence from another species corroborates our simulated flight speeds.
The results presented in this paper can be readily extended to bats using frequency-modulated (FM) calls. For obstacle avoidance in the horizontal plane, this extension follows directly from our results. Indeed, in our simulations the controller avoids obstacles in the horizontal plane by using first echo delay and IID extracted from a single narrow frequency band. Moreover, the bat only processes the onset of the echo (i.e. the first millisecond). This type of transient information is also available to bats using FM signals. The main difference between FM and CF bats in this respect is that FM bats have access to IIDs across multiple frequency bands. Bats navigating along hedgerows [57] or among the trunks of trees could make use of this horizontal obstacle avoidance mechanism.
To demonstrate that the proposed mechanism indeed extends to FM bats avoiding obstacles, we modeled an FM bat flying in heterogeneous artificial environments (identical to those used in Fig 5). The controller was adapted to use the head related transfer function [81] and emission directivity [82] of the FM bat Phyllostomus discolor at 60 kHz (atmospheric attenuation: 2 dB/m [35]). P. discolor uses frequency modulated calls which include frequencies between 40 and 90 kHz [83, 84]. However, in the current simulations, we simulated only one of the frequency channels available to this bat. i.e. we modelled a single frequency channel at 60 kHz. No cyclic ear movements were simulated. In addition, as FM bats do not compensate for Doppler shifts this behaviour was omitted. Apart from these changes, the controller was not altered. In flight, the calls of P. discolor have been reported to reach a peak intensity of 124 dB (cited in [85]). Hence, we used 120 dB as emission strength (gbat, Eq (1)) as before. The maximum gain of the HRTF was set to 6 dB [86].
To the best of our knowledge, ear movements of FM bats in flight have only been studied in the final approach during prey capture, e.g. [22, 87]. Hence, it is unknown whether FM bats exhibit ear motions while avoiding obstacles. However, as indicated above, ear movements are not necessary for successful obstacle avoidance in the vertical plane. Indeed, the controller with pinnae fixed in an off-axis position performed nearly as well as the default controller (see Fig 4). Therefore, we hypothesize that FM bats might be able to avoid obstacles in azimuth as well as elevation by turning their pinnae off-axis. We tested this by combining the controller with an HRTF obtained by rotating the left ear down by 15 degrees and the right ear up by 15 degrees (see Fig 12). This is the same rotation of the pinnae as used for the simulations of R. rouxii. We used the same configuration of the simulated ears for the Constrained controller. The default controller, on the other hand, has both ears co-located in the horizontal plane (see Fig 12).
Fig 13a-13c shows that, as expected, the controller using the P. discolor directionality can avoid obstacles in the horizontal plane. The only controller variants that were unable to avoid obstacles were Random A and Random B. Pointing the ears off axis did not have an (adverse) effect on the obstacle avoidance behaviour. The results in Fig 13d-13f show that equipping the FM bat controller with ears pointing off-axis results in increased obstacle avoidance performance. In contrast, having the ears co-located in the horizontal plane (i.e. the Default controller) leads to numerous collisions.
Rotating their ears into an off-axis position is only one way in which FM bats could compensate for the absence of cyclic ear movements during flight. They could also change the orientation of their heads and/or bodies between calls. In fact, this behaviour has been observed in CF bats when being prevented from rotating their pinnae. Mogdans et al. [3] reported that in their experiment, the CF bats with immobilized pinnae showed more vigorous head movements than before surgery and compared to the controls while hanging in the flight room. They also reported, as referred to above, that flight records of intact bats revealed they sometimes passed vertical wires with the head tilted off the horizontal plane. Evidence for changes in head orientation in FM bats has been reported for Eptesicus fuscus which has been shown to be able to shift its beam from call to call, e.g. [88]. Likewise, pipistrelle bats were found to exhibit extensive scanning behaviour in azimuth and elevation while flying through natural habitats [89]. This behaviour in combination with the mechanism proposed above, i.e. rotate towards the ear receiving the weakest echo, would support obstacle avoidance based on the same cues used by our controller.
The sensorimotor strategy proposed in this paper can be readily incorporated into a behavior-based control architecture. This type of controller, originally proposed for robots by Brooks [90] and inspired by neuroscience [50], decomposes complex behavior into a number of independent sensorimotor loops (reviewed in refs. [50, 91, 92]). Each sensorimotor loop controls a single behaviour such as obstacle avoidance, approaching targets or corridor following. All sensor data is fed into each loop. However, loops only extract the information necessary for the behaviour they control. An action selection mechanism (e.g. mutual inhibition of behaviours [90]) ensures that only a single sensorimotor loop drives the actuators [93] at each point in time.
Brooks proposed the behaviour-based control architecture as an alternative to so-called deliberate control architectures. These controllers process the sensor data to derive a general representation of the world first. Once a general and complete representation has been derived, planning and reasoning algorithms are employed next to determine the most suitable action sequence [49]. However, deriving a representation that supports all required actions has proven to be the most challenging aspect of deliberate controllers. Indeed, experience in robotics has learned this is only possible for highly simplified environments. Today, no autonomous robot operating in realistic environments is operated by an entirely deliberate control architecture [91]. In contrast, behaviour based controllers avoid having to compute explicitly an internal representation of the world. Indeed, in the words of Brooks, in a controller consisting of multiple sensorimotor loops
We argue that the fact that behaviour-based control does not depend on the extraction of a general representation of the environment makes it an appealing candidate as a control strategy in echolocating bats. Indeed, the sparseness and unreliability of localization cues makes deriving a general representation of the world very difficult, if not impossible under many real world conditions. A behaviour-based control architecture would circumvent this issue by only relying on extracting (and possibly storing [94]) those cues necessary for a particular sensorimotor loop.
Furthermore, behaviour-based control architectures readily allow for redundancy. Each behaviour (e.g. obstacle avoidance [90]) can be controlled by multiple, independent sensorimotor loops each exploiting different cues. For example, in the current paper we have proposed a sensorimotor loop for obstacle avoidance based on IID and time of flight cues derived from the onset of the first echo. However, we acknowledge that bats may use many more echo cues than the ones we have exploited in this paper. Also, they are likely to integrate more echo information across calls, i.e. base their decisions on echo-stream information. In particular, CF bats might be using the complete echo for extracting IID, use Doppler shifts or use the FM parts of the echoes. FM bats, on the other hand, are very likely to use spectral cues whenever available. Each of these cues could be extracted, evaluated, stored and mapped to motor commands by a set of dedicated sensorimotor loops taking precedence through an adequate action selection mechanism. This would lead to a high level of robustness as, in case a particular sensorimotor loop fails to extract the relevant cues, other loops will take over motor control.
In summary, we tentatively propose that many aspects of bat echolocation—including prey capture, obstacle avoidance and navigation—could be modeled by a behaviour-based control architecture consisting of a set of sensorimotor loops each extracting and exploiting a subset of cues from the echoes. Indeed, other sensorimotor loops proposed in the past fit readily in this framework, e.g. the prey capture strategies proposed by Kuc [20] and Walker et. al. [15] or the models of target approach proposed by Lee et. al. [95] and Bar et al. [96]. In the case of obstacle avoidance, we consider the proposed obstacle avoidance behaviour to be a robust sensorimotor loop to which both FM and CF bats can fall back on in case less reliable cues are unavailable. We maintain that a behaviour-based controller would result in a robust echolocator capable of exploiting a wide range of cues whilst keeping computational demands limited by avoiding the need to reconstruct a general representation of the environment from noisy and complex echoes.
Proposing a behaviour-based architecture as a model for echolocation based control in bats implies that future research should not only focus on identifying sensorimotor loops underlying different behaviours but also on how these loops interact and how context-dependent action selection is achieved. Indeed, a behaviour-based controller offers a framework in which to analyze the bats’ flexibility in exploiting a variety of (multimodal) cues under changing circumstances, e.g. [74].
In conclusion, we propose that Interaural Intensity Differences calculated on the onset of the first echo, in combination with first echo delay, constitute a sufficient set of stable and robust cues for avoiding obstacles in a 3D world—without the need to reconstruct the 3D layout of the reflectors from complex and noisy echo signals. Our simulations suggest that exploiting these cues would allow both FM and CF bats to perform this basic echolocation subtask with a limited computational load and minimal latency providing a hard real-time response capability.
|
10.1371/journal.ppat.1002060 | Stromal Down-Regulation of Macrophage CD4/CCR5 Expression and NF-κB Activation Mediates HIV-1 Non-Permissiveness in Intestinal Macrophages | Tissue macrophages are derived exclusively from blood monocytes, which as monocyte-derived macrophages support HIV-1 replication. However, among human tissue macrophages only intestinal macrophages are non-permissive to HIV-1, suggesting that the unique microenvironment in human intestinal mucosa renders lamina propria macrophages non-permissive to HIV-1. We investigated this hypothesis using blood monocytes and intestinal extracellular matrix (stroma)-conditioned media (S-CM) to model the exposure of newly recruited monocytes and resident macrophages to lamina propria stroma, where the cells take up residence in the intestinal mucosa. Exposure of monocytes to S-CM blocked up-regulation of CD4 and CCR5 expression during monocyte differentiation into macrophages and inhibited productive HIV-1 infection in differentiated macrophages. Importantly, exposure of monocyte-derived macrophages simultaneously to S-CM and HIV-1 also inhibited viral replication, and sorted CD4+ intestinal macrophages, a proportion of which expressed CCR5+, did not support HIV-1 replication, indicating that the non-permissiveness to HIV-1 was not due to reduced receptor expression alone. Consistent with this conclusion, S-CM also potently inhibited replication of HIV-1 pseudotyped with vesicular stomatitis virus glycoprotein, which provides CD4/CCR5-independent entry. Neutralization of TGF-β in S-CM and recombinant TGF-β studies showed that stromal TGF-β inhibited macrophage nuclear translocation of NF-κB and HIV-1 replication. Thus, the profound inability of intestinal macrophages to support productive HIV-1 infection is likely the consequence of microenvironmental down-regulation of macrophage HIV-1 receptor/coreceptor expression and NF-κB activation.
| Human intestinal macrophages, unlike lymphoid tissue macrophages, brain microglia and genital (vaginal) macrophages, are profoundly incapable of supporting productive HIV-1 infection. Intriguingly, all macrophages are derived exclusively from blood monocytes, which are HIV-1 permissive after differentiation into monocyte-derived macrophages. Therefore, the unique non-permissiveness of intestinal macrophages to HIV-1 must be conferred by the intestinal mucosal microenvironment. Here we report that intestinal stroma potently blocked up-regulation of HIV-1 receptor/coreceptor CD4 and CCR5 expression during monocyte differentiation into macrophages and macrophage nuclear translocation of NF-κB, which is a critical requirement for HIV-1 transcription. These two mechanisms work collaboratively to render intestinal macrophages non-permissive to HIV-1. Harnessing this natural antiviral defense may provide a novel strategy to exploit for the prevention of infection in HIV-1 permissive cells.
| Macrophages play crucial roles in the establishment, pathogenesis and latency of human immunodeficiency virus-1 (HIV-1) infection [1], [2], [3] through their ability to support viral replication [4], [5], transmit virus [6] and act as a viral reservoir [6], [7], [8], [9]. In this connection, macrophages throughout the body, including lymphoid tissue macrophages [10], [11], brain microglia [12] and genital (vaginal) macrophages [13], are permissive to HIV-1. In sharp contrast, resident macrophages in the human small intestine are profoundly incapable of supporting productive HIV-1 infection [13], [14], [15], although intestinal macrophages are derived exclusively from blood monocytes [16], which when differentiated into monocyte-derived macrophages are HIV-1 permissive [4], [5], [17], [18]. The unique non-permissiveness of intestinal macrophages to HIV-1 stands in marked contrast to the ability of intestinal CD4+ T cells to support productive viral infection and undergo early, rapid and profound depletion during primary HIV-1 and SIV infection [19], [20], [21], [22], [23], [24], [25], [26].
After their recruitment into the lamina propria, pro-inflammatory blood monocytes differentiate into non-inflammatory intestinal macrophages through stromal transforming growth factor β (TGF-β)-mediated Smad-induced IκBα and nuclear factor kappa B (NF-κB) inactivation, as we recently reported [27], [28]. In further contrast to blood monocytes, intestinal macrophages are markedly down-regulated for receptors that mediate inflammatory responses, including LPS, Fcγ and Fcα receptors [27], [28], [29], triggering receptor expressed on myeloid cells-1 (TREM-1) [30], [31], as well as CD4, CCR5 and CXCR4 [13], [14], [15]. Since CCR5 expression correlates directly with the differentiation of monocytes into macrophages [32], [33], [34], the reduced expression of CCR5 on intestinal macrophages raises the possibility that the non-permissiveness of intestinal macrophages to HIV-1 is related to reduced HIV-1 receptor/co-receptor expression. However, our detection of proviral DNA in isolated intestinal macrophages exposed to HIV-1 in vitro [14] suggests post-entry restriction also may be involved in the inability of intestinal macrophages to support HIV-1 replication.
To elucidate the mechanism that renders intestinal macrophages non-permissive to HIV-1, we exposed blood monocytes and monocyte-derived macrophages to conditioned media from cultured lamina propria stroma isolated from normal human jejunum to determine the effect of the lamina propria microenvironment on CD4/CCR5 expression and HIV-1 permissiveness. Our results indicate that the inability of primary human intestinal macrophages to support HIV-1 replication is likely due not only to the marked down-regulation of CD4 and CCR5 but also to the inability of intestinal macrophages to activate NF-κB, a critical requirement for HIV-1 transcription.
CCR5-tropic HIV-1 strains are predominant among the transmitted/founder viruses isolated from acutely infected persons [35], [36], [37]. Since the gastrointestinal mucosa is the largest reservoir of macrophages in the body [38], and macrophages are an important HIV-1 target cell, we initiated studies to define the HIV-1 receptor phenotype and permissiveness of purified intestinal macrophages to macrophage-tropic HIV-1. Intestinal macrophages and blood monocytes were isolated from the same donors, purified and analyzed for expression of the HIV-1 primary receptor CD4 and the coreceptors CCR5 and CXCR4. As shown in Table 1, very low proportions of intestinal macrophages expressed CD4 (1.0%), CCR5 (0.8%) and CXCR4 (2.1%), and a barely detectable proportion (0.3%) expressed both CD4 and CCR5 (P = 0.0001 to P = 0.039), consistent with our earlier finding of markedly diminished CD4, CCR5 and CXCR4 expression on intestinal but not vaginal macrophages [13]. The low levels of CD4 and CCR5 expressed on intestinal macrophages corresponded to low levels of receptor/co-receptor-specific mRNA [13]. In contrast, modest proportions of blood monocytes expressed CD4 (11.6%), CCR5 (2.9%) and CXCR4 (14.1%), and 2.2% of the monocytes were CD4+CCR5+, indicating that 3- to 10-fold fewer intestinal macrophages expressed the receptors compared to autologous blood monocytes (Table 1).
We previously showed that isolated intestinal macrophages do not support HIV-1 replication [13], [14], [15]. The low level of CD4, as well as CCR5, on intestinal macrophages (Table 1) raised the possibility that a restriction in HIV-1 entry could contribute to the cells' non-permissiveness to HIV-1. To address this possibility, we sorted autologous CD4+ intestinal macrophages and blood monocytes by magnetic activated cell sorting (MACS), cultured the cells for 4 days (>98% viable), inoculated each population with equivalent amounts of highly fusigenic and macrophage-tropic R5 viruses, including NA420 B33, NA20 B59 or NA353 B27, which infect cells with extremely low levels of CD4 and/or CCR5 expression [39], and monitored viral replication by p24 release over 20 days. As shown in Figure 1, 95% of both the intestinal macrophages and blood monocytes were HLA-DR+CD13+. Among the sorted CD4+ intestinal macrophages, 34.1% expressed CCR5 and levels of p24 were barely detectable only on day 12, whereas among the sorted CD4+ blood monocytes, 26% expressed CCR5 and large amounts of p24 were released by the monocyte-derived macrophages up to day 20 (Figure 1). Importantly, neither the exposure of intestinal macrophages to pro-inflammatory stimuli, including lipopolysacharride, interferon-γ or tumor necrosis factor-α, nor culture for up to 2 weeks prior to inoculation with virus, induced HIV-1 permissiveness in the macrophages (data not shown). These findings indicate that even CD4+ intestinal macrophages that express CCR5 are refractory to HIV-1, implicating a post-entry mechanism for down-regulated HIV-1 permissiveness. However, the profound low level of CD4 and CCR5 expression on the total intestinal macrophage population (Table 1) raised the possibility that the mucosal microenvironment of the jejunum caused the down-regulation of CD4 and CCR5, thereby also contributing to the reduced permissiveness of intestinal macrophages to CCR5-tropic HIV-1.
Intestinal macrophages are terminally differentiated and express very low levels of CD4 and CCR5, but they are derived from blood monocytes [16], which, during and after differentiation into adherent macrophages, express high levels of CD4 and CCR5. Since factors released by the intestinal extracellular matrix (stroma) down-regulate an array of innate response receptors on blood monocytes [27], we examined whether stromal factors present in conditioned media derived from normal intestinal stroma (S-CM) [27], [28] also down-regulate CD4 and CCR5 expression on blood monocytes during and after their differentiation into macrophages. Compared to monocytes differentiated into adherent macrophages during 2 days culture in media alone, monocytes differentiated into macrophages in the presence of S-CM (10–500 µg protein/mL) displayed a marked dose-dependent decrease in surface CD4 and CCR5 (Figure 2A). In contrast, when monocytes were first differentiated for 4 days into adherent macrophages and then exposed for 2 days to S-CM, CD4 and CCR5 expression was not down-regulated (Figure 2B). Thus, intestinal stromal products prevent differentiation-induced upregulation of CD4 and CCR5 expression on monocyte-derived macrophages but do not down-regulate receptor/co-receptor expression after the cells have differentiated into macrophages. These findings offer an explanation for the near absence of CD4 and CCR5 on terminally differentiated intestinal macrophages, which are derived exclusively from circulating monocytes that have recruited into the lamina propria.
Since undifferentiated monocytes do not support productive HIV-1 infection, we next determined whether monocyte-derived macrophages exposed to lamina propria stromal products supported HIV-1 replication. Monocyte-derived macrophages were cultured for 2 days in the presence of varying concentrations of S-CM, after which the cultures were inoculated with R5 virus (NA353 B27). As shown in Figure 3A, the pre-incubation of monocyte-derived macrophages with S-CM prior to the inoculation of HIV-1 caused a dose-dependent decrease in p24 production during a 20-day culture period. However, when monocyte-derived macrophages were pre-incubated with conditioned media from purified cultures of intestinal epithelial cells (EC-CM) [40] or intestinal mononuclear cells (MNL-CM) [27] derived from the same donor tissue as the S-CM, HIV-1 replication was not inhibited (Figure 3B). Furthermore, S-CM also caused a dose-dependent decrease in viral replication when S-CM and virus were added simultaneously to the monocyte-derived macrophage cultures (Figure 3C). These findings suggest that extracellular matrix products, rather than intestinal epithelial cell or lamina propria mononuclear cell products, inhibit productive HIV-1 infection in intestinal macrophages and that the down-regulation in viral replication is not the exclusive consequence of the low level of CD4 and CCR5 expression on the macrophages.
To further distinguish between reduced HIV-1 entry and down-regulated viral replication, we pseudotyped HIV-1 with VSV-G envelope to bypass HIV-1 receptor/co-receptor-dependent entry. As predicted, treatment of monocyte-derived macrophages with S-CM for up to 24 hours did not impair the entry of VSV-G pseudotyped virus into the cells (data not shown) but caused a dose-dependent reduction in single-round replication of VSV-G pseudovirons, as shown by immunofluorescence and flow cytometry in Figure 4, upper panels. The same pre-treatment of monocyte-derived macrophages with S-CM also inhibited infection of YU2 pseudovirons in a dose-dependent manner (Figure 4, lower panels). These results further indicate that S-CM inhibition of R5 replication was not due only to down-regulated CD4 and CCR5 expression but also involved post-entry restriction in viral replication.
We have shown that stromal TGF-β inactivates NF-κB in monocyte-derived macrophages by dysregulating NF-κB signal proteins and inducing IκBα, the cytoplasmic negative regulator of NF-κB [28]. Because NF-κB is required for HIV-1 transcription [41], we investigated whether stromal TGF-β-mediated down-regulation of NF-κB also inhibits the ability of monocyte-derived macrophages to support HIV-1 replication. Monocyte-derived macrophages were cultured in triplicate with increasing concentrations of S-CM and inoculated with R5 HIV-1 (NA353 B27) at a multiplicity of infection (MOI) of 1. After 2 hours, cells were visualized by confocal microscopy for the translocation of phosphorylated NF-κB p65 (pNF-κB p65) into the nucleus and the cytoplasmic and nuclear intensity of NF-κB. On day 12, the supernatants in parallel cultures were analyzed for the level of p24. As shown in Figure 5A, exposure of monocyte-derived macrophages to increasing concentrations of S-CM caused a dose-dependent decrease in NF-κB p65 translocation into the nucleus and a dose-dependent decrease in p24 production. However, when S-CM at an inhibitory concentration of 250 µg protein/mL was pre-incubated for 1 hour with anti-TGF-β antibodies at a concentration of 100 µg/mL, S-CM inhibition of both the nuclear translocation of NF-κB p65 and HIV-1 p24 production was reversed, whereas pre-incubation with irrelevant IgG (100 µg/mL) antibody had no effect on S-CM inhibitory activities (Figure 5B). Furthermore, incubation of the cells with activated, recombinant human TGF-β (rhTGF-β at a concentration of 10 pg/mL had little or minimal effect on NF-κB translocation or p24 production (Figure 5C). However, rhTGF-β 50 pg/mL, which approximates the concentration of TGF-β in S-CM 250 µg/mL, inhibited NF-κB translocation and activity, as well as p24 production, similar to that of S-CM 250 µg/mL (Figure 5C). Moreover, we previously showed (flow cytometry, ELISA, immunocytochemistry and Western blot) that LPS-exposed intestinal macrophages and S-CM-treated blood monocytes did not phosphorylate p65, had very low levels of p50, did not translocate p50 or p65 into the nucleus and expressed markedly reduced levels of NF-κB signal proteins (28). Expression of p50 and p65 genes also were markedly reduced in intestinal macrophages compared to autologous blood monocytes (28). These findings are consistent with minimal, if any, transcriptionally active p50/p65 heterodimer and together implicate stromal TGF-β-mediated down-regulation of NF-κB activation in the inhibition of HIV-1 replication by stromal factor-differentiated macrophages in vitro and intestinal macrophages in vivo.
We have shown that macrophages isolated from normal human small intestine are highly refractory to productive HIV-1 infection [13], [14], [15], supporting observations that memory CD4+ T cells rather than macrophages are the predominant mononuclear target cell in the intestinal mucosa during primary HIV-1 infection [19], [20], [21], [22], [23], [24], [25], [26]. We also have shown that in contrast to intestinal macrophages, vaginal macrophages are permissive to macrophage-tropic HIV-1 [13]. Since tissue macrophages throughout the body are derived from blood monocytes, our findings suggest that the lamina propria of the intestinal mucosa is a unique microenvironment capable of influencing HIV-1 permissiveness in blood monocytes recruited to the intestinal mucosa. Consistent with this concept, we present new evidence that products released by the intestinal extracellular matrix inhibit up-regulation of CD4 and CCR5 during the differentiation of blood monocytes into macrophages. However, the low level of CD4 and CCR5 expression on intestinal macrophages is not the exclusive cause of the cells' non-permissiveness to HIV-1, since (1) the very small subset (1%) of intestinal macrophages that express CD4, a proportion of which also express CCR5, did not support HIV-1 replication; (2) intestinal stromal products also decreased HIV-1 replication when stromal products were added simultaneously to cultures of monocyte-derived macrophages, i.e., before the induction of CD4 and CCR5 down-regulation; and (3) stromal products inhibited single-round gene expression of VSV-G pseudotyped virus, which enters cells independent of CD4 and CCR5. In this connection, we previously showed that unsorted intestinal macrophages with undetectable CD4 also do not of support HIV-1 replication (13, 14). Having previously shown that stromal TGF-β differentiates pro-inflammatory blood monocytes into non-inflammatory cells with the phenotype and function of intestinal macrophages [27] through Smad-induced IκBα expression and NF-κB signal dysregulation [28], we show here that a critical consequence of stromal TGF-β-induced NF-κB inactivation is the profound inability of monocyte-derived macrophages to support HIV-1 replication.
TGF-β is reported to both inhibit and stimulate HIV-1 replication, depending on the cell type, level of cell differentiation, virus strain, timing of treatment and presence of other cytokines [42], [43], [44]. In intestinal mucosa, latent TGF-β is produced by many different types of cells, including epithelial cells, mast cells, T regulatory cells, T cells undergoing apoptosis, and stromal cells. TGF-β constitutively released by these cells binds to the lamina propria extracellular matrix binding domains and upon activation and release regulates multiple macrophage defense and immune functions, consistent with an elaborate and finely tuned system of cross-talk that we have described previously [16]. Here we show that among these functions is the down-regulation of NF-κB activity and thus HIV-1 replication in monocyte-derived macrophages. These data suggest that TGF-β, at least in part, mediates the profound non-permissiveness of intestinal macrophages to HIV-1.
NF-κB plays a critical role in HIV-1 replication in T cells [41] and cells of the monocyte lineage [45]. In addition to stimulating the initiation of HIV-1 transcription [46], [47], [48], NF-κB also has been implicated in promoting HIV-1 transcriptional elongation [49], [50]. Importantly, NF-κB is constitutively activated in HIV-1-infected monocytes [51], possibly through upstream activation of the IKK complex by HIV-1 regulatory/accessory proteins [52], [53] or HIV-1-induced (via NF-κB activation) cytokines [54]. The activation of IKK leads to the phosphorylation and proteosomal degradation of IκBα and IκBβ, thereby releasing NF-κB for translocation into the nucleus to bind NF-κB-binding sites in the enhancer region of the HIV-1 long terminal repeat and host gene promotor sites. Thus, we conclude that stromal TGF-β inactivates NF-κB in monocyte-derived macrophages and that this inactivation likely contributes to the profound blockade in HIV-1 expression in intestinal macrophages, a highly unique population of mononuclear phagocytes [55], [56].
The HIV-1 non-permissiveness of intestinal macrophages due to NF-κB inactivation is consistent with our recent finding that stromal TGF-β dysregulation of NF-κB signaling causes inflammation anergy in intestinal macrophages [28]. Importantly, long-term culture of intestinal macrophages in the absence of stromal factors does not restore inflammatory capability [27], [28] and, as reported here, did not promote the emergence of HIV-1 permissiveness, indicating prolonged, if not permanent, down-regulation of these functions in intestinal macrophages. Also, exposure of intestinal macrophages to pro-inflammatory stimuli, including lipopolysaccharide (LPS), interferon-γ (IFN-γ) and tumor necrosis factor-α (TNF-α), does not induce inflammatory function [27], [28] and did not restore replication competence. These findings suggest that in primary HIV-1 infection, resident macrophages in healthy intestinal mucosa are incapable of de novo HIV-1 replication.
In contrast to primary HIV-1 infection, in late stage disease HIV-1-infected blood monocytes may recruit to intestinal mucosa that is either inflamed or infected with opportunistic pathogens. In such a microenvironment, dysregulated homeostasis permits viral replication to continue after the monocytes take up residence in the lamina propria, as we have reported for esophageal macrophages in patients with AIDS and opportunistic mucosal infections [57]. We also have reported that cytomegalovirus blocks stromal inhibition of HIV-1 infection of macrophages and that this inhibition is mediated, at least in part, by cytomegalovirus-induced monocyte production of TNF-α, which acts in trans to enhance HIV-1 replication [58]. However, the very low levels of TNF-α (<2.9 pg/mL) in S-CM generated from normal mucosa and inflamed Crohn's mucosa [59] suggest that TNF-α is not involved in stromal down-regulation of intestinal macrophage permissiveness to HIV-1.
In the present study, we investigated HIV-1 permissiveness in intestinal macrophages using highly macrophage-tropic R5 viruses, including NA420 B33, NA20 B59 and NA353 B27 [39], [59], in order to maximize the possibility of infecting intestinal macrophages. Interestingly, infectious molecular clones of transmitted founder viruses derived from acutely infected persons are R5-tropic but fail to replicate efficiently in monocyte-derived macrophages [60], [61]. Although we have not yet examined the ability of these molecular clones to infect intestinal macrophages, such infection seems unlikely, since intestinal macrophages do not activate NF-κB, a requirement for HIV-1 gene transcription during macrophage differentiation [45].
The findings presented here do not exclude the possibility that HIV-1 restriction factors other than TGF-β are present in the stroma and thus S-CM. S-CM was used in a range of 10-1000 µg protein/mL, corresponding to TGF-β in the range of 1–150 pg/mL. Although rhTGF-β at a concentration of 10 pg/L had little or minimal effect on NF-κB translocation or p24 production (Figure 5C), rhTGF-β 50 µg/mL, which approximates the concentration of TGF-β in S-CM 250 µg/mL, inhibited NF-κB translocation and viral replication (p24 production), similar to that of S-CM 250 µg/mL. Apolipoprotein B mRNA-editing enzyme catalytic polypeptide-like 3G (APOBEC3G), which causes dC-to-dU mutations in viral DNA, is reported to be induced by LPS in dendritic cells and by IFN-α in monocyte-derived macrophages [62], [63]; however, we have been unable to detect APOBEC3G in resting or IFN-α-treated intestinal macrophages. Also, higher levels of anti-HIV-1 miRNAs have been reported to inhibit HIV-1 in monocytes [64], [65], but the role of miRNA as a restriction factor in monocytes is controversial [66], [67]. A cellular restriction factor that is neutralized by primate lentiviral Vpx protein was recently detected in quiescent monocytes, but its reduction as the cells differentiate into macrophages makes it an unlikely restriction factor in terminally differentiated intestinal macrophages [68]. Other potential restriction factors, including p21 [69], [70] and interferon-induced C/EBPβ [71], [72], have been proposed but have not yet been investigated in mucosal macrophages.
A confounding issue regarding post-entry restrictions in intestinal macrophages is that such restrictions would be unique to macrophages residing in the intestinal mucosa, since macrophages in the vaginal mucosa are highly replication competent [13]. Although the extracellular matrix could release products that induce yet-to-be-identified anti-viral restrictions, the findings presented here implicate stromal TGF-β-induced NF-κB inactivation as contributing to the non-permissiveness of macrophages in the human small intestine. These findings help explain the overwhelming absence of productive infection in intestinal macrophages, in sharp contrast to the highly productive infection in intestinal T cells, in primary HIV-1 infection. The ability of intestinal CD4+ T cells to support robust HIV-1 replication is well established in our in vitro [13] and in vivo studies [9], [10], [11], [12], [13], [14], [15], [16], [17], [18], [19], [20], [21], [22], [23], [24], [25], [26]. Furthermore, TGF-β does not inhibit HIV-1 expression in a chronically infected T cell line or in primary T cell blasts infected in vitro with HIV-1 [42]. The discordance between intestinal T cell and macrophage support for HIV-1 replication in the presence of down-regulatory stromal TGF-β is currently under investigation in our laboratory. Thus, the unique dysregulation in NF-κB signaling induced in monocytes by extracellular matrix products, especially TGF-β, when the cells take up residence in the intestinal mucosa, offers a mechanism by which the host down-regulates mucosal macrophages for harmful pro-inflammatory responses and permissiveness to viruses in which transcription is NF-κB-dependent. Harnessing this natural anti-viral defense mechanism may provide a novel strategy to exploit for the prevention of infection in HIV-1 permissive cells.
All tissue and cell protocols were approved by the Institutional Review Board of the University of Alabama at Birmingham. Written informed consent was provided by study participants.
Macrophages were isolated from segments of intestinal mucosa of otherwise healthy subjects undergoing elective gastric bypass by enzyme digestion and purified by counterflow centrifugal elutriation, as previously described [73], [74], [75]. Circulating blood monocytes from the same donors were purified by gradient sedimentation followed by magnetic anti-PE bead isolation of anti-CD14-PE-treated cells per the manufacture's manual (Miltenyi Biotec). All studies were performed using fresh cells. Macrophages and monocytes were routinely >98% pure and 98% viable by propidium iodide staining. CD4+ monocytes and intestinal macrophages were isolated by magnetic CD4+ microbead separation.
Macrophage-tropic viruses were prepared as previously described [13], [76], [77]. Briefly, replication competent clones of highly macrophage-tropic R5 viruses, including NA420 B33, NA20 B59 and NA353 B27 [39], were transfected into 293T cells by Fugene 6 (Roche), according to the manufacture's protocol. After 60 hours, the supernatants were harvested, clarified by low speed centrifugation (1,000 g, 10 minutes), filtered (0.45 µm filter), titrated using JC53BL cells [78], aliquoted and stored at −80°C.
YU2 envelope (Env) or vesicular stomatitis virus glycoprotein (VSV-G) HIV-1 pseudovirions that express GFP upon infection were kindly provided by D. Levy, NYU and constructed as follows. Briefly, the env gene was deleted and the gfp gene was inserted between the env and nef genes of the pNL4-3 clone. An internal ribosome entry site (IRES) element was inserted between the gfp and nef genes to rescue nef gene expression [79]. To generate the YU2 Env or VSV-G GFP reporter pseudovirions, the clone was co-transfected with the YU2 Env or VSV-G expression plasmid into 293T cells and harvested, as described above.
Using our previously described protocols [40], [73], [74], epithelium and lamina propria mononuclear cells (MNLs) were removed by enzyme digestion from segments of normal human jejunum from otherwise healthy subjects undergoing elective gastric bypass, and purified by elutriation. The epithelial cells (EC) (10×106/mL), lamina propria MNLs (10×106/mL), and cell-depleted lamina propria stroma (1 g wet wt stromal tissue/mL), respectively, were cultured in RPMI for 24 hours without serum, and the EC-conditioned media (EC-CM), MNL-CM and stroma-CM (S-CM) were harvested, sterile-filtered (0.2 mm Syringe Filter; Corning Inc.) and frozen at −70°C, as previously described [27], [28]. Cell depletion from lamina propria stroma was confirmed by immunohistochemistry [73]; intestinal macrophages expressed barely detectable CD14 [13]. Conditioned media did not alter monocyte-derived macrophage viability during incubation for as long as 4 days as assessed by flow cytometric analysis of propidium iodide uptake. S-CMs were normalized to 500 µg/mL RPMI. Endotoxin and protein content were determined by ELISA (endotoxin ELISA: Cambrex Bio Science; protein ELISA: Pierce Protein Research Products/Thermo Scientific). Only endotoxin-free EC-CM, MNL-CM and S-CM were used in the experiments.
Intestinal macrophages and monocytes were incubated with optimal concentrations of PE-, APC-, or FITC-conjugated antibodies to HLA-DR, CD13, CD4, CCR5 (BD Pharmingen), or control mAbs of the same isotype at 4°C for 20 minutes, washed with PBS, fixed with 1% paraformaldehyde and analyzed by flow cytometry. Data were analyzed with FlowJo software (Tree Star, Inc.). To examine the effect of S-CM on CD4 and CCR5 expression in monocyte-derived macrophages, blood monocytes were cultured in 48-well plates at 5×105 cells/well in RPMI plus macrophage colony-stimulating factor (M-CSF) serum and S-CM at final concentrations of 0, 10, 100 and 500 µg/mL for up to 3 days and analyzed for CD4 and CCR5. Student's t-test was used to determine the statistical significance of the difference of expression levels of these receptors between intestinal macrophages and autologous blood.
Sorted intestinal macrophages and monocytes from 2 donors were cultured in triplicate in 96-well plates at 2×105 cells/well in RPMI plus M-CSF and serum for 4 days. Cultures then were inoculated with NA20 B59, NA353 B27 or NA420 B33 at an MOI = 1, cultured for the indicated duration with 100 µL of supernatant, harvested every 4 days and stored at −70°C until assayed for p24 by ELISA (PerkinElmer).
To examine the effect of S-CM on macrophage permissiveness to HIV-1, MACS-sorted monocytes were cultured for 4 days in RPMI plus M-CSF to generate monocyte-derived macrophages, after which S-CM was added at final concentrations of 10, 100 and 500 µg protein/mL. Control cultures of monocyte-derived macrophages were incubated in media alone. Two days later, culture supernatants were removed, and triplicate cultures were inoculated with NA353 B27 (MOI = 1) for 2 hours, cultured for 20 days, and the kinetics of p24 production was determined as above. Parallel triplicate cultures of monocyte-derived macrophages were inoculated simultaneously with NA353 B27 (MOI = 1) plus S-CM (final concentrations of 0, 10, 100 and 500 µg protein/mL) for 2 hours, and viral replication was monitored as above.
Cultures of monocyte-derived macrophages prepared as above were inoculated with NA353 B27 (MOI = 1) plus S-CM or with S-CM only. Cells treated with S-CM only were harvested after 2 hours, cytospun onto glass slides and stained for NF-κB p65. Cells infected with virus were cultured, and supernatants were harvested on day 12 and assayed for p24 by ELISA. Parallel monocyte-derived macrophages were inoculated for 2 hours in triplicate with NA353 B27 (MOI = 1) plus S-CM 250 µg protein/mL pre-treated with 0, 25 or 100 µg/mL of anti- TGF-β for 1 hour at 37°C. Analysis of viral replication and NF-κB p65 staining were performed as above.
A final aliquot of monocyte-derived macrophages prepared as above was cultured for 6 days, inoculated in triplicate with NA353 B27 (MOI = 1) plus rhTGF-β (R&D Systems) or rhTGF-β only at final concentrations of 0, 10, or 50 pg/mL for 2 hours. Evaluation of NF-κB p65 intensity and viral replication were performed as above.
Cells cytospun onto glass slides were fixed and permeabilized with Cytofix/Cytoperm (BD Biosciences) for 20 minutes. After washing with PBS, cells were blocked with casein protein (DAKO) for 1 hour and incubated with rabbit anti-NF-κB p65 or isotype control antibodies (Santa Cruz Biotechnology) for 90 minutes, washed with PBS, incubated with donkey anti-rabbit IgG-FITC (Jackson ImmunoResearch Laboratories) for 30 minutes, washed with PBS and counterstained with DAPI nuclear stain. Cells were visualized by confocal microscopy, and the cytoplasmic and nuclear fluorescence intensity of NF-κB was converted to histograms using IPLab image analysis software version 3.6 (BD Biosciences Bioimaging).
For comparison of the effects of treatment on NK-κB activity, NF-κB intensity was normalized to the blue signal in the nucleus. Five images were analyzed per sample and mean intensities were generated. For comparison of the effects of treatment on HIV-1 replication, p24 value of each treatment was normalized to the media control group with the replication level of the media control group defined as 100%. Statistical significance was determined by Student's t-test.
Data is expressed as mean ± SD or ± SEM, and statistical significance between groups was determined using Student's t-test. P values ≤0.05 were considered significant.
|
10.1371/journal.pntd.0007100 | Clinical and molecular epidemiology of Crimean-Congo hemorrhagic fever in Oman | Crimean-Congo hemorrhagic fever (CCHF) is a serious disease with a high fatality rate reported in many countries. The first case of CCHF in Oman was detected in 1995 and serosurveys have suggested widespread infection of humans and livestock throughout the country.
Cases of CCHF reported to the Ministry of Health (MoH) of Oman between 1995 and 2017 were retrospectively reviewed. Diagnosis was confirmed by serology and/or molecular tests in Oman. Stored RNA from recent cases was studied by sequencing the complete open reading frame (ORF) of the viral S segment at Public Health England, enabling phylogenetic comparisons to be made with other S segments of strains obtained from the region.
Of 88 cases of CCHF, 4 were sporadic in 1995 and 1996, then none were detected until 2011. From 2011–2017, incidence has steadily increased and 19 (23.8%) of 80 cases clustered around Eid Al Adha. The median (range) age was 33 (15–68) years and 79 (90%) were male. The major risk for infection was contact with animals and/or butchering in 73/88 (83%) and only one case was related to tick bites alone. Severe cases were over-represented: 64 (72.7%) had a platelet count < 50 x 109/L and 32 (36.4%) died. There was no intrafamilial spread or healthcare-associated infection. The viral S segments from 11 patients presenting in 2013 and 2014 were all grouped in Asia 1 (IV) lineage.
CCHF is well-established throughout Oman, with a single strain of virus present for at least 20 years. Most patients are men involved in animal husbandry and butchery. The high mortality suggests that there is substantial under-diagnosis of milder cases. Preventive measures have been introduced to reduce risks of transmission to animal handlers and butchers and to maintain safety in healthcare settings.
| Crimean-Congo hemorrhagic fever, an often fatal tick-borne viral disease, has made an impact in the Sultanate of Oman—affecting nationals and expatriates alike—for the past 20 years. In this retrospective review of the epidemiology and outcomes of cases in Oman from 1995 to 2017, we identified 4 sporadic cases in 1995 and 1996, then none until 2011, followed by a steady increase until 2017. The mortality rate of 32 of 88 cases (36.4%) is high in comparison to studies from other countries and this could be explained by under-diagnoses of milder cases in the Sultanate. Transmission is commonly associated with animal husbandry and butchering and 88% cases were infected by contact with animals, whereas transmission by tick bite is more commonly recorded in some countries. A proportion of cases (23.8%) were clustered around the Eid-Al-Ahda festival which has, from 2011–2017, occurred in the summer months, which have a higher risk of transmission. This additional risk has been noted and preventive measures have been introduced to reduce the risk of transmission to animal handlers and butchers.
| Crimean-Congo hemorrhagic fever (CCHF) is a serious and often fatal infection caused by the CCHF virus (CCHFV). Ixodid ticks, especially Hyalomma spp, act as both reservoirs and vectors. This virus has the greatest geographic range of any tick-borne virus and there are reports of viral isolation and/or disease from more than 30 countries in Africa, Asia, Eastern and Southern Europe, and the Middle East [1–3]. Numerous domestic and wild animals, such as cattle, goats, sheep and small mammals, such as hares and rodents, serve as asymptomatic amplifying hosts for the virus [4].
CCHFV can be transmitted between animals and humans by Hyalomma ticks. It can also be transmitted by direct contact with blood and other body fluids of viremic humans and animals and has the potential to cause population-based outbreaks [5–6]. Clinical features commonly include fever of abrupt onset, myalgia, headache and thrombocytopenia, and can progress to hemorrhage, multiorgan failure and death. The levels of liver enzymes, creatinine phosphokinase, and lactate dehydrogenase are raised, and bleeding markers are prolonged [7–8]. The crude mortality rate of CCHF differs from country to country, ranging from 2–80% [1]. Early diagnosis and supportive management are essential for a favorable outcome.
CCHFV is a negative-sense single-stranded RNA virus classified within the Orthonairovirus genus of the Nairoviridae family. The CCHFV genome is comprised of single-stranded negative-sense RNA divided into 3 distinct segments designated small (S), medium (M), and large (L). Comparisons of full S segment sequences have shown that CCHFV forms 7 distinct clades, each with strong geographical associations [1,9–11]. Subtle links between distant geographic locations, shown by phylogenetic analysis, may have originated from the international livestock trade or from long-distance carriage of CCHFV by infected ticks via bird migration [5,10,12–13].
Oman is situated in the southeastern corner of the Arabian Peninsula, bordering the Kingdom of Saudi Arabia, United Arab Emirates, and Yemen. The summer is hot and humid with temperatures reaching as high as 49°C and the winter relatively cooler and with some rain. The total population is 4,615,269 individuals, of whom 54.6% are Omanis and the remainder are expatriates [14].
Cases of CCHF were first detected in Oman in 1995 when there were 3 unrelated sporadic cases, followed by a further case in 1996 [15, 16]. Cases related to animal movement and slaughter were also reported the following year from Western Saudi Arabia [17–18] and from the UAE [15,19, 20], where an imported case had previously resulted in fatal infections of health care workers in 1979 [21]. A survey conducted in 1996 in Oman revealed asymptomatic seropositivity for CCHFV exposure in 1/41 (2.4%) of Omanis compared to 73 (30.3%) of 241 non-Omani citizens with occupational animal contact [22]. However, no further human cases of CCHF were reported in Oman until 2011 [23], and since then there has been a steady increase [24]. A recent survey has shown infection in a variety of animals and ticks in Oman [25].
Limited data are available on the prevalent clade(s), or group(s) of organisms from a single ancestor, of CCHFV in the Arabian Peninsula. Sequencing of the S, M, and L segments of CCHFV isolated from the 1996 patient in Oman (recorded as Oman 1997 in GenBank) showed that it belonged to Asia lineage 1 (clade IV) [10,26], as was the virus isolated from a patient who returned to India with CCHF acquired in Oman in 2016 [27]. Virus isolates from 4 patients in the UAE in 1994 and 1995 also align with the Asia 1 (clade IV) lineage, as did contemporaneous isolates from Hyalomma ticks obtained from livestock imported into the UAE from Somalia [9,19,28]. A further human isolate in the UAE in 1994/95 aligned with lineage Africa 1 (clade III) [9, 19, 28].
The aims of this study are to describe the clinical and epidemiological features and outcomes of cases of CCHF diagnosed in Oman between 1995 and 2017. We also investigated the local molecular epidemiology of CCHFV by partial and complete S segment sequencing of stored CCHFV isolates from patients recently diagnosed in Oman.
A retrospective descriptive record-based review and analysis of CCHF cases was conducted over the period 1995 through 2017. CCHF has been listed as a notifiable disease in Oman since 1995 and surveillance forms from suspected cases are submitted by all healthcare providers to the Communicable Diseases Department at Ministry of Health headquarters. Blood samples obtained from suspected cases were submitted at the same time to the Central Public Health Laboratory (CPHL) at the MoH in Muscat, Oman. All CCHF cases reviewed and included in this study were detected by this routine communicable disease surveillance combined with the CPHL results during the study period.
A generic national form is used for initial notification of a suspected case of CCHF; once the diagnosis is confirmed, a more detailed form is submitted that includes patient identifiers, demographic and geographic variables, relevant exposure history, key clinical features, and some clinical laboratory test results. The form is not unique to CCHF, so prompts for some specific CCHFV-related exposures and laboratory variables are missing. The report format remained similar until 2017 when the paper form was replaced by an electronic version.
Data were systematically extracted from the surveillance forms of all laboratory confirmed cases of CCHF. Demographic variables included age, sex, nationality, location, and date of notification. Risk factors included history of tick bite, occupational exposure and contact with tissues, blood or other biological fluids from an infected animal, or contact with a case within 14 days prior to the onset of symptoms. Clinical data included presence/absence and duration of fever, headache, myalgia, nausea, vomiting, diarrhea, petechial rash, and bleeding from sites including gums, nose, lung, gastrointestinal tract, or skin. Key laboratory variables include platelet counts, hemoglobin, urea and electrolytes, and liver function tests. The main clinical outcomes were death or survival.
The case definition for a suspected case in Oman is: an illness with sudden acute onset with the following clinical findings: a fever ≥ 38.5°C (> 72 hours to < 10 days) associated with severe headache, myalgia, nausea, vomiting, and/or diarrhea; thrombocytopenia < 50 x 109/L; hemorrhagic manifestations which develop later and may include petechial rashes, bleeding from the gums, nose, lungs, gastrointestinal tract, etc.; history of tick bite, occupational exposure, contact with fresh tissues, blood, or other biological fluids from an infected animal [24].
Good laboratory practice and a high level of effective biosafety precautions are required by laboratory staff handling materials from suspected CCHF cases due to the potential for sample-to-person, or indirect, transmission [6,29]. Blood samples collected from suspected cases of CCHF admitted to all MoH and non-MoH health care institutions in Oman are sent to CPHL in triple pack containers, using the most direct and timely route available. These samples are considered urgent and results are provided within 24 hours of their arrival at CPHL. National guidelines are in place to instruct local laboratories where suspected cases are admitted on safe handling of all material collected for any diagnostic purpose [24,30].
Both serum and plasma samples are requested for CCHFV testing. Plasma is preferred for molecular testing using a commercial CCHFV real time reverse transcription polymerase chain reaction (rRT-PCR) kit (In vitro Diagnostics, Liferiver Shanghai ZJ Bio-Tech Co., Ltd. Shanghai, China). At CPHL, the plasma extraction takes place inside a gloved box using a manual extraction system, QIAamp Viral RNA Kit (QIAGEN, Hilden, Germany). The samples are first treated with AVL buffer (QIAGEN) to inactivate infectious viruses and RNases. Intact viral RNA is then purified by selective binding and washing steps. The screening RT-PCR reaction is based on a one step real-time RT-PCR. Briefly, CCHFV RNA is converted into cDNA and a thermostable DNA polymerase is used to amplify specific CCHFV S segment sequence targets by standard thermocycling in a PCR as per manufacturer instruction. The kit contains an internal control to identify possible PCR inhibition.
A positive result from a RT-PCR screening for CCHFV RNA is used to confirm infection. In such cases, the serum sample is not tested further. If the RT-PCR is negative, heat inactivated serum (56°C water bath for 30 minutes) is tested for CCHFV antigen and IgM and IgG antibodies using a commercial kit (Vector-Best, Novosibirsk, Russia). For samples that are negative for all parameters, a convalescent serum is requested for CCHF IgG testing. The CPHL takes part in regular internal and external quality assurance reviews in association with WHO EMRO and WHO Quality Management Standards.
All available stored serum samples, collected from 21 CCHF patients in 2013 and to 2014, were inactivated with AVL buffer and sent to PHE Porton Down, England, UK. At PHE, AVL samples were processed with a standard QIAamp Viral RNA Kit. Eluted RNA was evaluated for the presence of CCHFV RNA using an in-house RT-PCR assay [31]. Sequencing was performed using standard CCHFV S segment sequencing primers as described previously [32]. Assembled sequence data for the S segment of each sample were manipulated and analyzed using the Lasergene suite of programs (DNAStar, Maddison, WI, USA). For phylogenetic analysis, sequences were aligned using the Clustal W computer program (The European Bioinformatics Institute, Wellcome, UK) [33] and output in PHYLIP Format (scikit-bio). To construct maximum-likelihood phylogenetic trees, quartet puzzling was applied using the program, Tree-Puzzle, at the Institut Pasteur [34, 35]. The Tamura-Nei model of substitution was adopted, as has been performed in other phylogenetic studies demonstrating reassortment [36]. Phylogenetic trees were drawn using the program TreeView (JAM Software GmbH, Trier, Germany) [37]. The values at the tree branches represent the puzzle support values. S segment sequences were submitted to GenBank.
The data analysis was conducted at the MoH Department of Surveillance in Muscat. A descriptive analysis compared age, sex, nationality, location, and date of cases. Risk factors and clinical and laboratory parameters were also tabulated. Missing data items (positive or negative) were omitted from analysis. Statistical comparisons were performed using SPSS 11.0 package program (SPSS Inc, Chicago, IL, USA).
Ethical approval was sought from the MoH, Oman. The study is considered free from ethical constraints as it is a secondary analysis of the data collected routinely for the purpose of public health surveillance and reporting. No personal identifying information accompanied the samples sent to PHE.
A total of 88 cases were reported between 1995 and 2017. Of these, 82 (93.2%) were confirmed by RT-PCR and 4 by CCHFV IgM alone. Two further probable cases (both fatal) in 2011 and 2016 were included on the basis of typical clinical and laboratory features as per electronic records. There were 3 isolated cases in January, May, and June 1995 with a further case in 1996, and then no cases were reported until 2011. Since then, there has been a steady increase in numbers, peaking at 20 cases in 2015 (Fig 1). Annual notifications of suspected cases were not systematically recorded until 2011 and data about notifications of suspected cases and possible missed cases are incomplete. In the years 2001 to 2011 inclusive, there were 35 notifications of possible VHF cases, of which 2 were proven CCHF (in 2011) and at least 23 were confirmed to be cases of dengue (2 fatal).
The patients had a median (range) age of 33 (15–68) years and 79 (90%) were male. The most common nationality affected was Omani 51 (59%) followed by Bangladeshi 18 (21%), Pakistani 7 (8%), Yemeni 3 (4%), Indian 4 (5%), Somali 2, and Sri Lankan 1. Cases occurred in all governorates (wilayats) except Musandam and Al Wustah (Fig 2). There was no geographic or source-related clustering of cases; however, several cases followed Eid Al Adha, a festival associated with animal sacrifice. In the years 2013–2017, 19/80 (23.8%) of all cumulative cases had their onset within 3 weeks after Eid Al Adha (Fig 3). There was also a smaller peak of cases in the spring weeks 6–19 (Fig 3). The main exposure risk identified was animal/fresh tissue exposure in 73/88 (83%), with only 1 case attributed to tick bite alone. Exposure risk was not identified in 14 (15.9%) (Table 1).
Clinical features in 88 patients included fever in 80 (90.9%), hemorrhagic features 41 (46.6%), vomiting 32 (36.4%), myalgia 30 (34.1%), diarrhea 20 (22.7%), respiratory symptoms 17 (19%), abdominal pain 11 (12.5%), other symptoms in 29 (33%). Severe thrombocytopenia (platelet count < 50 x 109/L) was present in 64 (72.7%). There were 32 deaths, resulting in a cumulative case fatality rate of 36.4%. The case fatality rate in Omanis was 16/53 (30.2%) and in Bangladeshis was 10/18 (55.6%) (P>0.05).
Of the 21 serum samples that were sent to PHE, 20 were RT-PCR positive using an in-house assay. However, of these, only 12 samples provided suitable cycle threshold values (the cycle threshold being 28 or under) to warrant further sequencing of CCHFV S segments and only 12 samples provided sequencing data which spanned the entire ORF of the S segment. Sequence data have been submitted to GenBank and sequences have been assigned the following accession numbers: MH037279 (Oman 2012-40S), MH037280 (Oman 2013-116S), MH037281 (Oman 2014-828P), MH037282 (Oman 2014-979P), MH037283 (Oman 2014-602P), MH037284 (Oman 2013-825P), MH037285 (Oman 2013-92S), MH037286 (Oman 2013-108S), MH037287 (Oman 2013-179P), MH037288 (Oman 2014-860P), MH037289 (Oman 2014-624S), and MH037290 (Oman 2014-747P). Sequences were compiled with a range of other CCHFV S segment ORF sequences and used to make the maximum likelihood phylogenetic tree shown in Fig 4.
This report summarizes the clinical, epidemiological, and virological findings in 88 people with symptomatic CCHF throughout the Sultanate of Oman in the past 2 decades. Cases were detected by passive surveillance, starting with a few sporadic reports in 1995 and 1996, followed by no cases until 2011. Since then there has been a sustained increase in yearly reports, of which 19/80 (23.8%) have clustered around the Eid Al Adha festival, occurring in summer months in the years 2013 to 2017, with a possible smaller peak in the spring months. Ninety percent of all patients were male with a median age of 33 years. Both Omanis and citizens of other nationalities were affected, the predominant risk factors being exposure to animals and meat products, especially involvement in butchering or slaughtering. Diagnosis was confirmed by RT-PCR in 82 (93.2%) cases and by serology alone in 4 (4.5%). Stored viral RNA from 12 patients presenting in 2013 and 2014 was sequenced for the entire S segment ORF of each of the 12 samples, all grouped in the Asia 1 (IV) clade. The cumulative mortality was 36.4%, and no cases of healthcare related or intrafamilial spread of infection were reported.
This is the largest series of cases of CCHF reported from a GCC country, and provides the first data about locally prevalent strains of CCHFV in almost 20 years. The findings raise a number of questions about the origin and distribution of CCHFV in Oman and neighboring countries, the reasons for the high observed mortality, and the appropriate human and veterinary public health responses in Oman and other GCC states.
The clinical features of the cases were similar to those reported in other countries [13,38–40]. The most common symptoms reported were fever, fatigue, headache, loss of appetite, myalgia, and abdominal pain. Hemorrhagic manifestations were described in 34/67 (50.8%) and severe thrombocytopenia (< 50 x109/L) was present in 64/88 (72.7%) at presentation. There is no internationally agreed case definition for CCHF, but at least 3 scoring systems to assess severity of illness have been proposed [41–43]. Mortality is known to be associated with older age, presence of underlying illness, and CCHFV viral load at presentation [8]. Case numbers were too small to show a link with mortality in our series and details about the latter 2 risks were not recorded. Representative case fatality rates elsewhere include 5% in Turkey, 17.6% in Iran, and 15% in Pakistan [1,44–45]. However, the CFR of 15% in Pakistan was reported from a center with substantial experience, whereas overall mortality rates of up to 41% have been reported more recently in Pakistan, especially during outbreaks [46].
The lower mortality in Turkey and Iran could be explained by improved surveillance and early diagnosis of CCHF in patients with fever and thrombocytopenia, following prolonged campaigns to raise awareness in both healthcare personnel and the general public in those countries. A serosurveillance study conducted in Oman in 1996 showed that none of the 74 antibody-positive individuals identified recalled ever being hospitalized for an illness resembling CCHF with associated fever and bleeding, suggesting that there is a substantial incidence of subclinical CCHF human infections in Oman [3, 12–13,22]. Serosurveillance studies in other countries have shown seroprevalence rates of approximately 10–13% in high risk human populations [5, 47]. Based on mortality data from Turkey we believe that under-diagnosis of mild cases has skewed the mortality data in Oman.
This is the first study to describe the complete sequence of the S segment ORFs of a series of CCHFV isolates from the region. The results largely confirm findings from partial sequencing of sporadic isolates from the UAE [19, 28] and Oman [10,27] since the mid 1990’s. The phylogenetic relationship of these sequences with other published sequences from the region is depicted in Fig 4. The similarity of all these sequences to the human cases in Oman and the UAE over 2 decades is striking, and these sequences align with those from Pakistan. In the future, with the advance of cheaper sequencing technologies, it will be valuable to compare full length genomes from multiple locations with clinical data. This may help address hypotheses about alternative strain pathogenicity, including the relative contribution of segment reassortment in CCHF disease [10, 48].
The average age of patients in Oman was 33 years, and 90% were men. Most infections were acquired while butchering or slaughtering animals or from other close animal tissue and blood exposure as in earlier and more recent cases in Dubai [20,49] and the Kingdom of Saudi Arabia [17]. This contrasts with the situation in Turkey, Kazakhstan, and Iran, where tick bites are the most commonly reported risk factor [11, 39–40,44]. Slaughtering animals during Eid Al Adha is known to pose a particularly high risk of infection [5]. Sporadic unregulated slaughtering without using appropriate personal protective equipment still occurs in Oman during Eid Al Adha. It is also common for non-professional individuals to become involved and for butchers to freelance, going from house to house to sacrifice animals, as people find it more convenient to have the sacrifice performed at farms and backyards. During skinning and subsequent tanning of the hides, ticks can bite humans. We examined the effect of Eid as a possible cause for the apparent increase in cases over the past 7 years as the festival has moved back into the summer months when tick activity is most prominent. However, this is not the only factor. Fig 3 demonstrates that there is the expected clustering of cases after Eid Al Adha, but this accounts for only 23.6% of the total cases. Similar findings have been reported in Pakistan [50] and the data suggest that climatic factors affecting tick activity are most important in promoting seasonal variation in human infection risk together with extra added risk at the time of Eid.
These data change our perceptions about the duration and origin of CCHFV activity in Oman and neighboring countries. Previously, it had been postulated that sporadic cases were related to the importation of infected livestock from other countries. A large amount of livestock is imported to Oman every year: in 2016, over 1.5 million farm animals were imported, including sheep (85.2%), cattle (7%), and goats (5.6%). The origins of these included Armenia, Australia, Djibouti, India, Iran, Jordan, Pakistan, Somalia, Sudan, Turkey and the UAE. A 21-day quarantine procedure is in effect for animals arriving from other countries by sea or land. Once in Oman, animals are distributed to sales centers, feedlots, and distribution points throughout the country. Livestock are held in large holding pens and not segregated according to country of origin or time from entry into Oman. Spread of infection could result from unrestricted entry of tick-infested and potentially viremic domestic animals during religious holidays; the abundance of virus-infected ticks within stockyards and holding pens; the uncontrolled movement of livestock animals infested with CCHFV-carrying Hyalomma ticks to ranches, farms, and markets throughout the country; and the indiscriminate mixing and crowding of tick-infested and potentially viremic animals with uninfected and tick-free animals [3,13,22].
There was partial support for the possibility of intermittent importation of CCHFV with livestock into the UAE in the 1990s, where ticks were found on animals with different clades of CCHFV S segment corresponding to African as well as Asia 1 clades [9,19,28]. However, human serosurveys in Oman in 1996 [22] and the finding of the virus in ticks and animals throughout Oman in 2013–2014 [25] suggest that all areas of the Sultanate have had a substantial burden of CCHFV infection for at least 2 decades, probably related to all the risks mentioned above. Moreover, all virus isolates from humans in Oman and the UAE have had remarkably similar S Segments, apart from the nosocomial outbreak in Dubai in 1979 from an Indian index case [21]. In contrast, several different S segments are circulating in Somalia, Iran, and Turkey [9,10,51]. This suggests that the Asia 1 S segment of CCHFV has been circulating in Oman for more than 20 years. It will be of interest to fully sequence the complete genomes of the recent isolates from ticks in Oman (and elsewhere) to explore this hypothesis further [25]. Reports of CCHFV antibody positivity in earlier human serosurveys in Kuwait [52] and intermittent occupational-related outbreaks in the UAE [19, 20, 28], and KSA [17,18], since then suggest that this is also the case throughout GCC countries.
The Oman MoH has undertaken a number of activities and initiatives to educate and inform the public about the risks of CCHF infection associated with slaughtering. A joint strategic initiative was developed in collaboration with the Ministry of Agriculture and Fisheries and Ministry of Municipalities and Water Resources. Education and information on prevention of CCHF in different languages has been targeted at those involved in slaughtering and handling animals. This includes placing advertisements on social media platforms, TV, radio, billboards, magazines, and newspapers before and during Eid Al Adha. Knowledge about CCHF is increasing in Oman with hospitals now following guidelines for the management of suspected cases of CCHF [53]. In addition, guidelines have been produced for culturally acceptable safe burials [24, 54]. It is reassuring that no healthcare related infections were detected in this series.
The data suffer from the limitations of a retrospective study that spans over 20 years, based on notifications of suspected illness and laboratory reports. In particular, the completeness of notification has been highly variable and is likely to have underestimated the incidence of symptomatic infections. Data on notifications of suspected cases that later turned out to be negative for CCHFV have not been systematically recorded and it is likely that the gap in notified cases between 1997 and 2011 is due to missed diagnoses and underreporting, rather than absence of cases. However, the records of confirmed cases at CPHL are thought to be complete. Conversely, the increase in notifications since 2011 may be due to a genuine increase in cases and/or be due to increased physician awareness and hence case recognition and reporting.
This is the largest reported series of CCHF from any of the GCC countries to date and brings together all published viral sequences in this region. The implication is that CCHF is endemic and under-recognized in Oman and surrounding countries and that prospective studies are needed to determine how often less severe cases of fever and thrombocytopenia are presenting in Oman. Proven and suspected cases have been reported in expatriate travelers returning from Oman to India [27] and Pakistan [55] and the possibility of CCHF should be considered in febrile travelers arriving from GCC countries, especially if they have been involved in animal slaughtering [56]. Oman has responded by improving its notification systems and laboratory support. Active local and regional programs of health promotion and human illness prevention need to be maintained together with surveillance and control of infection in animals and local tick vectors.
|
10.1371/journal.pntd.0005602 | Using G6PD tests to enable the safe treatment of Plasmodium vivax infections with primaquine on the Thailand-Myanmar border: A cost-effectiveness analysis | Primaquine is the only licensed antimalarial for the radical cure of Plasmodium vivax infections. Many countries, however, do not administer primaquine due to fear of hemolysis in those with glucose-6-phosphate dehydrogenase (G6PD) deficiency. In other settings, primaquine is given without G6PD testing, putting patients at risk of hemolysis. New rapid diagnostic tests (RDTs) offer the opportunity to screen for G6PD deficiency prior to treatment with primaquine. Here we assessed the cost-effectiveness of using G6PD RDTs on the Thailand-Myanmar border and provide the model as an online tool for use in other settings.
Decision tree models for the management of P. vivax malaria evaluated the costs and disability-adjusted life-years (DALYs) associated with recurrences and primaquine-induced hemolysis from a health care provider perspective. Screening with G6PD RDTs before primaquine use was compared to (1) giving chloroquine alone and (2) giving primaquine without screening. Data were taken from a recent study on the impact of primaquine on P. vivax recurrences and a literature review. Compared to the use of chloroquine alone, the screening strategy had similar costs while averting 0.026 and 0.024 DALYs per primary infection in males and females respectively. Compared to primaquine administered without screening, the screening strategy provided modest cost savings while averting 0.011 and 0.004 DALYs in males and females respectively. The probabilistic sensitivity analyses resulted in a greater than 75% certainty that the screening strategy was cost-effective at a willingness to pay threshold of US$500, which is well below the common benchmark of per capita gross domestic product for Myanmar.
In this setting G6PD RDTs could avert DALYs by reducing recurrences and reducing hemolytic risk in G6PD deficient patients at low costs or cost savings. The model results are limited by the paucity of data available in the literature for some parameter values, including the mortality rates for both primaquine-induced hemolysis and P. vivax. The online model provides an opportunity to use different parameter estimates to examine the validity of these findings in other settings.
| A single infection with Plasmodium vivax can cause multiple episodes of illness due to dormant liver parasites called hypnozoites. Primaquine is the only drug currently available to treat hypnozoites but is under-used because it can cause life-threatening red blood cell damage in people who have an inherited condition called glucose-6-phosphate dehydrogenase (G6PD) deficiency. In other locations, primaquine is given without testing for G6PD deficiency, putting patients at risk of potentially fatal hemolysis. New rapid diagnostic tests provide the opportunity to screen for G6PD deficiency prior to giving patients primaquine. Our study describes a cost-effectiveness analysis conducted using data gathered from the Thailand-Myanmar border. Our results show that screening for G6PD deficiency followed by primaquine treatment provided a few days of disability-free health per patient treated. This was achieved for similar costs as not giving primaquine to anyone or cost savings when compared to giving primaquine without screening. In addition to the health gains provided to patients, the safe use of primaquine will be a critical tool to eliminate malaria. We provide an interactive cost-effectiveness tool online that can be adapted to other locations to examine the potential costs and benefits of using rapid diagnostic tests for G6PD in different scenarios.
| Plasmodium vivax is an important public health concern, particularly in Asia and South America, where it is now responsible for the majority of malaria cases. While traditionally regarded as a benign disease, P. vivax malaria has been associated with severe and fatal outcomes [1, 2]. As countries move toward malaria elimination and the overall incidence of malaria declines, the proportion of cases that are due to P. vivax infections increases [3, 4]. A single infection of P. vivax can lead to multiple relapses due to its ability to form dormant liver stage parasites called hypnozoites. These relapses are indistinguishable from new infections and repeated episodes can lead to a cumulative risk of anemia and malnutrition [5, 6]. In short latency relapse settings, the majority of P. vivax cases are thought to be due to relapses [7].
Primaquine is the only drug currently licensed for the radical cure of P. vivax; however, it can cause severe hemolysis in individuals with glucose-6-phosphate dehydrogenase (G6PD) deficiency, a common genetic disorder [8] that is positively associated with P. vivax incidence [9]. The prevalence of G6PD deficiency varies from less than 1% to more than 30%, with a mean of 8% in countries where malaria is endemic; equivalent to 350 million people worldwide [10]. G6PD deficiency is largely asymptomatic until individuals are exposed to oxidative stress from an external source, including certain drugs, such as primaquine, and foods, most notably fava beans [9]. The degree of enzyme deficiency varies widely depending upon the genotypic variant which varies with geographical region. A recent review found only 14 documented deaths attributable to primaquine use [11]; however, fatalities may have gone unreported [12].
The WHO recommends that primaquine be used for the radical cure of P. vivax infected patients who can be tested for G6PD deficiency [13]. The gold standard for diagnosing G6PD deficiency is the spectrophotometric assay, a test that requires a laboratory setting and specialized staff [14, 15]. The Fluorescent Spot Test (FST), which is the most widely used assay for G6PD deficiency, is easier to perform but requires basic laboratory equipment, electricity and a cold chain, rendering it difficult to use in remote settings. Thus, routine testing for G6PD deficiency prior to prescribing primaquine generally is not part of antimalarial policy in most countries [16]. Recently, the CareStart G6PD (Access Bio, Somerset, NJ, USA) lateral flow rapid diagnostic test (RDT) has become available for point of care testing. This phenotypic test has high sensitivity for an enzyme activity cut off of 30% [17, 18]; hence false negative results would rarely lead to a G6PD deficient individual receiving primaquine with an attendant risk of hemolysis. Unlike the other G6PD RDT by BinaxNOW (Alere, Orlando, FL, USA), the CareStart RDT can be used in settings where the temperature is above 25°C, a common necessity in P. vivax endemic settings[19, 20].
The availability of point of care G6PD tests is of clinical and public health importance so that P. vivax patients have safe access to primaquine treatment for the prevention of relapses and the resulting health complications [20]. Here we evaluate the cost-effectiveness of using G6PD RDTs on the Thailand-Myanmar border and present our model as an interactive web tool that can be adapted to other settings.
A cost-effectiveness analysis [21] using a health care provider perspective was conducted with decision tree models for P. vivax infections using R statistical software [22] over a 1 year time horizon. The model was parameterized for the north-western border of Thailand with Myanmar (Tak Province), with the benefit of data on recurrences from a recent clinical trial at the Shoklo Malaria Research Unit (SMRU), which provides free of charge care to migrants and refugees [23] (S1 Appendix). In this population of migrants and refugees, the prevalence of G6PD deficiency was documented to be 9–18% [24]. The most common genetic variant was the Mahidol variant (88%) with Chinese-4, Viangchan, Açores, Seattle, and Mediterranean variants also present [24]. Low, unstable P. vivax transmission is seen in this area [3] with a frequent relapse pattern [25]. In recent years, the overall number of malaria cases has been decreasing while the prevalence of P. vivax in the population has remained relatively stable at 9% [3]. Routine practice along the border is to administer 14 days of supervised therapy with or without G6PD screening to patients able to attend the clinic; in practice this is a small proportion of the patients.
The testing of G6PD status with CareStart G6PD RDT before administering primaquine (“screening strategy”) was compared to a strategy in which no G6PD test is performed and primaquine is not used at all (“chloroquine strategy”). In addition, the screening strategy is compared with a strategy where primaquine is given to all patients without testing for G6PD deficiency (“primaquine strategy”) (Table 1). The chloroquine strategy and primaquine strategy were not directly compared to each other because (1) it is unlikely that in settings where the chloroquine strategy is used switching to the primaquine strategy would be a viable option due to the evident concerns about safety and (2) it is unlikely that settings where the primaquine strategy is used would consider changing to the chloroquine strategy due to its inability to achieve radical cure.
Recurrences were recorded over a one year time period in patients who were treated with chloroquine alone, as compared with those who were treated with chloroquine plus 14 days of supervised primaquine for each P. vivax episode (0.5 milligrams (mg)/kilogram (kg)/day) [23]. The relative risk of having at least one recurrence following primaquine treatment was 0.22 as compared to those receiving chloroquine alone. For those who had at least one recurrence, the mean number of recurrences was 3.54 in the chloroquine arm and 1.16 in the primaquine arm. The model applied the inclusion criteria of the clinical trial, which was restricted to patients who were six months and older, not pregnant and presenting with uncomplicated P. vivax malaria [23]. The mean age in the clinical trial was 21 years; this was used for the disability-adjusted life-year (DALY) calculations for years of life lost (S1 Appendix).
The analysis and results were completed separately for males (Fig 1) and females (Fig 2) to account for their differences in risks and outcomes. Firstly, as pregnant females would not be prescribed primaquine due to the unknown G6PD status of the fetus, the screening and primaquine strategies modeled the inclusion of a pregnancy test for all women of childbearing age who were unaware that they were pregnant. Those who were identified or known to be pregnant would be treated with chloroquine only and would not have a G6PD RDT. Secondly, since G6PD is an X-linked disorder, males who have deficiency are hemizygous while females can be either homozygous or heterozygous with a range of G6PD expression levels. Accordingly, G6PD deficiency was divided into two groups: severe (<30% enzyme activity) and intermediate (30–69% enzyme activity) (S2 Appendix). Generally, only females can have intermediate deficiency and the outcomes in this group were taken from heterozygotes. The G6PD RDT with a cut off of 30% activity does not detect heterozygous females with intermediate activity [16, 18]; accordingly, some women with intermediate deficiency who are identified as G6PD normal with currently available RDTs could be at risk of severe hemolysis when prescribed hemolytic drugs.
Table 2 shows the parameters used in the model. Both the screening and primaquine strategies include the cost of supervised therapy in order to reflect the additional costs required for the gains in effectiveness seen in the clinical trial. For the screening strategy, weekly supervised primaquine therapy for 8 weeks was given to those who tested G6PD abnormal [26] and the effectiveness was taken from the trial results for 14 day therapy (S1 Appendix). For each recurrence, the cost and DALY value used were taken from clinical episodes, severe malaria episodes and episodes that resulted in death and weighted proportionally. The probability for severe P. vivax was taken from a meta analysis of clinical studies using those with severe anemia, but other symptoms due to severe P. vivax were not included [2].
The probability of having a hemolytic episode that requires a transfusion in individuals with severe and intermediate G6PD deficiency treated with primaquine was taken from a study of children treated with Dapsone in Africa [27]. While this population may be different in terms of age and G6PD variant from those being treated with primaquine on the northwestern border of Thailand with Myanmar, this was the best available data on transfusion risk. The probability of hemolysis requiring transfusion for severe G6PD deficiency was taken from the proportion of hemizygotes and homozygotes in the study while the probability for females with intermediate deficiency was taken from the proportion of heterozygotes. It was assumed that 10% of patients requiring a transfusion did not receive one; of those, 10% died as a result of not receiving a transfusion. It was assumed that the decision to give a transfusion was made by a physician and that the costs are included in the cost of transfusion.
Costs of commodities and service delivery were taken from Myanmar and Thailand and supplemented by international sources when needed. Costs are reported in 2014 United States Dollars (US$). The cost of supervised therapy was taken from data on annual costs of a community health worker in Myanmar [32], assuming one half-day of pay per observation. The cost of hospitalization for a blood transfusion was included for severe hemolytic episodes which did not lead to death. Table 1 describes the costs for recurrences.
The DALY weights were taken from the 2010 Global Burden of Disease Study [38]. These weights were combined with life tables for Myanmar [36] and assumptions about the length of illness to calculate the DALY burden for each strategy.
In instances where the screening strategy averted DALYs while costing more money the incremental cost-effectiveness ratios (ICER) was calculated:
ICER = Costs– CostbDALYb−DALYs
Where Cost is the total cost of the strategy and DALYs is the total DALYs of the corresponding strategy. While the gross domestic product per capita for Myanmar is approximately US$1200 [39], it has been argued that a lower willingness to pay threshold may be appropriate lower income countries [40]; consequently, a threshold of US$500 was chosen to reflect the resource limitations of healthcare facilities serving migrant and refugee communities.
A one-way sensitivity analysis was conducted to examine the impact of parameter values on the overall outcome. Low and high values were taken from 95% confidence intervals (CIs) when available. When not available, the point estimate was varied by 50% and given wider intervals when necessary to reflect the uncertainty (Table 2). Results that varied from the base case by more than US$0.05 or 0.0002 DALYs averted were reported.
A probabilistic sensitivity analysis (PSA) was conducted to incorporate the uncertainty of all parameters over 1000 sampling iterations using the parameter ranges used in the one-way sensitivity analysis. Table 2 lists the distributions used in the PSA. The sum of squared differences was minimized from the specified ranges to produce the shape values for the beta and gamma distributions and random numbers were generated from these distributions. The mean number of recurrences for each iteration was calculated from 100 bootstrapped data points that were randomly sampled from the data set with replacement (S1 Appendix). The PSA produced a mean estimate and 95% credible intervals (CrIs) for the costs, DALYs and incremental results.
A key concern is adherence to primaquine regimens by the patients as well as compliance to guidelines by prescribers, which is collectively referred to as “adherence” here. In order to account for this, a two-way sensitivity analysis examined the interplay of costs and benefits depending on adherence to the primaquine strategy (whether primaquine was administered to the patient and the full course taken) and screening strategy (whether a G6PD RDT plus primaquine was administered to the patient and the full course taken). This cohort analysis assumed that at 0% adherence all individuals have a relative risk and mean number of recurrences equivalent to receiving chloroquine only. The proportion of individuals in the population who are adherent increases steadily until 100% adherence, which assumes that recurrences are equivalent to the base case. Costs of supervised primaquine and G6PD screening were also varied accordingly. Assumptions about adherence in individuals with G6PD deficiency who receive 14 day primaquine remain the same as the base case analysis.
Costs and DALYs for each strategy are shown in Table 3 and the cohort results are in Table 4. On the Thailand-Myanmar border, the screening strategy averted more DALYs than the chloroquine strategy: 0.026 for males and 0.024 for females. These gains were produced for similar costs. The base case ICERs were US$6.3 and US$11.7 per DALY averted for males and females respectively.
Fig 3 shows the results of the one-way sensitivity analysis in males (see S3 Appendix for all results). The screening strategy always averted more DALYs than the chloroquine strategy (S3C and S3D Appendix). Costs for the screening strategy were highest when radical cure had a low impact on recurrences, when the costs of supervised therapy and the G6PD RDT were increased, and also when the cost of a recurrence was decreased (S3A and S3B Appendix). The only assumptions that made the screening strategy cost over US$500 per DALY averted were lowering the number of recurrences after chloroquine to 1 (US$3678.7 in males and US$3724.5 in females) and assuming the same relative risk of having at least one recurrence in females (US$1223.6).
The mean costs and DALYs and CrIs estimated by the PSA are shown in Table 4. The screening strategy costing more than the chloroquine strategy with mean incremental cost of US$0.8 (95%CrI: –17.4 to 19.7) and 0.026 DALYs averted (95%CrI: 0.007 to 0.117) per male (Fig 4A). At a willingness to pay threshold of US$500, the screening strategy had an 81.2% probability of being cost-effective (Fig 5A). The PSA resulted in a mean incremental cost of US$0.75 (95%CrI: –15.0 to 20.0) with 0.023 DALYs averted (95%CrI: 0.006 to 0.122) per female (Fig 4C) and a 77.6% probability of being cost-effective at a willingness to pay threshold of US$500 (Fig 5C). The ICERs were US$31.3 per DALY averted for males and US$32.4 for females.
Again, the screening strategy resulted in better health outcomes in the base case with 0.011 DALYs averted in males and 0.004 in females (Table 3). The health gains in females were more modest due to their overall lower probability of hemolysis requiring transfusion. In addition, the screening strategy produced cost savings of US$7.1 and US$2.2 per male and female initially treated, respectively (Table 3). The simulation output indicated that one death due to hemolysis would be expected for every 6682 males and 15,994 non-pregnant females treated using the primaquine strategy. This would be reduced to one death per 668,164 males and 201,198 non-pregnant females treated with the screening strategy (Table 4).
The one-way sensitivity analysis showed that changes in the parameter values had a smaller impact on the costs when comparing screening and primaquine strategies, especially for females, and results that consistently averted DALYs (S3E–S3H Appendix). The parameters related to mortality, the need for transfusion and the prevalence of G6PD deficiency having the highest impact on DALY results (S3G and S3H Appendix). The screening strategy was cost saving with the exception of raising the G6PD RDT cost to US$10.0, which caused an incremental cost for the screening strategy of US$235.9 and US$1602.1 for males and females, respectively (S3E and S3F Appendix). The screening strategy remained cost saving even at low levels of G6PD deficiency (7%).
The PSA showed a mean cost savings of US$7.3 (95%CrI: -15.4 to 3.4) and 0.012 DALYs averted (95%CrI: 0.001 to 0.113) in males (Fig 4B), and a mean cost savings of US$2.2 (95%CrI: –6.2 to 6.7) and 0.004 DALYs averted (95%CrI: 0.000 to 0.029) in females (Fig 4D). The screening strategy had a 97.7% probability of being cost-effective for males at a willingness to pay threshold of US$500.0 (Fig 5B). For females, the probability was 91.1% (Fig 5D). Table 3 shows the cost and DALY estimates from the PSA.
The two-way analysis (Fig 6) demonstrated that the screening strategy would be cost-effective in scenarios where it is used to maintain or increase the number of patients who are adherent to their primaquine regimens. The impact of switching to the screening strategy was slightly less in females due to the exclusion of pregnant women from primaquine treatment and the low sensitivity of the G6PD RDT in women with intermediate G6PD deficiency.
Due to the extensive heterogeneity and parameter uncertainty around key parameter estimates, notably relapse patterns [25], G6PD variants and prevalence [10] and costs, a web-based interface was built using the R-Shiny application so that the model could be adapted to other settings as need be. See website (https://malaria.shinyapps.io/g6pd_screening/).
Point of care G6PD RDTs offer the opportunity for the safe uptake of primaquine for the prevention of recurrences. Our findings suggest that on the Thailand-Myanmar border the use of G6PD RDTs to identify patients with G6PD deficiency before supervised primaquine is likely to provide significant health benefits (equivalent to between 1 and 9 days of perfect health) compared to giving chloroquine alone or giving 14 day primaquine without G6PD testing. Furthermore, the use of point of care G6PD RDTs will potentially save costs or, at most, increase them moderately.
Primaquine is currently the only licensed hypnozonticidal drug, but healthcare professionals who treat P. vivax cases are often more concerned with avoiding the immediate risk of hemolysis than with protecting the patient from the risks associated with future relapses. In other settings, primaquine may be administered without G6PD testing, putting individuals with G6PD deficiency at risk of severe hemolysis, although the degree of risk will depend upon local G6PD variants and their prevalence. In the scenario presented, the use of G6PD RDTs will save costs while averting DALYs compared to a policy in which primaquine is administered without G6PD testing. While our results give a high probability of cost savings when switching from the primaquine strategy to the screening strategy, this should not deter radical cure without screening in settings where screening is unavailable as the primaquine strategy averted more DALYs than the chloroquine strategy.
Our model is based on supervised primaquine therapy and hence our findings may not be applicable to other settings where unsupervised primaquine is the norm and adherence to a complete course of treatment and thus effectiveness may be low [41]. Our two-way analysis on adherence (Fig 6) enabled comparison between settings with varying adherence and how this impacts upon cost effectiveness. The screening strategy averts more DALYs than the primaquine strategy, even at relatively high primaquine strategy adherence and low screening strategy adherence. Shorter drug courses, such as 7 day primaquine and tafenoquine, should contribute to higher adherence levels and reduced costs for the primaquine and screening strategies.
Overall, the screening strategy was less cost-effective in women as compared to men. This reflects a greater proportion of women who are excluded from receiving primaquine due to pregnancy and the lower risk of severe hemolysis in females with intermediate G6PD deficiency. Since men represented 65% of patients in the trial that the recurrence data were derived from (S1 Appendix), the overall cost-effectiveness estimates per person presenting with P. vivax malaria would likely be closer to the results for males.
The cost-effectiveness of using G6PD RDTs is also dependent on the diagnostic accuracy of the test. Our model draws on studies conducted on the Thailand-Myanmar border, which demonstrated a high sensitivity in healthy volunteers. While other studies have shown similar results [16, 17] in healthy volunteers, a recent study in Brazil found that the sensitivity of the CareStart G6PD RDT dropped to 50% in patients with malaria compared to 80% in those who did not [15]. A recent cost-effectiveness analysis of male patients with P. vivax malaria in Brazil used a low sensitivity for the CareStart G6PD RDT (46%) but still found it to be more cost-effective than both the BinaxNOW test and routine care; where the analysis also involved the prescription of primaquine without having a G6PD test [42]. This study, however, used the endpoints ‘adequately diagnosed case’ and ‘hospitalization avoided’ instead of DALYs. The Brazilian population was given a 94% probability of hospitalization when primaquine was given to G6PD deficient men. Our model differs in that we assume a lower rate of hospitalization due to severe hemolysis. We also include results for both genders and report DALYs, enabling comparisons with interventions for other diseases.
Our study has a number of limitations, mostly related to our model assumptions. The cost-effectiveness of the screening strategy would be increased if it included the onward transmission of P. vivax or the longer term impact of repeated episodes, such as anemia, malnutrition and all-cause mortality. This is particularly relevant in areas such as the Thailand-Myanmar border where the estimated proportion of recurrences due to relapses is estimated to be 78%. The cost-effectiveness may decrease if some individuals were not able to metabolize primaquine, if healthcare workers were not able to utilize G6PD RDTs or supervise primaquine regimens, if the prevalence of G6PD deficiency in those presenting with P. vivax was lower to that in the general population, if the diagnostic accuracy of the G6PD test were lower, if healthcare facilities providing care for hemolytic episodes were not accessible or if the operational costs of implementing a switch to the screening strategy were included. These parameter limitations are similar to those highlighted in a recent review of the costs and cost-effectiveness of P. vivax control and elimination [43].
Finally, our model is limited by the paucity of data available in the literature for some parameter values, including the mortality rate for those who have a primaquine-induced hemolytic episode requiring transfusion but do not receive them. Our assumptions of primaquine induced mortality were derived from previous risks of mortality in patients treated with Dapsone in Africa and equated to a population risk of 1 in 6,682 administrations to males and 1 in 15,994 administrations in females. These risks are significantly higher than the risks documented in a previous review [11] but the screening strategy averted more DALYs than the primaquine strategy at a lower level of primaquine-induced mortality, though this is likely due to the utilization of weekly primaquine by the screening strategy. Other variables, such as the prevalence of G6PD deficiency will vary greatly depending on the epidemiological setting. Whilst it would be beneficial to gather more robust parameter estimates on which to base informed policy decisions, this should be tempered by the feasibility of gathering such data and the potential benefits of implementing appropriate policies sooner, especially in the context of elimination.
Although our model is relatively simple, it provides a useful starting point for policy makers to compare the risks and benefits of using G6PD RDTs to enable the safe and effective use of primaquine. To assist in this process we provide an online tool with which policy makers and healthcare providers can vary the assumptions made in the model in keeping with local scenarios and as additional data becomes available (https://malaria.shinyapps.io/g6pd_screening/). As the only licensed antimalarial for the radical cure of P. vivax infections, primaquine will be a critical tool for the elimination of all malaria [44] and for the health gains provided to patients. The currently available G6PD RDTs can identify G6PD deficient males, making the screening strategy an attractive option regardless of current practice. In situations where blood transfusions are not accessible, further information may be required on the prevalence of G6PD deficiency and associated risk of hemolysis in females with intermediate G6PD deficiency who test normal by current G6PD RDT methods [28]. Despite the initial cost, point of care RDTs avert DALYs by reducing recurrences while diminishing the hemolytic risk in G6PD deficient patients.
|
10.1371/journal.pntd.0006090 | Quantitative proteomic analysis of amastigotes from Leishmania (L.) amazonensis LV79 and PH8 strains reveals molecular traits associated with the virulence phenotype | Leishmaniasis is an antropozoonosis caused by Leishmania parasites that affects around 12 million people in 98 different countries. The disease has different clinical forms, which depend mainly on the parasite genetics and on the immunologic status of the host. The promastigote form of the parasite is transmitted by an infected female phlebotomine sand fly, is internalized by phagocytic cells, mainly macrophages, and converts into amastigotes which replicate inside these cells. Macrophages are important cells of the immune system, capable of efficiently killing intracellular pathogens. However, Leishmania can evade these mechanisms due to expression of virulence factors. Different strains of the same Leishmania species may have different infectivity and metastatic phenotypes in vivo, and we have previously shown that analysis of amastigote proteome can give important information on parasite infectivity. Differential abundance of virulence factors probably accounts for the higher virulence of PH8 strain parasites shown in this work. In order to test this hypothesis, we have quantitatively compared the proteomes of PH8 and LV79 lesion-derived amastigotes using a label-free proteomic approach.
In the present work, we have compared lesion development by L. (L.) amazonensis PH8 and LV79 strains in mice, showing that they have different virulence in vivo. Viability and numbers of lesion-derived amastigotes were accordingly significantly different. Proteome profiles can discriminate parasites from the two strains and several proteins were differentially expressed.
This work shows that PH8 strain is more virulent in mice, and that lesion-derived parasites from this strain are more viable and more infective in vitro. Amastigote proteome comparison identified GP63 as highly expressed in PH8 strain, and Superoxide Dismutase, Tryparedoxin Peroxidase and Heat Shock Protein 70 as more abundant in LV79 strain. The expression profile of all proteins and of the differential ones precisely classified PH8 and LV79 samples, indicating that the two strains have proteins with different abundances and that proteome profiles correlate with their phenotypes.
| Leishmaniasis is an antropozoonosis caused by Leishmania parasites that affects around 12 million people in 98 different countries. Cutaneous leishmaniasis caused by Leishmania amazonensis can have different clinical forms and severities depending on the parasite strain. We have here shown that two Leishmania amazonensis strains, named PH8 and LV79, which have different virulence in mice, also have different protein signatures. In fact, samples from these strains can be distinguished based on the abundance of all proteins detected and of the differential ones. Differential proteins identified in this work may be employed in the future to predict virulence of parasite strains or isolates.
| Leishmaniasis is an antropozoonosis that affects around 12 million people in 98 different countries in Europe, Africa, Asia and America [1]. More than 1,5 million new cases are reported every year, 0,7 to 1,2 of them of the tegumentary forms and 0,2 to 0,4 million of the visceral form [1]. The clinical form of the disease depends mainly on the Leishmania species and on the immunologic status of the host [2]. In Brazil, Leishmania (Viannia) braziliensis and Leishmania (Leishmania) amazonensis are the species most frequently involved in tegumentary leishmaniasis [3]. The human L. (L.) amazonensis symptomatic infection frequently leads to the localized cutaneous leishmaniasis (LCL), with moderate cellular hypersensitivity, and more rarely to the diffuse cutaneous leishmaniasis (DCL), associated with anergy to parasite’s antigens [3].
The parasite has two main forms: promastigotes, transmitted by an infected female phlebotomine sand fly, and amastigotes, which live and replicate in phagolysosomes of phagocytic cells, mainly macrophages [4,5]. Macrophages are important cells of the immune system, capable of directly killing intracellular pathogens and triggering adaptive responses against them [6]. When activated, these cells produce cytokines and reactive oxygen species, nitric oxide, lysosomal enzymes and proteases with microbicidal effects [5]. Leishmania, however, can evade these mechanisms and replicate inside macrophages due to parasite´s virulence factors [7,8].
The importance of specific virulence factors may vary according to the Leishmania species. Protein A2, LACK (homolog of receptor for activated C kinase) and cathepsin L-like cysteine protease B (CPB), for instance, are considered important factors for L. (L.) donovani, L. (L.) major and L. (L.) amazonensis, respectively [2]. Inositol phosphosphingolipid phospholipase C-like (ISCL) is also considered an essential factor for L. (L.) major survival inside the acid phagolysosome [9]. Curiously, while L. (L.) major ISCL knock out parasites lost virulence in BALB/c mice, L. (L.) amazonensis ko parasites had similar virulence compared to wild type in this mouse strain [10].
Lipophosphoglycan (LPG) and major surface glycoprotein GP63 are by far the most studied Leishmania virulence factors. LPG is the most abundant molecule in promastigote´s surface [11]. It inhibits macrophage nitric oxide production, signal transduction and apoptosis, delays phagolysosome maturation and induces RNA double strand-dependent protein kinase (PKR), which increases parasite growth [12–14]. Although essential for L. (L.) major and L. (L.) donovani infectivity, LPG is not necessary for L. (L.) mexicana infection in vitro and in vivo [15,16]. The zinc-metalloprotease GP63 is an important antigen in promastigotes, also expressed (at lower levels) in amastigotes [17]. GP63 facilitates Leishmania infection and survival since it degrades extracellular matrix, decreases kinase and upregulates phosphatase activity in infected macrophages, and enhances the resistance to antimicrobial peptides. Besides, GP63 cleaves C3 to C3b and C3bi, increasing parasite resistance to complement-mediated lysis, and directly cleaves the pro-inflammatory factors AP-1 and NF-κB (reviewed in [11,17]). Interestingly, it was recently shown that cysteine peptidase B, an important virulence factor for L. (L.) mexicana and L. (L.) amazonensis [18], regulates the levels of LPG and GP63 in L. (L.) mexicana [19].
While some factors are restricted to the parasite surface, others can be secreted. GP63, elongation factor 1 alpha (EF-1α), frutose-1,6-bisphosphate aldolase, secreted acid phosphatase (SAcP), heat shock proteins (HSPs) 10 and 70 and tryparedoxin peroxidase, among others, are produced and secreted by amastigotes [8,20]. Not only GP63, as previously mentioned, but also EF-1α, aldolase and SAcP, interact with macrophage kinases and phosphatases, reducing cell activation and microbicidal capacity [8]. Cysteine peptidases may either accumulate inside amastigotes or be secreted in exosomes, depending on the Leishmania species. These important virulence factors have roles both inside the parasite and in the host [11].
It is well known that Leishmania species differ in terms of virulence, as illustrated by the fact that several mouse lineages are resistant to L. (L.) major and susceptible to L. (L.) amazonensis [2,21]. It is also known that strains of the same Leishmania species may show different infectivity and metastatic phenotypes in vivo [22–24]. Although proteome comparison has been extensively employed for the identification of proteins involved in resistance to drugs [25–29], few studies have used this strategy to identify virulence factors. One of them compared different clones of L. (V.) guyanensis and identified two proteins associated with metastatic capacity [22]. Another study analyzed two strains of L. (L.) infantum with different infectivity in vivo and found that proteins such as KMP-11, heat shock proteins, tryparedoxin peroxidase (CPx) and peroxidoxin were differentially expressed [23]. A recent work compared L. (V.) braziliensis isolates from mucosal and cutaneous lesions of the same patient and observed overexpression of prostaglandin f2-alpha synthase and HSP70 in cutaneous isolates [24].
We have previously shown that LV79 strain of L. (L.) amazonensis develop small lesions in C57BL/6 mice. In fact, LV79 lesions in this mouse strain increase until six weeks after inoculation and decrease thereafter, although parasites can still be found in lesions until thirteen weeks post infection [30]. On the other hand, PH8 strain was shown to generate lesions of increasing size in the same mouse strain [31]. In the present work, we show that promastigotes from LV79 and PH8 strains induce different lesion development in BALB/c and C57BL/6 mouse strains, and that amastigotes from PH8 are more infective. Differential abundance of virulence factors probably accounts for the higher virulence of PH8 amastigotes. In order to test this hypothesis, we have quantitatively compared the proteomes of PH8 and LV79 lesion-derived amastigotes using a label-free proteomic approach.
The comparison of the proteomes of lesion-derived amastigotes from the two strains identified proteins such as CPx, SOD and HSP70 as significantly more abundant in LV79 amastigotes, and GP63 as more abundant in PH8 parasites. The expression profile of all proteins and of the differentially expressed ones precisely classified PH8 and LV79 samples, indicating that protein abundance profiles correlate with the phenotypes of the two strains.
All animals were used according to the Brazilian College of Animal Experimentation (CONEP) guidelines, and the protocols were approved by the Institutional Animal Care and Use Committee (CEUA) of the University of São Paulo (protocol number 001/2009). Euthanasia was performed in CO2 camera.
Promastigotes of Leishmania (L.) amazonensis LV79 (MPRO/BR/72/M 1841) and PH8 (IFLA/BR/67/PH8) strains were cultured at 24°C in M199 medium supplemented with 10% fetal calf serum (FCS). Parasites were sub-cultured every 7 days to inoculums of 2 × 106/mL.
For differentiation of amastigotes into promastigotes, lesion-derived parasites were counted using Neubauer chamber and transferred to M199 medium with 10% FCS at densities of 103, 104 and 105 parasites/mL. Cells were incubated at 24°C for 4 days and promastigote densities were determined.
Four to 8-week-old BALB/c and C57BL/6 mice maintained in our facilities were infected in the left hind footpads with 2 × 106 promastigotes of L. (L.) amazonensis LV79 or PH8 in the beginning of stationary-phase (day 5, see S1 Fig) in a final volume of 20μL. Footpad thickness was measured weekly using a caliper (Mitutoyo Corporation, Japan). For histological analysis, we employed five BALB/c animals for each parasite strain. Animals were euthanized, infected paws were removed and control footpads were removed from uninfected mice with similar ages. Fragments of these tissues were fixed in 10% buffered formalin for 18 h, washed and dehydrated in graded concentrations of ethanol, diaphanized and embedded in paraffin. The 4 μm paraffin sections were stained with hematoxylin and eosin.
For immunohistochemistry, sections were deparaffinized, blocked with 5% BSA in PBS for 30 minutes, and incubated in 0,1% sodium azide, 3% H2O2 in methanol for 30 minutes for blocking endogenous peroxidase. After incubation with rabbit anti-Leishmania serum (gently provided by Prof. Mauro Cortez) 1:1000 in PBS 2% (w/v) BSA for 18h at 4°C, slides were washed in PBS and incubated with secondary anti-rabbit peroxidase-conjugated antibody (Imuny, Brazil) 1:2000 in PBS 2% (w/v) BSA for two hours. Slides were then washed in PBS, incubated with DAB (DAKO, Denmark) for 2 minutes, washed in water and counterstained with hematoxylin. Samples were dehydrated and diaphanized, mounted with Permout (Sigma) and analyzed in Nikon Eclipse E200 LED microscope with Moticam 580 (Motic) camera.
Amastigotes were purified as previously described [32]. Briefly, lesions were minced and homogenized in 5mL PBS using a tissue grinder (Thomas Scientific). After centrifugation at 50 x g for 10 min at 4°C, the supernatant was recovered and centrifuged at 1450 x g for 17 min at 4°C. Supernatant was then discarded and the pellet was washed three times with PBS followed by centrifugations at 1450 x g for 17 min at 4°C. After 3h of incubation in RPMI with 4% serum under rotation at room temperature to liberate endocytic membranes, amastigotes were further centrifuged, resuspended in 2mL of erythrocyte lysis buffer (155mM NH4Cl, 10mM KHCO3, 1mM EDTA, pH7,4) and incubated for 2 min in ice. Parasites were washed twice in PBS, resupended at 109 cells/300μL in PBS + Proteoblock (a protease inhibitor cocktail from Fermentas) and lysed by 8 cycles of freeze thaw in liquid nitrogen-42°C. Soluble proteins were obtained after centrifugation at 12.000 x g for 3 min and quantified by Bradford (Biorad).
MTT assay was performed using MTT (3-[4,5-dimethylthiazol-2-yl]-2,5- diphenyltetrazolium bromide (Sigma) as previously described [13]. Briefly, 2x107 lesion-derived amastigotes were resuspended in 100 μL PBS with 5mM glucose, transferred to 96 well plates and incubated with 20 μL of MTT (5mg/ml in PBS) at 34°C for 50 minutes. 100 μl of SDS 10% were added and absorbance at 595 nm (reference at 655nm) was measured in BioTek ELx800 equipment (Biotek, Winooski,VT, USA).
Trypsin-like activity in amastigote extracts was assayed as we recently described [33].
100 μg of soluble amastigote proteins from each sample were digested with trypsin. The resulting peptide mixture was analyzed on a LTQ Velos Orbitrap mass spectrometer (Thermo Fisher Scientific) coupled with LC-MS/MS by an EASY-nLC system (Thermo Fisher Scientific) through a nanoelectrospray ion source. Sample concentration and desalting were performed online using a pre-column (2 cm; 100 μm ID; 5 μm C18-A1; Thermo). Separation was accomplished on Acclaim PepMap 100 C18 column (10cm; 75um ID; 3um C18-A2; Thermo) using a linear gradient of A and B buffers (buffer A: A = 0.1% formic acid; Buffer B = 99% ACN, 0.1% formic acid) from 1% to 50% buffer B over 60 for a total of 77 min at a flow rate of 0.3 μL/min to elute peptides into the mass spectrometer. Columns were washed and re-equilibrated between LC—MS/MS experiments. Mass spectra were acquired in the positive-ion mode over the range m/z 400–1500 at a resolution of 30,000 (full width at half-maximum at m/z400) and AGC target >1 × e6. The 20 most intense peptide ions with charge states ≥2 were sequentially isolated to a target value of 5,000 and isolation width of 2 and fragmented in the linear ion trap using low-energy CID (normalized collision energy of 35%) with activation time of 10 ms. Dynamic exclusion was enabled with an exclusion size list of 500, exclusion duration of 30 s, and a repeat count of 1. Three biological replicates (amastigotes from three independent mice infections) were performed with two technical runs for LV79 and PH8.
For protein identification and quantification, raw files were imported into MaxQuant version 1.5.2.8 [34]. The database search engine Andromeda [34,35] was used to search MS/MS spectra against a database composed of Uniprot Mus musculus (release May 5th, 2016; 50,189 entries) and Leishmania sp (release May 5th 2016, 50,820 entries) databases. Database search employed the following parameters: (i) mass tolerance of 4.5 ppm and 0.5 Da for MS and MS/MS, respectively; (ii) trypsin cleavage at both ends and two missed cleavage allowed; (iii) carbamidomethylation of cysteine (57.021 Da) was set as a fixed modification, and oxidation of methionine (15.994 Da) and protein N-terminal acetylation (42.010 Da) were selected as variable modifications. All identifications were filtered to achieve a protein and peptide FDR of 1%. One peptide was set as the minimum number for protein identification, and all proteins identified with one peptide had this peptide as unique peptide that could unambiguously identify that protein. For protein quantification, a minimum of two ratio counts were required. All identifications were filtered to achieve a protein and peptide FDR of less than 1% as recommended in the proteomic community for large scale mass spectrometry-based experiments acquired in the data-dependent mode used in this study.
Protein quantification was based on the MaxQuant label-free algorithm using both unique and razor peptides for protein quantification, and at least 2 ratio counts were required for considering a protein quantification valid. Protein abundance was calculated based on label-free protein quantification (LFQ) values, which are normalized intensities calculated by the MaxQuant software [36]. LFQ-based quantification was shown to provide very accurate and robust quantification and has been validated in many diverse biological contexts [37]. Fold changes were calculated by dividing the average of the LFQ intensities from LV by the average of LFQ intensities from PH replicates.
Statistical analyses of the proteome data were performed using Perseus v.1.5.4.1 in the MaxQuant environment. First, proteins identified in the reverse database, potential contaminants and proteins identified only by site were excluded. The LFQ intensities were log2 transformed and the averages of the two technical replicates values for each independent experiment were calculated. T-test analysis was applied on the PV and PH groups with a p value set to p<0.05. Hierarchical clustering of significantly altered proteins was performed using the Z-score calculation on the log2 intensity values, and the results were represented as a heat map. Principal component analysis was constructed in the web-based chemometrics platform MetaboAnalyst 2.0 [38].
Western blots were performed as previously described [32] using 25μg of soluble amastigote proteins and 12% acrylamide gels. After incubation with ECL Prime Western Blotting Detection Reagent (GE healthcare) for five minutes, membranes were developed using ChemiDoc XRS+ (BioRad) and analyzed using Image Lab (BioRad) software. The results were normalized to actin band intensities.
Both LV79 and PH8 L. (L.) amazonensis strains cause lesions in BALB/c and C57BL/6 mice, but lesions were smaller and decreased with time in C57BL/6 mice (Fig 1A). On the other hand, BALB/c lesions were significantly larger than C57BL/6 for both parasite strains, as we have already described [30]. PH8 lesions were significantly larger than LV79 in both mouse strains (Fig 1B, 1C and 1D), and parasite loads tend to be higher in infections with this L. (L.) amazonensis strain (Fig 1E).
We also compared histological sections of PH8 and LV79 lesions in BALB/c mice. After twelve weeks of infection, BALB/c mice showed disrupted footpad structure and high abundance of infected macrophages for both parasite strains, and more abundant necrosis in LV79 lesions (Fig 2A and 2B). Immunohistochemistry indicated higher abundance of parasites (labeled in brown) in PH8 lesions (Fig 2C versus Fig 2D), corroborating the higher parasite recovery (Fig 1E) and lesion size (Fig 1D) observed for this L. (L.) amazonensis strain.
Infections shown in the previous experiments were initiated with promastigote cultures in stationary phase. To verify whether infections using amastigotes of PH8 and LV79 also generated lesions with significant different sizes, we isolated lesion-derived parasites and inoculated them in naïve BALB/c footpads. Before inoculation, we estimated parasite viability by MTT assay and analyzed trypsin-like activity, used as a measure of metacaspase activity, which is directly associated to parasite death [33]. We also compared parasite differentiation into promastigotes. In Fig 3A we show that lesion-derived amastigotes from PH8 strain have higher viability than LV79, and, accordingly, lower trypsin-like activity (Fig 3B). As expected, PH8 amastigotes generate cultures with higher numbers of promastigotes (Fig 3C). Lesions generated after inoculation of PH8 amastigotes were bigger than the ones generated by LV79 amastigotes, as shown in Fig 3D, 3E and 3F. To analyze if the larger sizes of PH8 lesions could be attributed to a higher number of viable parasites, we adjusted LV79 parasite numbers considering their viability, so that we would inoculate the same number of viable amastigotes for LV79 and PH8. As shown in Fig 3D, 3E and 3F, infections with normalized LV79 parasites still led to smaller lesions than PH8, indicating that the higher virulence of PH8 cannot be solely attributed to the increased viability of lesion amastigotes. In fact, only in infections using 5 or 10 times more LV79 amastigotes we observed a lesion development pattern similar to PH8´s (S2 Fig).
Differential abundance of virulence factors probably accounts for the higher virulence of PH8 amastigotes. In order to test this hypothesis, we have quantitatively compared the proteomes of PH8 and LV79 lesion-derived amastigotes using a label-free proteomic approach. Amastigote loads for LV79 strain in C57BL/6 mice lesions 13 weeks after infection are around 7 x 104 parasites/footpad, much lower than the 1.5 x 108 parasites/footpad of BALB/c, as we have recently shown [30]. This low parasite recovery precluded the use of C57BL/6-derived amastigotes for proteome analysis.
Three independent experiments (named 1, 2 and 3) were performed with BALB/c mice infected with stationary promastigotes of the two strains, and each amastigote sample was analyzed in technical duplicates. The total number of proteins identified in the Leishmania database, considering all experiments and replicates, was 301. Fig 4A indicates that 276 of the 301 proteins were detected in the proteomes of both strains, while 15 and 10 proteins were detected only in LV79 and PH8 amastigotes, respectively (S1 Table). Among the proteins identified in both samples, 12 were significantly more abundant in PH8 amastigotes and 25 in LV79 (Table 1). Among these 37 proteins, 16 had fold changes of at least 2 (ratios LV79/PH8 higher than 2 or lower than 0,5): 11 more abundant in LV79 and 5 more abundant in PH8, which are now depicted in bold in Table 1. Although most fold changes were not very high, they are robust since they have statistical significance after t-test of three independent experiments. These results indicate that among the 301 proteins identified, 20% (62 proteins) were either exclusively detected or increased in one of the strains. It is important to mention that among the 301 proteins, 218 (72%) were common across all experiments (PH8 and LV79) and replicates. We also observed that the R2 correlation value of the quantified protein signals between individual replicates was excellent, with a range of 0.929–0.975, indicating high reproducibility among replicates.
The pattern of expression of the 37 differential (but not exclusively detected) proteins precisely clustered PH8 and LV79 samples in two separate branches, as shown in Fig 4B. When we employed expression data of all identified proteins, including the two technical replicates of each sample, PH8 and LV79 samples still clustered (Fig 4D). Samples were also efficiently grouped based on principal component analysis (Fig 4C), indicating that the two strains have remarkable differences in terms of protein abundance.
Proteins with different abundance comparing PH8 and LV79 are involved in several cellular processes, among them metabolism/ ATP synthesis, signaling, proliferation/replication, translation, and oxidative stress (Table 1). These proteins included some known Leishmania virulence factors such as cysteine protease, tryparedoxin and tryparedoxin peroxidase (CPx), superoxide dismutase (SOD), GP63, heat shock protein 70 (HSP70) and elongation factor.
Proteins showing subtle differences are more difficult to validate in “semi-quantitative” Western blot assays, and for this reason we have chosen to validate proteins with ratios higher than 2: tryparedoxin peroxidase, with fold 4,62 in LV79/PH8, and GP63, with fold 0,34 in LV79/ PH8 (2,94 times more abundant in PH8). Both were analyzed using antibodies developed against Leishmania (anti-GP63, anti-CPx). The images and corresponding bar graphs shown in Fig 5 validate proteome analysis (Table 1): GP63 is indeed more abundant in PH8 proteomes, and CPx is more abundant in LV79 proteomes.
We have shown that BALB/c and C57BL/6 mice infected with promastigotes of LV79 and PH8 strains develop lesions with striking different sizes according to the parasite and mice strains. In comparison to BALB/c mice, C57BL/6 lesions were smaller and decreased with time for both parasite strains, different from the huge increasing lesions previously reported for PH8 [31]. This discrepancy may be attributed to C57BL/6 strain maintained in different animal facilities or to the parasite strain from different labs. Anyway, lesions caused by PH8 strains were significantly smaller than the ones induced by LV79 in both BALB/c and C57BL/6 mice.
Amastigotes from the two strains were compared in terms of protein abundance, as shown by proteome analysis. The 301 Leishmania proteins identified in this study represent a small fraction of the 6000 proteins predicted to be expressed in amastigotes, and several reasons may explain this fact. First, the study of lesion-derived amastigotes´ proteome presents some technical challenges due to the interference of host proteins, which are carried along amastigote purification and protein extraction steps even using a well stablished protocol such as ours. The presence of host proteins certainly diminishes our capacity of identifying a higher number of parasite proteins. In fact, after protein identification using a database composed of Uniprot Mus musculus and Leishmania sp, a total of 213 and 301 proteins were identified in the mouse and Leishmania databases, respectively, and 815 of the peptides detected belong to mouse proteins and 875 to Leishmania proteins. Moreover, we have analyzed the iBAQ values, which may be used as a measure of protein abundance [40], and are calculated by dividing the total intensity of a protein by the number of tryptic peptides between 6 and 30 amino acids in length. Comparing the total iBAQ value for Leishmania proteins to the total iBAQ value of mouse proteins, we found that Leishmania proteins accounted for the double of the iBAQ value of mouse ones. Besides, we did not perform sub-cellular fractionation or peptide fractionation prior to the LC-MSMS analyses. Instead, we only considered soluble proteins from a non-detergent-based protein extraction, since our main interest was on soluble amastigote virulence factors that could modulate macrophage infection and parasite survival. This strategy probably leads to a lower number of proteins compared to total extract preparations using detergents [41] or sub-cellular fractionation. At last, biological or chemical post-translational modifications as well single nucleotide polymorphism were not included as variable modifications in the MSMS search, which may represent a fraction of MSMS that was not identified.
Proteins considered as virulence factors in Leishmania such as CPx, SOD, GP63 and HSP70 were identified as differentially expressed between the two parasite strains. SOD, CPx and HSP70 are known to reduce oxidative damage in Leishmania. SODs are important in antioxidant defense in many organisms, metabolizing superoxide (O2-) into oxygen (O2) and hydrogen peroxide (H2O2). They are organized in three families based on the metal ion that supports activity: Ni, Cu complexed with Zn, and Mn or Fe [42]. Eukaryotes including mammals have Cu/ Mn/ ZnSODs, whereas FeSODs have been found in prokaryotes, protozoans, plants, and algae [43]. Different FeSOD species (FeSOD-A and FeSOD-B) have been characterized in L. (L.) chagasi, L. (L.) tropica, and L. (L.) donovani [44–46], and in this work we have identified a Fe SOD in L.(L.) amazonensis proteome similar to L. (L.) mexicana enzyme.
CPx has been shown to increase oxidative resistance in L. (L.) donovani [47], L. (L.) infantum [48] and L. (L.) amazonensis [49]. This enzyme also augments infection [47] and virulence [23] of L. (L.) donovani. High levels of the enzyme were reported in antimony resistant L. (L.) donovani [48], L. (L.) braziliensis and L.(L.) chagasi [29], in L. (L.) amazonensis resistant to arsenite [49] and in metastatic L. (V.) guyanensis [50]. HSP70 also protects Leishmania from toxic environmental conditions reducing heat-induced denaturation and cell death [51]. Indeed, HSP70 has been shown to be increased in L. (L.) infantum and L. (L.) donovani under heat shock or oxidative and nitrosative stresses [23,51], and the overexpression of this protein conferred increased resistance to H2O2 in L. (L.) donovani [51] and in L. (L.) amazonensis [9]. Like CPx, HSP 70 is overexpressed in antimonial resistant L. (L.) infantum and L. (V.) braziliensis parasites [29]. Besides, more virulent isolates of L. (V.) braziliensis showed increased HSP70 expression [24]. SOD, CPx and HSP70 were all more abundant in LV79 amastigotes. Interestingly, parasites from this strain generated smaller lesions and showed lower viability after isolation from lesions. It is possible that other virulence factors compensate for the lower expression of these three proteins and account for PH8 higher virulence and survival in the host, or that post translation modifications of one or some of these proteins generate more active protein species in PH8. In fact, we have previously described different species of CPx and HSP70 in L. (L.) amazonensis amastigotes [32], and HSP70 activity is known to be influenced by phosphorylation at specific residues [52].
Among the virulence factors mentioned above, only GP63 had higher abundance in the most virulent PH8 strain. Considering that this molecule favors binding of promastigotes to macrophages and intramacrophage survival and replication [53], as well as parasite survival in BALB/c mice [54], it is conceivable that a higher abundance of GP63 may contribute to PH8 virulence.
The results presented here show that amastigotes from L. amazonensis strains PH8 and LV79, which have different virulence in mice, also have proteins with different abundances. To our knowledge, this is the first gel free proteome of lesion-derived amastigotes. Despite the difficulties of working with lesion-derived parasites and the detection of a relatively low proportion of the predicted products, the comparison of PH8 and LV79 strains enabled the reproducible identification of several proteins that distinguish the two strains and that may be involved in virulence in L. amazonensis. In fact, samples from the same strain are efficiently grouped using expression data from all proteins and from the differentially expressed ones. These results indicate that PH8 and LV79 can be distinguished by comparison of protein abundances and that proteome analysis may be used to characterize Leishmania phenotype and eventually predict the virulence of other L. (L.) amazonensis strains or isolates.
|
10.1371/journal.ppat.1000870 | Impaired Innate Immunity in Tlr4−/− Mice but Preserved CD8+ T Cell Responses against Trypanosoma cruzi in Tlr4-, Tlr2-, Tlr9- or Myd88-Deficient Mice | The murine model of T. cruzi infection has provided compelling evidence that development of host resistance against intracellular protozoans critically depends on the activation of members of the Toll-like receptor (TLR) family via the MyD88 adaptor molecule. However, the possibility that TLR/MyD88 signaling pathways also control the induction of immunoprotective CD8+ T cell-mediated effector functions has not been investigated to date. We addressed this question by measuring the frequencies of IFN-γ secreting CD8+ T cells specific for H-2Kb-restricted immunodominant peptides as well as the in vivo Ag-specific cytotoxic response in infected animals that are deficient either in TLR2, TLR4, TLR9 or MyD88 signaling pathways. Strikingly, we found that T. cruzi-infected Tlr2−/−, Tlr4−/−, Tlr9−/− or Myd88−/− mice generated both specific cytotoxic responses and IFN-γ secreting CD8+ T cells at levels comparable to WT mice, although the frequency of IFN-γ+CD4+ cells was diminished in infected Myd88−/− mice. We also analyzed the efficiency of TLR4-driven immune responses against T. cruzi using TLR4-deficient mice on the C57BL genetic background (B6 and B10). Our studies demonstrated that TLR4 signaling is required for optimal production of IFN-γ, TNF-α and nitric oxide (NO) in the spleen of infected animals and, as a consequence, Tlr4−/− mice display higher parasitemia levels. Collectively, our results indicate that TLR4, as well as previously shown for TLR2, TLR9 and MyD88, contributes to the innate immune response and, consequently, resistance in the acute phase of infection, although each of these pathways is not individually essential for the generation of class I-restricted responses against T. cruzi.
| Innate and acquired immune responses are triggered during infection with T. cruzi, the etiologic agent of Chagas' disease, and are critical for host survival. Parasite burden is usually controlled by the time the adaptive response becomes operational. Nevertheless, T. cruzi manages to subsist within intracellular niches and establishes a chronic infection, leading to the development of cardiomyopathy in approximately one-third of infected individuals. Recently, Toll-like receptors (TLRs) have been shown to recognize T. cruzi molecules and mice lacking MyD88, the key adaptor for most TLRs, are extremely susceptible to infection. Although TLRs are known to link innate and adaptive responses, their role in the establishment of crucial effector mechanisms mediated by CD8+ T cells during T. cruzi infection has not yet been determined. We analyzed the induction of IFN-γ and cytotoxic activity in vivo in TLR2-, TLR4-, TLR9- or MyD88-deficient mice during infection, and found intact responses compared to WT mice. We also demonstrated that TLR4 is required for optimal production of inflammatory cytokines and nitric oxide and, consequently, for a better control of parasitemia levels. Understanding how TLR activation leads to resistance to infection might contribute to the development of better strategies to improve immune responses against this pathogen.
| T. cruzi is an intracellular protozoan parasite that causes Chagas' disease, an endemic disorder affecting 16–20 million people which remains a health problem in Latin America. Although both innate and acquired immune responses are triggered during early infection and are critical for host survival, around 5% of individuals die due to myocarditis during the acute phase of the disease. In most cases, despite of the immune response, T. cruzi manages to subsist within the host and in approximately 30% of infected individuals it establishes a lifelong chronic illness presenting different clinical forms, including cardiomyopathy and megasyndrome in the gut [1]. Immunopathology due to parasite persistence is considered a key element in the development of chagasic cardiomyopathy, although a secondary role for autoimmunity is not completely excluded.
Different members of the family of Toll-like receptors (TLRs), by recognizing diverse pathogen-associated molecular patterns (PAMPs) of bacterial, viral, fungal, and protozoan origin trigger the activation of innate immunity and the subsequent development of Ag-specific adaptive immunity [2]. To date, TLR2, TLR4, and TLR9 have been implicated in recognition of different T. cruzi-derived PAMPs [3]–[6]. TLR2 recognizes GPI-anchors of mucin-like proteins and the T. cruzi-released protein Tc52 [3], [4], whereas TLR4 is responsible for recognition of free glycoinositolphospholipids [5] and TLR9 is involved in recognition of the CpG motif present in T. cruzi DNA [6]. Mice deficient in MyD88, the adaptor molecule required for signaling events by most TLRs as well as IL-1R and IL-18R, show greatly enhanced susceptibility to infection with this protozoan parasite [7]. The susceptibility to infection of Tlr2−/, Tlr9−/− and Tlr2−/Tlr9−/− double knockout mice (all in the C57BL/6 background) has also been analyzed [6], [7]. Interestingly, although mice simultaneously lacking TLR2 and TLR9 are highly vulnerable to infection, their mortality rate is still less than that of Myd88−/− mice, pointing to the involvement of other TLRs and/or IL-1/IL-18 in the control of mortality.
In addition to MyD88-dependent activation, another transduction pathway is involved in signaling through TLR3 and TLR4. This pathway is mediated by the TIR domain-containing adaptor inducing IFN-γ (TRIF). Interestingly, Myd88−/−Trif−/− and Myd88−/−Ifnar−/− double knock out mice were even more sensitive to in vivo infection with T. cruzi than Myd88−/− mice, indicating that in addition to MyD88-dependent induction of proinflammatory cytokines, the TRIF-dependent production of type I IFN also contributes resistance to T. cruzi infection [8]. In accord with this observation, we have previously demonstrated that the lack of expression of functional TLR4 in mice of C3H background caused higher parasitemia and accelerated mortality to T. cruzi infection [5], although the mechanisms by which this occurs are not yet fully determined. However, since C3H WT mice are known to be more susceptible to T. cruzi infection when compared to mice of the C57BL strains, the direct comparison between the levels of susceptibility of C3H/HeJ (TLR4-deficient) mice and the other above mentioned Tlr−/− and Myd88−/− mice is difficult to interpret.
Therefore, one of the aims of the present work was to analyze the role of TLR4 in the C57BL background in the innate response to T. cruzi. For this, host cell invasion, parasite survival and release from infected macrophages, as well as nitric oxide (NO) production were quantified in C57BL/6 (WT) and TLR4-deficient cell cultures. We also evaluated the contribution of TLR4 to the in vivo control of parasitemia levels and survival, as well as to IFN-γ and TNF-α production in the B6 and B10 backgrounds. Importantly, the participation of TLR2, TLR4, TLR9 and MyD88 in the induction of crucial effector mechanisms of the adaptive response against T. cruzi was also investigated, measured as the Ag-specific IFN-γ production and cytotoxic response mediated by CD8+ T cells in infected mice.
In order to compare the anti-T. cruzi microbicidal activity of WT and TLR4-deficient macrophages, it was first necessary to investigate whether the infection rate and parasite load were equivalent in both cases. Therefore we first compared the capacity of T. cruzi trypomastigotes to infect TLR4-deficient and WT macrophage (MO) cultures in three different genetic backgrounds: C3H, C57BL/10 and C57BL/6. Strains C3H/HeJ and C57BL/10ScN are natural mutants in which the Tlr4 gene suffered mutations that result either in a residue substitution (P712H), rendering the receptor non-functional, or a deletion, with non-expression of TLR4, respectively [9]. Engineered Tlr4−/− in the B6 background was also previously described [10]. As shown in Figure 1A, no difference in the percentage of infected macrophages or in the number of parasites per macrophage after one hour of infection could be detected between cultures from the TLR4-deficient strains and their respective WT controls. However, when non-internalized parasites were extensively washed out after 1 h of interaction and the cultures were left to continue for three more hours, a significantly higher percentage of infected MO was found in TLR4-mutant cultures (Fig. S1). This result suggests the existence of an early microbicidal mechanism which is dependent on a functional TLR4. In agreement with that, the number of trypomastigotes released in the supernatant after the parasite completes its intracellular cycle in long-term cultures is significantly higher in the TLR4-deficient MO cultures (Fig. 1B–G). This is true for T. cruzi Y (Fig. 1C–G) and CL strains (Fig. 1B) and for both resident and elicited macrophages (Fig. 1F and 1G). Together these results indicated that although cell invasion by the parasite is not affected by the absence of a functional TLR4, T. cruzi growth is favored in TLR4-deficient MO, possibly due to a defective early anti-trypanosomacidal mechanism in TLR4-deficient MO.
The expression of fluorescent TLR4 in cell lines allowed us to map TLR4 subcellular location, demonstrating its presence on the cellular surface and in the Golgi, similar to the TLR4 distribution observed in human monocytes [11]. It is also known that early after cell invasion the T. cruzi localizes in a host cell vacuole which fuses with peripheral lysosomes. HEK293 cells stably transfected with TLR4-yellow fluorescent protein and MD-2 (HEK-TLR4YFP) were infected with labeled T. cruzi trypomastigotes and 2.5 h later we analyzed parasite-TLR4 co-localization by confocal microscopy. Staining these infected cells with a lysosome probe also revealed that T. cruzi-TLR4 co-localization occurs in acidic compartments (Figure 2).
Both NO and reactive oxygen species (ROS) have been shown to mediate T. cruzi killing [12]–[15]. Thus, we next analyzed the effects of adding NO and/or ROS inhibitors to the MO cultures during infection. Figure 3A shows that the addition of desferroxamine (DFO), an iron chelator which can also act as a free radical scavenger [16], causes a significant increase in the percentage of infected WT MO. This treatment abolishes the otherwise significant difference found in the percentage of infected MO between the non-treated WT and TLR4-deficient cultures. The same results are obtained when the inducible NO synthase (iNOS) inhibitor L-NMMA, or the combination of DFO and L-NMMA are added to these cultures. In order to further confirm the relevance of this early microbicidal mechanism, absent in TLR4-deficient MO, we tested the effect of inhibiting NO production in long term MO cultures, in which the number of parasites released by infected cells in the supernatant was evaluated several days after initial infection. As shown in Figure 3B, while iNOS inhibition had no effect in the number of parasites released by Tlr4−/− MO cultures, the addition of L-NMMA to the infected WT MO raised the number of trypomastigotes found in the supernatants to the levels observed in the Tlr4−/− MO cultures. In contrast, the addition of rIFN-γ an iNOS inducer to the MO cultures from the beginning of infection reduces the quantity of free trypomastigotes and results in equal numbers of released parasites from WT or TLR4-deficient MO (Fig. S2). Together, these results strongly suggest that the early trypanosomacidal mechanism absent in TLR4-deficient MO depends on ROS and NO induction.
We next compared parasitemia and mortality between different pairs of WT and TLR4-deficient mice (B10 or B6 versus B10/ScN or Tlr4 KO, respectively), after i.p. infection with 2×103 T. cruzi strain Y bloodstream trypomastigotes. Results in Figures 4A-C and 4D-F show that in both cases we found significantly higher parasitemia levels in TLR4-deficient mice, although the levels of parasites in the blood returned to very low or undetectable levels by day 11–12 post-infection and did not rise again, differently from what was previously described for C3H/HeJ mice, in which parasitemia levels were not controlled after day 15 pi [5]. We further monitored the mortality after infection and found that TLR4-deficiency in both B10 and B6 backgrounds results in higher lethality. Statistically significant differences, however, were consistently found only when comparing B10 and B10/ScN mice, while results with B6 and Tlr4−/− were more variable and did not reach statistical significance (Figure 4G and H). Of note, these results were obtained in male mice of 6–7 weeks of age, while in older TLR4-deficient mice the higher susceptibility could not be observed (data not shown). We have also performed experiments with lower (102) and higher (104 and 105) doses of infective T. cruzi forms/mice obtaining the same results (data not shown). Therefore, mice lacking TLR4 expression in a C57BL genetic background are more sensitive than their WT controls to infection with T. cruzi, although these strains do not display the uncontrolled parasitemia and the remarkable earlier mortality previously observed in the TLR4-mutant C3H/HeJ mice [5].
We then analyzed whether a lower NO production by the infected TLR4-deficient mice could explain their higher susceptibility to infection as suggested by the results obtained in vitro. In accordance with that hypothesis, the production of NO (inferred from nitrite levels in the supernatants) by spleen cells from infected TLR4-deficient mice was significantly reduced when compared with NO released by spleen cells from infected WT mice at day 10 post-infection (Figure 5A and B). Nitrite levels are also lower in the sera of TLR4-deficient mice, compared to WT animals, at this time point of infection (data not shown). Furthermore, the in vivo blockade of NO production in T. cruzi infected animals, by injection of the inducible NO synthase (iNOS) inhibitor aminoguanidine (AG) in the early phase of acute infection, brought the parasitemia and mortality of treated WT mice to the same levels obtained in treated Tlr4−/− animals (Figure 5C and D). In animals injected every other day with AG, following a previously reported protocol [17], parasitemia kept rising until treatment was stopped on day 13 pi, attaining 3 and 7 fold higher levels of what was usually observed in Tlr4−/− and WT non-treated animals, respectively (Figure 5C). This is due to the prevention of all NO generation, as for example in response to TLR2 and/or TLR9 signaling pathways, rather than exclusive inhibition of NO triggered by TLR4 engagement. Also, while non-treated infected animals usually die only after day 20 pi, earlier mortality was observed among AG-treated mice, with 50% mortality in both Tlr4−/− and WT AG-treated groups by day 12 pi (Figure 5D). Hence, these results suggest that the lower NO production due to the absence of TLR4 expression during the early phase of infection with T. cruzi is responsible for the higher sensitivity observed in Tlr4−/− mice.
As IFN-γ is thought to be the most important inducer of iNOS in macrophages and thus essential for mediation of NO-dependent parasite control during acute infection [18], we quantified IFN-γ production by spleen cells from WT and TLR4-deficient infected mice. As shown in Figure 6A, higher IFN-γ levels are indeed secreted by WT infected splenocytes at day 10 pi. We also compared the secretion of another crucial cytokine for iNOS expression and host resistance to T. cruzi, TNF-α [14] As shown in Figure 6B, the levels of TNF-α secreted by splenocytes from infected WT mice are also significantly higher than from TLR4-deficient mice. Both IFN-γ and TNF-α can be produced after T. cruzi-induced triggering of the innate immune response (mainly by NK/NKT cells and macrophages/DC respectively), as well as by CD8+ and CD4+ T lymphocytes later in the infection course, as part of the acquired response to the parasite. Since several previous studies have demonstrated the importance of IFN-γ secretion by CD8+ T cells in resistance to infection with T. cruzi [19], we asked if the frequency of Ag-specific IFN-γ-secreting cells would be altered in the absence of TLR4 expression. To do so, a previously defined H-2Kb-restricted epitope (PA8) derived from the amastigote surface protein-2 (ASP2), which is a member of the trans-sialidase family of surface proteins, was employed in ELISPOT assays [20]. However, as shown in Figure 6C no significant difference in the frequency Ag-specific IFN-γ secreting cells could be observed between WT and TLR4-deficient infected mice. Two other previously described trans-sialidase-derived peptides TSKB20 and TSKB18 [21] were alternatively employed in ELISPOT assays, giving the same results (not shown). The frequency of IFN-γ secreting CD4+ and CD8+ T cells in the spleens of WT and TLR4-deficient mice at day 13 pi was also investigated by intracellular staining and results are shown in Figure 6D-G. These data show that the frequencies of CD4+ and CD8+ T cells secreting IFN-γ in response to T. cruzi-derived antigens are not reduced in Tlr4−/− mice.
At this point, the present investigation was extended to compare the frequency of PA8/Kb-specific IFN-γ secreting lymphocytes between WT and mice which are deficient in other members of the TLR family, as well as Myd88−/− mice, whose susceptibility to T. cruzi infection was previously described [6], [7]. To our surprise, the frequency of these important effector cells of the acquired response is not altered in the spleens of Tlr2−/−, Tlr9−/− or Myd88−/− mice, as with Tlr4-deficient mice (Figure 7A). We also estimated the percentage of IFN-γ secreting CD8+ T cells in the spleen of infected Myd88−/− mice at day 10 pi by intracellular cytokine staining (ICS) following in vitro stimulation with PA8 peptide and obtained the same results, that is, no significant difference in the frequency of IFN-γ+CD8+cells between WT B6 and Myd88−/− mice (Fig. 7B–D). In order to further evaluate the adaptive response to T. cruzi in Myd88−/− mice, the percentage of IFN-γ secreting T cells was measured by ICS in both CD4+ and CD8+ subsets at days 11 and 13 pi. As shown in Figure 8, although the absence of MyD88 signaling strongly affects the percentage of CD4+ IFN-γ cells in these mice, MyD88 expression is not essential for the differentiation of IFN-γ producing CD8+ T cells specific against T. cruzi-derived epitopes. Finally, we tested whether expansion of specific CD8+ cytotoxic T cells was affected in any of the above Tlr−/− or Myd88−/− mice. For this purpose, we used a functional cytotoxic assay which measures the in vivo elimination of target cells (total splenocytes) coated either with PA8 (Figs. 9A and C), TSKB20 or TSKB18 (Fig. 9B) peptides, as previously described [21]. The phenotype of effector cells mediating peptide-specific in vivo cell killing was established earlier as being CD8+ T cells [22]. The kinetics of Ag-specific cytotoxic CD8+ T cell development during infection with the Y strain of T. cruzi in mice was also previously determined, showing that the maximum cytotoxicity (close to 100% specific lysis) is attained around day 15 pi and continued at a high level in B6 mice, even until 100 days after challenge [22]. As shown in Fig. 9B, at day 20 post-infection, no difference in peptide-specific cytotoxicity could be detected between Tlr4−/− and WT mice for any of the tested peptides. The same was true for Myd88−/− mice, in which the in vivo cytotoxicity assay was performed at an earlier post-infection time point (day 10 pi) due to their earlier mortality [7] (Fig. 9B). A summary of the cytotoxicity experiments is shown in Figure 9C, where the results of specific killing obtained with the PA8 immunodominant peptide in Tlr2−/−, Tlr4 −/−, Tlr9−/− or Myd88−/− mice are compared to B6 controls. No difference in the levels of specific cytotoxicity was observed in any of these deficient mice. Together, these results clearly indicate that deficiency in TLR2, TLR4, TLR9 or even MyD88 expression does not impair CD8+ T cell effector responses during infection with T. cruzi.
Different T. cruzi-derived molecules are able to induce host innate immune responses through the activation of different members of the TLR family, [3]–[6], [23], including glycoinositolphospholipids derived from the parasite membrane which induce a pro-inflammatory response through the TLR4 pathway [5], [23]. Thus, the documented high susceptibility of Myd88−/− mice to infection with T. cruzi could not be attributed to a single TLR, suggesting that different members of the TLR family act in concert in determining resistance to the pathogen [6]. Bafica and collaborators have shown that doubly deficient Tlr2−/− Tlr9−/− mice, although more susceptible than the single TLR2- or TLR9-deficient mice, do not display the acute mortality exhibited by Myd88−/− mice, suggesting that additional TLR/IL-1R family members are involved in the protection against infection with T. cruzi in mice [6]. In this context, the contribution of TLR4 signaling to control of the parasite burden in the C57BL/6 background was not investigated until the present study, as the only previous work on the subject was performed in mice of a different genetic background [5]. Importantly, the present work is the first to study the contribution of the different TLR and MyD88 pathways to the development of anti-T. cruzi responses mediated by CD8+ effector T cells, a critical element of the acquired immune response to the parasite.
The first question we addressed was to assess the role of TLR4 in T. cruzi internalization and triggering of very early microbicidal activity by macrophages. Infective T. cruzi trypomastigotes invade host cells using at least two different strategies; either by an active process recruiting host-cell lysosomes to the area of parasite cell contact or by an alternative pathway, in which the parasite infects phagocytic cells through conventional phagocytosis/endocytosis mechanism [24]–[26]. In both cases, the parasite may escape to the cytoplasm where it differentiates into the aflagellated amastigote form and begins intracellular replication. During cell invasion, T. cruzi interacts with different macrophage receptors to induce its own phagocytosis, but the nature of those receptors and the molecular mechanisms involved are still poorly understood. Although the general current view is that TLRs do not function directly as phagocytic receptors [27], a recent report indicated that during the invasion of T. cruzi, the activation of the Rab5-dependent phagocytic pathway is regulated by signals emanated through the parasite interaction with TLR2 in macrophages [28]. Our present results with Tlr4-deficient macrophages from three different mouse strains show that internalization of T. cruzi by macrophages is not affected by the absence of functional TLR4 expression. Some studies on the other hand, have demonstrated that TLR signaling by means of MyD88 can enhance phagosome acidification and function, the so-called phagosome maturation, which is required for effective sterilization of its contents [29]. In accord with those results, we found that after 2.5 h of infection, TLR4 and parasite co-localize into acidic compartments. Also, 4 h after infection, the percentage of TLR4-deficient macrophages infected with T. cruzi is significantly higher when compared to WT cells. The addition of iNOS or ROS inhibitors abolished the difference in the frequency of infected macrophages between cultures from TLR4-deficient and WT origin, indicating that this early trypanosomicidal mechanism triggered by TLR4 depends on the production of reactive nitrogen intermediates (RNI) and ROS, which have been described to participate in the microbicidal activity against T. cruzi and other pathogens [12]–[15]. Moreover, the fact that the simultaneous usage of iNOS and ROS inhibitors did not increase further the percentage of infected macrophages, suggests that the peroxynitrite anion (ONOO−), a strong oxidizing and against T. cruzi, formed by the reaction between nitric oxide (NO) and superoxide radical (O−2), may be the main species responsible for the elimination of T. cruzi, as described [30], [31]cytotoxic effector molecule. Therefore, TLR4 signaling triggers an important early parasiticidal event against T. cruzi, which is dependent on the formation of NO and ROS. Significantly lower production of NO was also found in splenocyte cultures from Tlr4−/− mice at day 10 post infection. In conformity to these results, we demonstrated that Tlr4−/− splenocyte cultures produce lower levels of the main iNOS inducer cytokines, IFN-γ and TNF-α. As with the in vitro results obtained with macrophage cultures, the inhibition of NO production during in vivo infection made WT and Tlr4−/− mice equally susceptible, as measured by mortality and parasite levels in the blood. Our results are in agreement with previous studies demonstrating that mice deficient for inducible nitric oxide synthase (iNOS) are highly susceptible to T. cruzi [32], and that the inhibition of iNOS from the beginning of infection lead to an increase in trypomastigotes in the blood and to high mortality [15], [33]. Together, our results point to a significant contribution of the TLR4 pathway to the innate immune response against T. cruzi infection, with the production of NO playing a major role.
We show that mice of either B10 or B6 genetic background with TLR4 deficiency presented significantly elevated parasite numbers in the blood compared to their WT controls after in vivo infection with T. cruzi. These results are in accordance with our previous work showing higher parasitemia levels of the Tlr4 mutant C3H/HeJ mice [5], although this is more pronounced in the latter lineage. Concerning mortality, however, the absence of functional TLR4 expression in B10 or B6 mice do not lead to the acute mortality previously observed in C3H/HeJ mice [5]. Therefore, the effects of TLR4 deficiency on susceptibility to infection with T. cruzi are more evident in the C3H background. Inbred strains of mice may vary from highly resistant to highly susceptible, as reflected by parasitemia levels and survival time and, following these criteria, C3H strains have been classified as “susceptible”. Classical genetic studies previously established that the resistance to T. cruzi is governed by multiple genetic factors, including H-2-linked gene(s) [34], [35] and the combination of different alleles in a group of loci confers resistance or susceptibility to infection. Therefore, analogous to the effects due to the absence of TLR2, which only become perceptible in mice with the concomitant deficiency on TLR9 [6], we have shown here that the susceptibility resulting from the absence of TLR4 is less pronounced in the resistant B6 and B10 backgrounds, compared to C3H strains.
IFN-γ is an important mediator of resistance to T. cruzi. Besides iNOS, IFN-γ regulates the expression of a large number of genes, including chemokines and chemokine receptors, which were shown to play a role in IFN-γ-mediated protection in T. cruzi infection [18], [32], [36]. Early during infection, IFN-γ is secreted by NK cells and other cell types, as part of the innate response, and later on the infection course by activated CD4+ and CD8+ T cells. Since TLRs have been implicated in the modulation of acquired immunity against several pathogens, we have herein addressed the question of whether the frequencies of IFN-γ secreting CD8+ and CD4+ T cells are altered in Tlr4−/ mice. Our data showed no significant difference in the frequencies of IFN-γ producing CD8+ or CD4+ T cells in the spleens of Tlr4−/− and WT mice, indicating that TLR4 deficiency does not interfere with these important effectors of the acquired response against T. cruzi. Therefore, the higher level of IFN-γ detected in the supernatants of WT splenocyte cultures at day 10 pi is probably contributed by cells of the innate response. A number of different cells types may account for that and we are currently evaluating their phenotype.
On the other hand, a significant reduction in the levels of the IFN-γ+CD4+ T cell population was observed in the spleens of Myd88−/− infected mice. This finding is in line with previous in vitro experiments [6], but contrasts to the results of a recent paper, where IFN-γ production by CD4+ T cells is shown to be preserved in Myd88−/− mice infected with the Tulahuen strain of T. cruzi [37]. The reason for this apparent discrepancy is not clear and could be due to the different strain of parasite used for infection or to the method employed for CD4+ T cell re-stimulation in vitro. Therefore, our results demonstrate that although the CD4+ T cell-mediated response is substantially diminished, unaltered frequencies of CD8+ IFN-γ T cells specific for Kb-restricted T. cruzi-derived peptides are present in the spleens of Myd88−/− mice compared to WT controls at days 10, 11 and 13 pi.
CD8+ T cell mediated responses are a critical component of protective immunity in T. cruzi infection, since in their absence mice quickly succumb to the infection or develop a more severe chronic disease (reviewed in [19]). Moreover, CD8+ T cells can be induced by vaccination to provide protection from lethal infection [38]. CD8+ T cells can control infection via a number of mechanisms: in addition to the already discussed secretion of IFN-γ inducing microbicidal activity in the host cell, the direct cytotoxic function against cells infected with T. cruzi is also a main effector response. Therefore, it is an important issue to define whether TLR-MyD88 mediated pathways can play a role in the priming and/or control of the cytotoxic T cell response against T. cruzi-infected targets. According to the present paradigm, this could be mainly achieved by the engagement of TLRs on antigen-presenting cells (APCs) such as dendritic cells (DCs), promoting upregulation of co-stimulatory molecules, enhancement of antigen processing and presentation, as well as secretion of Th1 polarizing pro-inflammatory cytokines by the DCs [2]. We demonstrated here, however, that the cytotoxic response mediated by CD8+ T cells against H-2Kb-restricted immunodominant peptides in T. cruzi infected mice is not dependent on TLR2, TLR4 nor TLR9 expression. Unexpectedly, the Ag-specific cytotoxic function was also preserved in Myd88−/− mice. As the cytotoxic response against the immunodominant PA8 T. cruzi epitope tested here was previously shown to be dependent on MHC class II restricted CD4+ T cells [39], our results indicate that although diminished in frequency, the residual response of CD4+ activated T cells observed in infected Myd88−/− mice is sufficient for their licensing function, which results in the development of parasite-specific CD8+mediated cytotoxic response. A first possible interpretation of these results is that none of the tested TLR and MyD88 pathways are involved in the generation of cytotoxic CD8+ T cells during T. cruzi infection. In fact, other signaling molecules and innate recognition systems might contribute to adaptive immunity to T. cruzi as the members of the Nod-like receptor protein (NLR) family [40]. Other examples are: 1) the release of pro-inflammatory bradykinin peptide by the parasite proteases during infection and consequent DC maturation induced by bradykinin B2 receptors (B2R) [41] and 2) the recently described DC maturation induced by NFATc1 activation and consequent IFN-γ production in a TLR-independent pathway [37]. However, to date, it was not determined if these TLR-independent pathways can fully account for the preserved CD8+ T cell cytotoxic response against T. cruzi-infected targets observed in Myd88−/− mice.
In our opinion, there is another plausible hypothesis to be considered for explaining the preserved CD8+ T cell cytotoxic response in Tlr2−/−, Tlr4−/−, Tlr9−/− or Myd88−/− mice: it is known that type I IFNs affect DC maturation [42], [43] and can also stimulate survival, development of cytolytic function, and production of IFN-γ by CD8+ T cells [44], [45]. Moreover, mice deprived of the type I IFN receptor, Ifnar−/−, develop higher parasitemia levels in comparison with control 129Sv mice [46] and doubly deficient Myd88−/−Ifnar−/− mice are highly susceptible to infection with T. cruzi [8]. Both TLR9 and TLR4 could induce type I IFN secretion through MyD88-dependent and -independent pathways, respectively. Therefore, TLR4 and TLR9 would be redundant concerning type I IFN production and might compensate for each other's absence in Tlr4−/− or Tlr9−/− mice. The TLR4-triggered TRIF pathway is also preserved in Myd88−/− mice and its activation would lead to type I IFN secretion and DC maturation, with the consequent normal adaptive responses against T. cruzi in these mice. Testing whether the CD8-mediated cytotoxicity against T. cruzi is affected in Tlr4−/−Tlr9−/− or Tlr4−/−Myd88−/− doubly deficient mice is one of our future goals. According to this hypothesis, cytotoxic CD8+ T cells would not be preserved in doubly deficient Myd88−/−Trif−/− mice, which is in agreement with the fact that these mice are even more susceptible to infection, as indicated by accelerated mortality when compared to single Myd88−/− mice [8]. Also, in opposition to Myd88−/−, the doubly deficient Myd88−/−Trif−/− mice are not able to control the levels of parasite in the blood [8].
The maintenance of CD8+ acquired responses against T. cruzi in Myd88−/− mice finds a parallel in studies of murine infection with Toxoplasma gondii, another intracellular protozoan parasite. As for T. cruzi, multiple TLR ligands were identified in T. gondii and Myd88−/− mice were shown to be highly susceptible to infection (reviewed in [47]). Interestingly, a recent work demonstrated that a robust and protective IFN-γ response can be elicited in Myd88−/− mice infected with an avirulent T. gondii strain [48]. Therefore, the MyD88 pathway is required for innate immunity to control infection with Toxoplasma, even though adaptive immunity against the pathogen can be triggered without the need for this TLR adaptor molecule. The same picture emerges from our present results with T. cruzi. The absence of a role for the MyD88 pathway in the generation of CD8+ adaptive responses during T. cruzi infection is also in line with other previous reports analyzing the immune response to other pathogens [49], [50], including the described protective CD8+ T cell response against the intracellular bacteria Listeria [51].
The preservation of CD8+ T cell mediated effector mechanisms in MyD88-deficient mice is in agreement with the fact that despite their high mortality, these mice do succeed in controlling the number of parasites in the blood, in contrast to the even more susceptible Ifng−/−, IfngR−/−, iNos−/− or Myd88−/−Trif−/− mice [8], [32], [36]. At present we do not know why Myd88−/−mice succumb earlier than WT mice and display 100% mortality to T. cruzi infection, notwithstanding their capacity of controlling blood parasitemia [7] and their preserved CD8+ T cell-mediated responses shown here. We would like to consider four non-exclusive possibilities: First, Myd88−/−mice, whose IFN-γ levels in serum were shown to be significantly lower [7], would also be affected by the fact that several genes, like iNOS and IP-10, have been shown to be 5- to 100-fold less extensively induced by IFN-γ in macrophages lacking MyD88 expression [52]. Second, the higher susceptibility of Myd88−/− mice could be directly attributed to the defective activation of CD4+ T cells demonstrated here, as this cell population has also been demonstrated to be essential for resistance to infection [53], probably through IFN-γ and TNF-α secretion. We do not know at the present what mechanism, absent in Myd88−/− mice, affects CD4+ T cell activation. Both attenuated DC maturation due to the absence of TLRs/MyD88 triggering and the absence of IL-1R/IL-18R signaling in CD4+ Myd88−/− T cells should be considered. Third, the lower levels of CD4+ helpers might also have indirect consequences as a defect in the B cell mediated response, which was also described to be necessary for resistance to the parasite [54]. In fact, although controversy exists, the requirement of TLR-MyD88 signaling for the generation of T-dependent antigen-specific antibody responses was proposed [55], [56] and, interestingly, antibody responses against different virus are altered or completely lost in Myd88−/− mice [57]–[59]. Finally, another possible consequence of a deficiency in CD4+ cell activation could be the defective migration of CTLs into peripheral sites of infection distinct of the spleen (liver and heart, for example), as recently demonstrated by Nakanishi et al. in a mouse model of herpes simplex virus (HSV) infection of the vagina [60].
In summary, the results obtained in the present study strongly argue in favor of a role for the TLR4 signaling in the innate immune response against T. cruzi displayed by B6 mice. Notably, we have also shown here that neither the absence of TLR2, TLR4 or TLR9 individually, nor the ablation of all MyD88-mediated pathways affect the development of cytotoxic and IFN-γ-producing CD8+ T cells, which are crucial effector mechanisms against this parasite. Determining precisely how TLR-TRIF-MyD88 activation contributes to trigger protective immunity against T. cruzi will be of critical relevance for vaccine development against this important human parasite.
All animal experiments were approved by and conducted in accordance with guidelines of the Animal Care and Use Committee of the Federal University of Rio de Janeiro (Comitê de Ética do Centro de Ciências da Saúde CEUA -CCS/UFRJ). Tlr2−/−, Tlr4−/−, Tlr9−/− and Myd88−/− mice were generated by and obtained from Dr. S. Akira (Osaka University, Japan) via Dr. R. T. Gazzinelli (Federal University of Minas Gerais, Brazil). Tlr2−/− and Tlr4−/− mice were maintained along C57BL/6 mice at the Laboratório de Animais Transgênicos (LAT, IBCCF°, UFRJ, RJ, Brazil). C3H/HeJ and C3H/HePas mice were from ICB, Universidade de São Paulo (USP, SP, Brazil). C57BL/10 and C57BL/10ScN mice were maintained at the Biotério of the Department of Immunology (IMPPG, UFRJ, RJ, Brazil). Tlr9−/− and Myd88−/− mice were maintained at the Centro de Pesquisas René Rachou (FIOCRUZ, MG, Brazil).
Mice used for experiments were sex- and age-matched, and housed with a 12-h light-dark cycle. Bloodstream trypomastigotes of the Y strain of T. cruzi [61] were obtained from BALB/c mice infected 7 days earlier. The concentration of parasites was estimated and each mouse (at least 4 per group) was inoculated intraperitoneally (i.p.) with 0.2 ml (2×103 trypomastigotes). Parasitemia was monitored by counting the number of bloodstream trypomastigotes in 5 µl of fresh blood collected from the tail vein. Mouse survival was followed daily.
Resident or elicited macrophages (obtained from the peritoneal cavity on day 4 after injection of 2.5 ml of 3% thioglycollate) were plated in triplicates and infected with trypomastigotes at a 1∶10 (macrophage:trypomastigote) ratio. After 1 h of infection, the cells were washed four times with PBS to remove the extracellular parasites and cultured in DMEM supplemented with 10% FBS (GIBCO, Invitrogen) for the indicated time periods at 37°C in an atmosphere containing 5% CO2. Trypomastigotes in the culture supernatants were counted microscopically in triplicates. Alternatively, extracellular parasites were removed by repeated washing after 1 h of infection and the cells were either washed, fixed and stained with Giemsa or cultured for a further 4 h in DMEM supplemented with 10% FBS before fixation and staining. In other experiments, macrophages were infected for 1 h in the presence of L-NMMA (1 mM) and/or DFO (100 µM); after washing, cells were cultured for further 4 h in DMEM supplemented with 10% FBS in the presence of L-NMMA (1 mM) and/or DFO (100 µM) and subsequently fixed and stained with Giemsa. The percentage of infected macrophages and the intracellular parasite numbers in 100 macrophages were counted under a light microscope.
The stable cell line of HEK293 cells expressing the fluorescent protein TLR4YFP and MD-2 constructs were described previously [11] and kindly donated by Dr D. Golenbock (University of Massachusetts Medical School, MA). Trypomastigotes were labeled with TO-PRO-3 (1 µl/ml, Molecular Probes, Invitrogen) for 30 min at RT, washed twice and then cultured with HEK- TLR4YFP cells at 10∶1 ratio for 1 hour at 37°C, 5% CO2. After repeated washing with PBS for extracellular parasite removal, cells were stained with LysoTracker Red probe (75 nM, Molecular Probes, Invitrogen) for 1 hour. Cells were then washed in PBS containing 1 mM MgCl2 and 1 mM CaCl2 and fixed in 3,7% paraformaldehyde/PBS for 15 min. Confocal microscopy was performed with a Zeiss Axiovert 200-M inverted microscope equipped with an LSM 510 Meta laser-scanning unit. Image analysis was performed with LSM 510 software (Zeiss).
The Griess reaction was performed to quantitate nitrite concentrations in the supernatant of macrophage or spleen cell cultures, as previously described [62]. Briefly, 50 µl of sample plus 50 µl of Griess reagent were incubated for 10 min at RT, followed by detection at 550 nm in an automated ELISA plate reader. The results are expressed in units of micromolar, and were determined comparing the absorbance readings of the experimental samples to a sodium nitrate standard curve.
Inhibition of iNOS in vivo was performed by injecting mice i.p. with 50 mg/kg of aminoguanidine/body weight, (AG, Sigma-Aldrich, St. Louis), diluted in sterile phosphate-buffered saline (PBS), every other day, as previously described [20]. Treatment started 4 h before infection with T. cruzi (performed as described above), and animals were treated until day 13 pi. Control mice received the same volume (200 µl) of PBS. A third group of mice received AG only.
The ELISPOT assay was performed essentially as described earlier [63]. Briefly, the preparation of plates was done by coating 96-well nitrocellulose plates Multiscreen HA (Millipore) with 60 µl/well of sterile PBS containing 10 µg/ml of the anti-mouse IFN-γ mAb R4-6A2 (BD Biosciences, San Jose, CA). After overnight incubation at RT, mAb solution was removed by sterile aspiration and the plates were washed three times with plain RPMI 1640 medium under sterile conditions. Plates were blocked by incubating wells with 100 µl RPMI medium containing 10% (v/v) FBS for at least 2 h at 37°C. Responder cells were obtained from spleens of B6 mice. Responder cells were ressuspended to a concentration of 106 viable cells per ml in RPMI medium (GIBCO, Invitrogen) supplemented with 10 mM HEPES, 2 mM L-glutamine, 5×10−5 M 2-β-mercaptoethanol, 1 mM sodium pyruvate, 100 U/ml of penicillin and streptomycin, 10% (v/v) FBS (all purchased from GIBCO, Invitrogen). B6 spleen cells adjusted to a concentration of 4×106 viable cells per ml were used as antigen presenting cells after incubation or not with the synthetic peptide at a final concentration of 10 µM for 30 min at 37°C. One hundred microliters of suspension containing responders or antigen presenting cells were pipetted into each well. The plates were incubated for 24 h at 37°C in an atmosphere containing 5% CO2. After incubation, the bulk of cultured cells was flicked out. To remove residual cells, plates were washed 3 times with PBS and 3 times with PBS-Tween. Each well received 75 µl of biotinylated anti-mouse IFN-γ mAb XMG1.2 (BD Biosciences) diluted in PBS-Tween to a final concentration of 2 µg/ml. Plates were incubated overnight at 4°C. Unbound antibodies were removed by washing the plates at least 6 times with PBS-Tween. Peroxidase-labeled streptavidin (KPL) was added at a 1∶800 dilution in PBS-Tween in a final volume of 100 µl/well. Plates were incubated for 1–2 h at RT and then washed three to five times with PBS-Tween and three times with PBS. Plates were developed by adding 100 µl/well of peroxidase substrate (50 mM Tris–HCl at pH 7.5 containing 1 mg/ml of DAB and 1 µl/ml of 30% hydrogen peroxide solution, both from Sigma). After incubation at RT for 15 min, the reaction was stopped by discarding the substrate solution and rinsing the plates under running tap water. Plates were dried at RT and spots were counted with the aid of a stereomicroscope (Nikon) or in the ImmunoSpot® Analyzer (Cellular Technology Ltd., Shaker Heights, OH, USA). Results of the ELISPOT assay are representative of two or more independent experiments.
Tissue culture trypomastigotes of the Y strain of T. cruzi were transformed to amastigotes in acidic DMEM/10% FCS for 24 h at 37°C, as previously described [64]. Parasites were pelleted, washed in PBS, and subjected to more than five rounds of freeze-thawing followed by sonication. Cellular debris were removed by centrifugation at 12,000 rpm, and the soluble fraction was boiled for 5 min to denature the proteins. Protein concentrations were determined using a Bio-Rad protein assay.
Splenocytes isolated from infected mice were cultured either with T. cruzi amastigote extract at 10 µg/ml (see above) or with PA8 peptide (VNHRFTLV) at 10 µM, or left unstimulated, for 5 h to 14 h at 37°C in the presence of brefeldin A (Sigma-Aldrich). Cells were surface stained with anti-CD8-PerCP and anti-CD4-FITC (BD Biosciences) and fixed for 10 minutes with a solution containing PBS, 4% paraformaldehyde at RT. Then, cells were permeabilized for 15 minutes with PBS, 0.1% bovine serum albumine, 0.1% saponin (Sigma-Aldrich). Intracellular cytokine staining was performed with anti- IFN-γ -PE (BD Biosciences). At least 10,000 gated CD8+ lymphocyte events were acquired. Analytical flow cytometry was conducted with a FACSCalibur (BD Biosciences) and the data were processed with CellQuest software (BD Biosciences).
For the in vivo cytotoxicity assays, splenocytes of the different mouse strains were divided into two populations and labeled with the fluorogenic dye CFSE (Molecular Probes, Invitrogen) at a final concentration of 5 µM (CFSEhigh) or 0.5 µM (CFSElow). CFSEhigh cells were pulsed for 40 min at 37°C with 1–2.5 µM of either H-2Kb -restricted ASP-2 peptide, also called PA8, (VNHRFTLV), H-2Kb- restricted TsKb-18 peptide (ANYDFTLV) or H-2Kb- restricted TsKb-20 peptide (ANYKFTLV). CFSElow cells remained unpulsed. Subsequently, CFSEhigh cells were washed and mixed with equal numbers of CFSElow cells before injecting i.v. 15–20×106 total cells per mouse. Recipient animals were mice that had been infected or not with T. cruzi. Spleen cells of recipient mice were collected 20 h after transfer, fixed with 2.0% paraformaldehyde and analyzed by cytometry, using a FACSCalibur Cytometer (BD Biosciences). Percentage of CFSElow (M1) and CFSEhigh (M2) cells were obtained using CellQuest software (BD Biosciences). Percentage of specific lysis was determined using the formula: 1 - ((M2infected/M1infected)/(M2naïve/M1naive)) ×100%.
Statistical analyses were performed using GraphPad Prism version 4.00 for Windows (GraphPad Software, San Diego California USA, www.graphpad.com). Data were compared using a two-tailed Student's t test and are expressed as mean ± SEM. Data were considered statistically significant if p values were <0.05. The LogRank test was used to compare the mouse survival rate after challenge with T. cruzi. The differences were considered significant when the P value was <0.05.
|
10.1371/journal.pcbi.1002293 | A Tale of Two Stories: Astrocyte Regulation of Synaptic Depression and Facilitation | Short-term presynaptic plasticity designates variations of the amplitude of synaptic information transfer whereby the amount of neurotransmitter released upon presynaptic stimulation changes over seconds as a function of the neuronal firing activity. While a consensus has emerged that the resulting decrease (depression) and/or increase (facilitation) of the synapse strength are crucial to neuronal computations, their modes of expression in vivo remain unclear. Recent experimental studies have reported that glial cells, particularly astrocytes in the hippocampus, are able to modulate short-term plasticity but the mechanism of such a modulation is poorly understood. Here, we investigate the characteristics of short-term plasticity modulation by astrocytes using a biophysically realistic computational model. Mean-field analysis of the model, supported by intensive numerical simulations, unravels that astrocytes may mediate counterintuitive effects. Depending on the expressed presynaptic signaling pathways, astrocytes may globally inhibit or potentiate the synapse: the amount of released neurotransmitter in the presence of the astrocyte is transiently smaller or larger than in its absence. But this global effect usually coexists with the opposite local effect on paired pulses: with release-decreasing astrocytes most paired pulses become facilitated, namely the amount of neurotransmitter released upon spike i+1 is larger than that at spike i, while paired-pulse depression becomes prominent under release-increasing astrocytes. Moreover, we show that the frequency of astrocytic intracellular Ca2+ oscillations controls the effects of the astrocyte on short-term synaptic plasticity. Our model explains several experimental observations yet unsolved, and uncovers astrocytic gliotransmission as a possible transient switch between short-term paired-pulse depression and facilitation. This possibility has deep implications on the processing of neuronal spikes and resulting information transfer at synapses.
| Synaptic plasticity is the capacity of a preexisting connection between two neurons to change in strength as a function of neuronal activity. Because it admittedly underlies learning and memory, the elucidation of its constituting mechanisms is of crucial importance in many aspects of normal and pathological brain function. Short-term presynaptic plasticity refers to changes occurring over short time scales (milliseconds to seconds) that are mediated by frequency-dependent modifications of the amount of neurotransmitter released by presynaptic stimulation. Recent experiments have reported that glial cells, especially hippocampal astrocytes, can modulate short-term plasticity, but the mechanism of such modulation is poorly understood. Here, we explore a plausible form of modulation of short-term plasticity by astrocytes using a biophysically realistic computational model. Our analysis indicates that astrocytes could simultaneously affect synaptic release in two ways. First, they either decrease or increase the overall synaptic release of neurotransmitter. Second, for stimuli that are delivered as pairs within short intervals, they systematically increase or decrease the synaptic response to the second one. Hence, our model suggests that astrocytes could transiently trigger switches between paired-pulse depression and facilitation. This property explains several challenging experimental observations and has a deep impact on our understanding of synaptic information transfer.
| Activity-dependent modification of synaptic transmission critically moulds the properties of synaptic information transfer with important implications for computation performed by neuronal circuitry [1]–[4]. Multiple mechanisms could coexist in the same synapse, regulating the strength or the efficacy of synaptic transmission therein in a way that depends on the timing and frequency of prior activity at that same synaptic terminal [5].
One widely studied mechanism responsible for the dependence of synaptic transmission on past activity has been dubbed presynaptic short-term plasticity [6]. Upon repetitive action potential stimulation, the response of a presynaptic terminal – usually assessed as the amount of neurotransmitter molecules released from this latter – will not follow with uniform strength but will be modified in a time- and activity-dependent manner, leading either to facilitation or to depression of synaptic release, or to a mixture of both [2]. Such stimulus-related variations of presynaptic response can span a time scale from few milliseconds to seconds from the stimulus onset [2], [7] and fade away after sufficiently prolonged synaptic inactivity [3], [5].
The ability of a presynaptic terminal to convey stimulus-related information is determined by the probability to release neurotransmitter-containing vesicles upon arrival of action potentials [3], [6]. The release probability depends on the number of vesicles that are ready to be released, i.e. the readily releasable pool, but also on the state of the calcium (Ca2+) sensor for the exocytosis of synaptic vesicles [8]. On the mechanistic level, both the finite size and the slow post-stimulus recovery of the readily releasable pool, that is the reintegration of the content of synaptic vesicles, give rise to the phenomenon of short-term presynaptic depression, with the extent of depression being determined by the frequency of prior synaptic stimulation [9]. The dependence of short-term facilitation on the pattern of synaptic activation is likely determined either by the slow removal of free presynaptic residual Ca2+ or by the slow unbinding of this latter from the Ca2+ sensor [3], although these issues are still debatable [10], [11].
Given the important role assumed by presynaptic short-term plasticity in neural computation [6], [12] and the variety of plastic responses – depression, facilitation or both – exhibited by central synapses [13], [14], it is important to unravel the mechanisms that might govern dynamical transitions between depressing and facilitating synapses. The goal of the present work was to investigate one such candidate mechanism: modulation of presynaptic plasticity by glial cells and astrocytes in particular.
Recent years have witnessed mounting evidence on a possible role of glial cells in the dynamics of neuronal networks [15]. In particular, the specific association of synapses with processes of astrocytes – the main type of glial cells in the hippocampus and the cortex [16]–[18] – together with the discovery of two-way astrocyte-neuron communication [19], [20], suggest an active role of these cells in modulation of synaptic transmission and information processing in the brain [21].
Astrocytes could modulate synaptic transmission at nearby synapses by releasing neurotransmitter (or “gliotransmitter”) in a Ca2+-dependent fashion [22]. In the hippocampus in particular, several studies have shown that astrocyte-released glutamate modulates short-term plasticity at excitatory synapses either towards depression or facilitation [23]–[25]. This is achieved by activation of presynaptic glutamate receptors [26] (see also Figure 1 for a schematic presentation). Thus, astrocytes are equipped with means to modulate the extent to which presynaptic terminal exhibits short-term depression or facilitation in response to sustained rhythmic stimulation [27].
We devised a biophysically plausible computational model to investigate the characteristics of astrocyte modulation of presynaptic short-term plasticity. Using the model, we were able to identify the parametric regime in which the synaptic response to action potential stimulation can switch from facilitating to depressing and vice versa. This ability to switch synaptic modus operandi depended critically on the characteristics of astrocyte-to-synapse signaling. These findings highlight the new potential role played by astrocytes in defining synaptic short-term plasticity and could explain contradicting experimental evidences.
Although based on experimental results in the hippocampus, [28]–[34], our description could also be extended to model other recognized neuron-glia signaling pathways such as GABAergic gliotransmission on interneuron-to-pyramidal cell synapses in the hippocampus [35], glia-mediated ATP release both on hippocampal synapses [36], [37] or in the hypothalamus [38] as well as in the retina [39], and glial modulation of neuromuscular transmission [40]–[42].
Regulation of synaptic transmission by astrocyte-released gliotransmitter is supported by an elaborate signaling network schematized in Figure 1. Here, we consider the well-characterized experimental case of glutamate-mediated astrocyte regulation of synaptic transmission in the hippocampus [27], [43]. At excitatory synapses there, astrocytes can respond to synaptically-released glutamate by intracellular Ca2+ elevations that in turn, may trigger the release of further glutamate from the astrocytes [22], [44]. This astrocyte-released glutamate (GA) diffuses in the extrasynaptic space and binds to presynaptic metabotropic glutamate receptors (mGluRs) or NMDA receptors (NMDARs) on neighboring presynaptic terminals [21], [30]. Glutamate activation of these receptors can modulate Ca2+ influx into the presynaptic terminal, affecting the release probability of glutamate-containing synaptic vesicles [26]. Thus, glutamate release from the presynaptic terminal is expected to increase the astrocytic intracellular Ca2+, eventually leading to glutamate release from that astrocyte. In turn, astrocytic glutamate modulates presynaptic Ca2+ and thus affects the amount of glutamate released from that same synapse in response to action potentials that will follow [27].
Astrocyte Ca2+ dynamics may also not be modulated by glutamate originating from the very presynaptic terminal that is regulated by the astrocyte, but rather by an exogenous source [45]. This could correspond to the heterosynaptic case whereby two distinct synapses, A and B, are contacted by processes from the same astrocyte [21]. Glutamate released by the presynaptic terminal of synapse A modulates astrocytic Ca2+, leading to modulation of glutamate release from the presynaptic terminal of synapse B. Alternatively astrocyte Ca2+ dynamics could be modulated by intercellular IP3 diffusion from neighboring astrocytes through gap junctions [46] or by exogenous stimulation of the astrocyte by different techniques or external stimuli [47], [48], or occur spontaneously [49], [50].
Although both homosynaptic and non-homosynaptic scenarios equally occur physiologically [21], [45], here we focus only on the latter. This approach, which is often adopted in the majority of experiments [30]–[32], [49], presents several advantages. First, it allows us to characterize the effect of astrocytic glutamate on short-term synaptic plasticity in general, that is, independently of the nature of synaptic inputs. Second, it uses Ca2+ signals to merely trigger glutamate exocytosis from the astrocyte. Thus we can focus on the timing of glutamate release without considering the complexity of the underlying Ca2+ dynamics [48] which can be ultimately modeled by simple stereotypical analytical functions (Text S1, Section I.2). Third, it can be used in the derivation of a mean-field description of synaptic transmission [51], [52] aimed at understanding regulation of short-term synaptic plasticity by a large variety of astrocytic glutamate signals impinging on the synapse, without the need to consider an equally large number of cases.
We first studied the effect of astrocytic glutamate release on the transfer properties of our model synaptic terminal. Because the response of a synapse to action potential critically depends on the value of U0 (equation 3), which in turn could be modulated by astrocytic glutamate binding to presynaptic glutamate receptors (equation 5), we expected that the steady-state frequency response of a synapse () could also be modulated by the astrocyte-synapse signaling. Since both geometry of synaptic bouton and diffusion of glutamate in the extracellular space are beyond the scope of the present work, we implicitly assumed, based on experimental evidence [31], that the release site of astrocytic glutamate apposes targeted presynaptic glutamate receptors. When the intracellular Ca2+ in the astrocyte crossed over the threshold of glutamate exocytosis (Figure 4A, top, dashed red line), the extracellular concentration of glutamate in proximity of presynaptic receptors first increased rapidly and then decayed exponentially at rate Ωc, as a result of the concomitant uptake by astrocytic glutamate transporters and diffusion away from the site of exocytosis (Figure 4A, middle) (see also Text S1, Section I.4; Figure S5).
For α = 0, equations (5–6) predict that this glutamate peak should lead to a sharp decay of U0, followed by a slower recovery phase (Figure 4C, left). Using equation (5), we can also predict the resulting dependence of the steady-state synaptic response on the input frequency (Figure 4C, middle). In the absence of astrocytic glutamate release (thick dashed black line), monotonously decreases for increasing input frequency fin for the merely depressing synapse considered in this figure. At the release of astrocytic glutamate (Figure 4A, middle), the peak of bound presynaptic receptors (Figure 4A, bottom) and the resulting sharp drop of U0 (Figure 4C, left, black mark) induce a strong decrease of the steady-state amount of released resources at low to intermediate input frequencies (0.1–10 Hz) (Figure 4C, middle, thick red line). In addition, the steady-state response loses its monotonicity and displays a peak frequency characteristic of facilitating synapses (see “Mechanisms of short-term presynaptic plasticity” in “Methods”). The curve then slowly transforms back to its baseline form (thin colored lines) and the peak synaptic input frequency appears to progressively shift toward smaller input frequencies (thick dashed arrow). Hence, for α = 0, the limiting frequency (equation 4) is predicted to sharply increase following astrocytic glutamate release and then to slowly relax back to smaller values (Figure 4C, right).
The exact opposite picture instead describes the scenario of α = 1 (Figure 4B). In this case, the parameter U0 increases upon astrocytic glutamate release (Figure 4B, left) causing a dramatic increase of the steady-state response for a range of frequencies within 0.1–10 Hz (Figure 4B, middle). Accordingly, the limiting frequency of the synapse dramatically reduces following astrocytic glutamate release, and slowly recovers back to its baseline value (Figure 4B, right). Taken together, the above results of the mean-field analysis predict that, depending on the parametric scenario, astrocyte can either transiently decrease, when α = 0, or increase, if α = 1, the release of a model synapse.
To assess the validity of these predictions, we show in Figure 5 the responses of two different model synapses (A: depressing; B: facilitating) to Poisson spike trains delivered at frequency fin (Figure 5, top panels for specific realizations of such spike trains). To simplify the presentation, we considered the case in which a single Ca2+ peak (Figure 5, middle) is sufficient to trigger the release of glutamate from the astrocyte. The synaptic response under different scenarios of astrocytic glutamate modulation (A: α = 0; B: α = 1) is then compared to the “Control” scenario obtained for the model synapses without astrocyte. In the case of α = 0 (Figure 5A, bottom) the amount of resources released by the model synapse steeply decreased at the onset of glutamate release from the astrocyte (green area) and slowly, i.e. tens of seconds, recovered to the levels comparable to those of the control scenario (blue area). The opposite effect was observed instead for α = 1 (Figure 5B). The synaptic response in this case was strongly augmented by astrocytic glutamate (magenta area) and then slowly decayed back to the levels obtained in control conditions (Figure 5B, bottom).
Collectively our mean field analysis (Figure 4) and simulations (Figure 5) suggest that glutamate release by the astrocyte can induce STD or STP of synaptic response to action potentials (Figure S6). Which one between these scenarios occurs depends on the value of the “effect” parameter α that lumps together both the density and the biophysical properties of presynaptic receptors targeted by astrocytic glutamate. These results are consistent with a large body of experimental observations in the hippocampus, where astrocyte-released glutamate could transiently decrease [33] or increase the synaptic response to stimulation [30]–[32], [34], [49].
The character of synaptic information transfer is shaped by several factors [2]. Synaptic strength at any given moment is determined by an earlier “activation history” of that same synapse [3], [5]. Structural and functional organization of presynaptic bouton affects the release and reintegration of neurotransmitter vesicles, ultimately defining the filtering feature (depressing or facilitating) of a synapse in response to spike train stimulation [3], [68]. Existing models of synaptic dynamics assume “fixed plasticity mode”, in which the depression/facilitation properties of a synapse do not change with time. However, in biological synapses, plasticity itself seems to be a dynamic feature; for example, the filtering characteristics of a given synapse is not fixed, but rather can be adjusted by modulation of the initial release probability of docked vesicles [13]. Using a computational modeling approach, we showed here that astrocytes have the potential to modulate the flow of synaptic information via glutamate release pathway. In particular, astrocyte-mediated regulation of synaptic release could greatly increase paired-pulse facilitation (PPF) at otherwise depressing synapses (and thus switch the synapse from depressing to facilitating); conversely, it could reinforce paired-pulse depression (PPD) at otherwise facilitating synapses (therefore switching the synapse from facilitating to depressing). These findings imply that astrocytes could dynamically control the transition between different “plasticity modes”. The present model also lends an explanation to several pieces of experimental data, as we detail below.
In agreement with experimental results [26], [27], our model suggests that the type of presynaptic glutamate receptors targeted by astrocytic glutamate critically determines the type of modulation that takes place. The modulatory action of an astrocyte is lumped in our model into the so-called “effect” parameter α: lower values of α make the action of an astrocyte depressing with respect to the overall synaptic release but also increase paired-pulse facilitation. On the contrary higher α values make the effect of an astrocyte facilitating but at the same time paired-pulse depression is enhanced. Interestingly, some recent experiments on perforant path-granule cell synapses in the hippocampal dentate gyrus, show that facilitation of synaptic release mediated by astrocyte-derived glutamate correlates with a decrease of paired-pulse ratio [31]. Our model provides a natural explanation of these experimental results.
Several lines of experimental evidence suggest that different types of glutamate receptors may be found at the same synaptic bouton [26]. The different types of receptors have different activation properties and hence could be recruited simultaneously or in a complex fashion [30], [41]. Thus it is likely that α could take intermediate values between 0 and 1. In one particular scenario, concurrence of astrocyte-mediated depression and facilitation could also lead these two effects to effectively cancel each other so that no apparent modulation of synaptic release is observed. Interestingly, in some recent studies, the Ca2+-dependent release of glutamate from astrocytes was reported not to affect synaptic function [69], [70], thus questioning the vast body of earlier experimental evidence pointing to the contrary. In our model we posit that an apparent lack of astrocytic effect on synaptic function could arise when the “effect” parameter α exactly matches the basal release probability of that presynaptic terminal, that is when α = U0* (in which case equation 5 becomes U0(Γ) = α, meaning that U0 does not depend on Γ anymore). This scenario would lead to concurrence of astrocyte-mediated depression and facilitation with no net observable effect on synaptic transmission.
Whether de facto astrocytes decrease or increase synaptic release likely depends on the specific synapse under consideration and the functional implications that such different modulations could lead to [15], [21], [30]. In the former case for example, enhanced PPF could be not functionally relevant if release of neurotransmitter is strongly reduced by astrocyte glutamate signaling. In such situation the astrocyte would essentially shut down synaptic transmission, hindering the flow of information carried by presynaptic spikes [71]. On the other hand, for astrocyte-induced facilitation, an increase of released neurotransmitter could correspond to a similar increase of transmission probability [72]. However, the associated modulations of paired-pulse plasticity could also account for complex processing of specific – i.e. temporal vs. rate – features of input spike trains [2], [51], [57] that could not otherwise be transmitted by the single synapse, that is without the astrocyte.
In a recent line of experiments on frog neuromuscular junction, it was observed that glial cells could govern the outcome of synaptic plasticity based on their ability to bring forth variegated Ca2+ dynamics [40], [41]. In other words, different patterns of Ca2+ oscillations in perisynaptic glia were shown to activate different presynaptic receptors and thus to elicit different modulatory effects on neurotransmitter release [41]. This scenario would call for a future modification of our model to include a dependence on astrocytic Ca2+ dynamics of the effect parameter α. Nevertheless such observations are generally bolstered by our study. Our model predicts the existence of a threshold frequency for Ca2+ oscillations below which PPD (PPF) is predominant with respect to PPF (PPD) and above which the opposite occurs. This supports the idea that different spatiotemporal Ca2+ dynamics in astrocytes, possibly due to different cellular properties [73]–[76], could provide specialized feedback to the synapse [40]. Moreover, our model suggests that different types of presynaptic glutamate receptors might not be necessary to trigger different modulations of synaptic transfer properties. The fact that the frequency of Ca2+ oscillations could bias synaptic paired-pulse plasticity subtends the notion that not only the nature of receptors, but also the dynamics of their recruitment by gliotransmitter could be a further critical factor in the regulation of synaptic plasticity [27], [41]. This latter could eventually be dictated by the timing and the amount of released glutamate [27], [44] as well as by the ultrastructure of astrocytic process with respect to synaptic terminals which defines the geometry of extracellular space [18], [77] thus controlling the time course of glutamate therein [78].
Remarkably, the threshold frequency of Ca2+ oscillations that discriminates between PPD and PPF falls, in our analysis, within the range <2.5 min−1 of spontaneous Ca2+ oscillations displayed by astrocytes in basal conditions independently of neuronal activity [32], [49], [50], [79], [80], hinting a possible role for these latter in the regulation of synaptic physiology. Spontaneous Ca2+ oscillations can indeed trigger astrocytic glutamate release [32], [79]–[82] which could modulate ambient glutamate leading to tonic activation of presynaptic receptors [26], [83]. In this fashion, spontaneous glutamate gliotransmission could constitute a mechanism of regulation of basal synaptic release. Notably, in a line of recent experiments, selective metabolic arrest of astrocytes was observed to depress Schaffer collateral synaptic transmission towards increasing PPF, consistently with a reduction of the basal synaptic release probability as predicted by our analysis [49]. The latter could be also relevant in the homosynaptic case of astrocytic glutamate exocytosis evoked by basal activity of the same presynaptic terminal that it feeds back to [21], [27], [84]. In such conditions, the ensuing influence of astrocytic glutamate on synaptic release correlates with the incoming synaptic stimulus also through Ca2+ dynamics in the astrocyte [85], unraveling potentially new mechanisms of modulation of synaptic transmission and plasticity.
Although we focused on regulation of astrocyte at single synapses, our analysis could also apply to synaptic ensembles [51], [54] that could be “contacted” either by the same astrocytic process [30], [32], [80] or by different ones with locally synchronized Ca2+ dynamics [81]. In this case, modulation of the release probability by the astrocyte would support the existence of “functional synaptic islands” [86], namely groups of synapses, intermittently established by different spatiotemporal Ca2+ dynamics, whose transmission mode and plasticity share common features. The implications that such dynamic astrocyte-synapse interaction might have with regard to information flow in neural circuitry remain to be investigated.
Due to their capacity to modulate the characteristics of synaptic transmission, astrocytes could also alter the temporal order of correlated pre- and postsynaptic spiking that critically dictates spike-timing dependent plasticity (STDP) at the synapse [87]. Thus, astrocyte modulation of short-term plasticity could potentially contribute to ultimately shape persistent modifications of synaptic strength [49], [88], [89] underlying processing, memory formation and storage that provides the exquisite balance, subtlety and smoothness of operation for which nervous systems are held in awe [90]. Future combined physiological and computational studies will determine whether or not this is the case.
|
10.1371/journal.pbio.1001717 | HDAC4 Reduction: A Novel Therapeutic Strategy to Target Cytoplasmic Huntingtin and Ameliorate Neurodegeneration | Histone deacetylase (HDAC) 4 is a transcriptional repressor that contains a glutamine-rich domain. We hypothesised that it may be involved in the molecular pathogenesis of Huntington's disease (HD), a protein-folding neurodegenerative disorder caused by an aggregation-prone polyglutamine expansion in the huntingtin protein. We found that HDAC4 associates with huntingtin in a polyglutamine-length-dependent manner and co-localises with cytoplasmic inclusions. We show that HDAC4 reduction delayed cytoplasmic aggregate formation, restored Bdnf transcript levels, and rescued neuronal and cortico-striatal synaptic function in HD mouse models. This was accompanied by an improvement in motor coordination, neurological phenotypes, and increased lifespan. Surprisingly, HDAC4 reduction had no effect on global transcriptional dysfunction and did not modulate nuclear huntingtin aggregation. Our results define a crucial role for the cytoplasmic aggregation process in the molecular pathology of HD. HDAC4 reduction presents a novel strategy for targeting huntingtin aggregation, which may be amenable to small-molecule therapeutics.
| Huntington's disease (HD) is a late-onset neurodegenerative disorder caused by protein-folding defects in the huntingtin protein. Mutations in huntingtin can result in extra-long tracts of the amino acid glutamine, resulting in aberrant interactions with other proteins and also causing huntingtin proteins to self-associate and -aggregate. The pathology of HD is therefore associated with nuclear and cytoplasmic aggregates. HDAC4 is a histone deacetylase protein traditionally associated with roles in transcription repression. The HDAC4 protein contains a glutamine-rich domain and in this work we find that HDAC4 associates with huntingtin in a polyglutamine-length-dependent manner and that these proteins co-localise in cytoplasmic inclusions. Importantly, reducing HDAC4 levels delays cytoplasmic aggregate formation and rescues neuronal and cortico-striatal synaptic function in mouse models of HD. In addition, we observe improvements in motor coordination and neurological phenotypes, as well as increased lifespan in these mice. Nuclear huntingin aggregates or transcription regulation, however, remained unaffected when HDAC4 levels were reduced to enable these effects. Our results thus provide valuable insight into separating cytoplasmic and nuclear pathologies, and define a crucial role for cytoplasmic aggregations in HD progression. HDAC4 reduction presents a novel strategy for alleviating the toxicity of huntingtin protein aggregation, thereby influencing the molecular pathology of Huntington's disease. As there are currently no disease-modifying therapeutics available for Huntington's disease, we hope that this HDAC4-mediated regulation may be amenable to small-molecule therapeutics.
| Huntington's disease (HD) is a progressive, inherited neurological disorder characterized by severe motor, cognitive, behavioural, and physiological dysfunction for which there is no effective disease-modifying treatment [1]. The disease is caused by the expansion of a CAG repeat to more than 35 CAGs within exon 1 of the HTT gene. At the molecular level, mutant huntingtin (HTT) containing an expanded polyQ stretch has a propensity to self-aggregate to produce a wide-range of oligomeric species and insoluble aggregates and exerts a gain of toxic function through aberrant protein–protein interactions [2]. Therefore, as with other neurodegenerative diseases such as Alzheimer's disease, Parkinson's disease, and the prion diseases, the polyglutamine (polyQ) disorders including HD are associated with the accumulation of misfolded proteins leading to neuronal dysfunction and cell death.
Transcriptional dysregulation is part of the complex molecular pathogenesis of HD, to which abnormal histone acetylation and chromatin remodelling may contribute [3]. The imbalance in histone acetylation was proposed to be caused by the inactivation of histone acetyltransferases, which led to the pursuit of histone deacetylases (HDACs) as HD therepeutic targets [4],[5]. There are 11 mammalian Zn2+-dependent HDACs divided into three groups based on structural and functional similarities: class I (HDACs: 1, 2, 3, 8), class IIa (HDACs: 4, 5, 7, 9), class IIb (HDACs: 6, 10), and HDAC11 as class IV [6]. Initial genetic and pharmacological studies performed in flies, worms, and HD mouse models have suggested that HDAC inhibitors may have a significant therapeutic potential [4],[5].
The preclinical evaluation of the HDAC inhibitor suberoylanilide hydroxamic acid (SAHA) demonstrated a dramatic improvement in the motor impairment that develops in the R6/2 HD mouse model [7]. Initially, SAHA was shown to inhibit class I and II HDACs at nanomolar concentrations, although it is predominantly a class I inhibitor [8]. More recently, SAHA was shown to lead to the degradation of HDACs 4 and 5 via RANBP2-mediated proteasome degradation in cancer cell lines [9]. Following on from this, we demonstrated that in addition to its deacetylase activity and the known effect on decreasing Hdac7 mRNA levels [10], SAHA treatment results in a reduction in HDAC2 and HDAC4 in brain regions of both WT and R6/2 mice, without affecting their transcript levels in vivo. This was associated with a reduction in aggregate load and the restoration of cortical Bdnf transcript levels in R6/2 mice [11].
It is well-established that HDAC4 acts as a transcriptional repressor that shuttles between the nucleus and cytoplasm. Phosphorylated HDAC4 is retained in the cytoplasm through its association with 14-3-3 proteins [12]. The N-terminal region of HDAC4 contains a MEF2 binding site and represses the transcription of MEF2-dependent genes important in the regulation of neuronal cell death [13]. In this context, MEF2 can act as a neuronal survival factor, and its inhibition has been linked to the death of neurons in several cell culture systems [14]. Crystallization of the N-terminal domain of HDAC4 suggested that HDAC4 might have the propensity to self-aggregate through its glutamine-rich domains, consistent with cell-culture studies [15]. Interestingly, regions containing high glutamine content in proteins have been observed to facilitate interactions with other glutamine-rich proteins, leading to the spontaneous assembly of insoluble toxic amyloid-like structures in mammalian cells [16].
In this study, we identified a novel mechanism by which HDACs can modify HD pathogenesis in vivo and found that HDAC4 reduction delays the HTT aggregation process. We demonstrated that HDAC4 associates with mutant exon 1 and full-length HTT in vivo in a polyQ-length-dependent manner and co-localizes with cytoplasmic inclusions in the brains of HD mouse models. HDAC4 knock-down inhibited aggregate formation in both the R6/2 (N-terminal fragment) and HdhQ150 (full-length knock-in) HD mouse models. This delay in aggregation occurred in the cytoplasm, consistent with the subcellular localisation of HDAC4 in mouse brain. We found no evidence for HDAC4 translocation to the nucleus during disease progression, and HDAC4 knock-down had no effect on HTT aggregation in the nucleus and no impact on global transcriptional dysregulation. HDAC4 reduction led to a marked restoration of the membrane properties of medium spiny neurons (MSNs) and of corticostriatal synaptic transmission. This was associated with an improvement in neurological phenotypes and extended survival. These data provide a clear demonstration that cytoplasmic pathogenic mechanisms contribute to HD-related neurodegenerative phenotypes and identify HDAC4 as a therapeutic target for HD.
In order to investigate whether HDAC4 is involved in the molecular pathogenesis of HD, we used a genetic approach to reduce HDAC4 levels in both the R6/2 and HdhQ150 knock-in HD mouse models. R6/2 mice are transgenic for a mutated N-terminal exon 1 HTT fragment [17]. The HdhQ150 mice have an expanded CAG repeat knocked in to the mouse huntingtin gene (Htt) [18],[19], which is partially mis-spliced with the result that these mice express mutant versions of both an exon 1 HTT and a full-length HTT protein [20]. Because Hdac4 knock-out (Hdac4KO) mice die in early postnatal life [21], the HD mutation could not be transferred onto an Hdac4 null background. Therefore, we crossed males for each of the HD mouse models to Hdac4HET females (Figure 1A). Analysis of the progeny indicated that Hdac4 mRNA levels were decreased to approximately 50% in both Hdac4HET and double-mutant mice in both crosses (Figure 1B).
Since HDAC4 functions as a transcriptional corepressor [22], it was important to check whether Hdac4 knock-down regulated the expression of the R6/2 transgene, as this would be expected to modulate the onset and progression of disease in R6/2 mice. Therefore, we used Taqman qPCR to demonstrate that exon 1 HTT mRNA was not altered in the cortex (Figure 1C), cerebellum, nor striatum (Figure S1A) of Dble::R6/2 mice. Similarly, we showed that endogenous Htt levels were unchanged as a consequence of HDAC4 reduction in both R6/2 and HdhQ150 mice (Figures 1D and S1B). In addition, given that CAG repeat length is linked to aggregation kinetics and disease progression, we ensured that the CAG repeats were maintained at comparable levels throughout the course of this study (Table S1).
Alteration of HDAC4 levels has been shown to modulate Hdac9 in muscle cells [23] and Hdac5 in primary mouse hepatocytes [24]. Therefore, we used Taqman qPCR to show that Hdac4 knock-down did not affect the levels of the other 10 Hdacs in brain regions of mice that did or did not express the HD mutation (Figures 1E and S1C and S1D).
Bioinformatic predictions of HDAC4 structure suggested that HDAC4 has an N-terminal coil–coil domain within which it possesses short polyQ tracts that might convey an increased propensity for amyloid formation [15] as was confirmed in cultured cells [13]. Hence, we hypothesized that HDAC4 might exhibit pro-aggregation properties in HD mouse models. In order to investigate the molecular consequences of HDAC4 knock-down in HD mice, we employed the Seprion ligand ELISA to quantify aggregate load [25] and time-resolved Förster resonance energy transfer (TR-FRET) to measure soluble mutant HTT levels [26]. TR-FRET detects a FRET signal between two appropriately labelled antibodies. In this case, 2B7 (epitope: 1–17 amino acids of HTT) is paired with MW1 (epitope: polyQ in nonaggregated HTT).
In R6/2 mice, the level of soluble exon 1 HTT decreases with disease progression as a consequence of its aggregation. The Seprion ELISA revealed that HDAC4 knock-down reduced the aggregate load in the cortex (Figure 2A), brain stem, hippocampus, and cerebellum (Figure S2A) of Dble::R6/2 mice at 4 and 9 wk but that this effect had diminished by 15 wk of age. Accordingly, TR-FRET demonstrated that reduced HDAC4 levels led to an increase in soluble exon 1 HTT in the cortex (Figure 2B), brain stem, hippocampus, and cerebellum (Figure S2B) of Dble::R6/2 mice at 4 and 9 but not at 15 wk of age. This shift in the ratio between soluble and aggregated exon 1 HTT levels can be visualised on the western blots in Figure 2D. We performed Seprion ELISA to determine whether similar results could be obtained in the HdhQ150 knock-in mice. A significant reduction in the aggregate load was observed in the striatum, cortex and cerebellum of Dble::HdhQ150 mice at 6 and 10 mo of age (Figure 2C). Taken together, these data show that HDAC4 knock-down significantly reduced aggregate load and increased levels of soluble mutant HTT in HD mouse models, reflecting a delay in the aggregation process.
We used Hdac4KO P3 brain tissue to confirm that the commercially available antibodies Sigma (DM-15), Santa Cruz (H-92), and Cell Signalling (CS2072) detected an HDAC4 specific signal (Figure S2C and unpublished data). On western blotting of nuclear and cytoplasmic fractions of mouse brain, we found that only trace amounts of HDAC4 could be detected in the nuclear fraction (Figure 2E). Therefore, to investigate whether the reduction in aggregation occurred in the cytoplasm, as would be consistent with the presence of HDAC4, we perfomed cellular fractionation on 4 wk and 9 wk brain tissue. We then resolved detergent insoluble high-molecular weight aggregates from the nuclear and cytoplasmic lysates by agarose gel electrophoresis (AGERA), prepared western blots, and performed immunodetection with the S830 antibody (epitope: exon 1 HTT). We found that HDAC4 knock-down reduced the aggregate load in the cytoplasm but not in the nucleus of Dble::R6/2 mice at both 4 (Figure S2D) and 9 (Figure 2F) wk of age. Consistent with this, we found that the cytoplasmic steady-state levels of HDAC4 were significantly reduced in Dble::R6/2 mice as compared to R6/2 at both 4 (Figure S2E) and 9 (Figure 2G) wk of age. The purity of the cellular fractions was validated by immunoblotting with α-tubulin and histone H3 antibodies (Figures 2F and S2D).
The R6/2 colony was maintained by backrossing to (CBA/Ca×C57BL/6J)F1 mice and the Hdac4 knock-out mice had been bred to the same F1 background for more than six generations. We know from having bred R6/2 mice for 99 generations and from multiple experiments that the differential segregation of CBA/Ca and C57BL/6J alleles has no effect on HD-related phenotypes in R6/2 mice [10],[27]–[33]. However, the Hdac4 null allele had been created on a 129S mouse strain background, and it is inevitable that even after backcrossing to (CBA/Ca×C57BL/6J)F1 mice for multiple generations, the Hdac4 null allele would be retained in a genomic region of 129S DNA that had not been removed by recombination. Therefore, it was possible that the observed effects might be due to genetic variation in the 129S Hdac4-linked haplotype rather than through a reduction in HDAC4. To rule out this scenario, we identified an Hdac4-linked single nucleotide polymorphism (SNP) that was polymorphic between 129S and both the C57BL/6J and CBA/Ca strain backgrounds. We crossed R6/2 mice to (129S×C57BL/6J)F1s and identified R6/2 progeny that either did not carry or were heterozygous for the 129S SNP (n = 7/genotype). The heterozygous mice contained the 129S haplotype with a wild-type Hdac4 allele. Seprion ELISA on 9-wk-old cortex showed that the 129S Hdac4-linked haplotype did not modify aggregate load in R6/2 mice (Figure S2F), confirming the role for HDAC4 in aggregate reduction.
To further understand the nature of the reduction in aggregate load by HDAC4 knock-down, we reasoned that HDAC4 might associate with HTT. Hence, we employed an in vitro GST pull-down assay and found that HDAC4 interacted specifically with exon 1 HTT containing a 53 polyQ tract but not with its 20 polyQ counterpart (Figure 3A). To determine whether HDAC4 associates with endogenous HTT, we immunoprecipitated HTT (2B7, epitope 1–17) or HDAC4 (DM-15) from brain lysates of 8-wk-old WT (7Q), HdhQ150 heterozygous, and HdhQ150 homozygous mice and immunoblotted with the MW1 (soluble mutant HTT), MAB2166 (soluble wild type and mutant HTT), and H-92 (HDAC4) antibodies. We found that mutant but not wild-type HTT co-immunoprecipitated with HDAC4 (Figure 3B). To investigate the effect of polyQ length on this interaction, we repeated the experiment with lysates from 8-wk-old heterozygous knock-in mice carrying polyQ tracts of 20 (HdhQ20) or 80 (HdhQ80). We found that HDAC4 could immunoprecipitate HTT containing 80 but not 20 glutamines (Figure 3C), consistent with the in vitro pull-down data. The sequence of HDAC4 is very similar to the class IIa member HDAC5. Therefore, to investigate the specificity of these interactions, we repeated the in vitro and in vivo immunprecipitations with an antibody specific to HDAC5. Although there was a weak interaction between exon 1 HTT and HDAC5 by in vitro pull-down (Figure 3A), this was not specific to mutant HTT, as was the case for HDAC4, and HDAC5 failed to immunoprecipitate HTT from brain lysates (Figure 3D).
To further explore this association between HDAC4 and mutant HTT, we performed immunohistochemistry to determine whether HDAC4 co-localized with nuclear and/or cytoplasmic inclusions. For this purpose, we validated a number of commercially available antibodies and found CS2072 to be specific for HDAC4 by fluorescent immunolabelling as it gave no signal on HDAC4KO P3 brain sections (Figure S3A). Consistent with our western blot results, HDAC4 was localised to the cytoplasm appearing as a punctate pattern in adult brains (Figures 3E and S3B and S3C). This cytoplasmic localisation of HDAC4 is supported by its co-localisation with the synaptic markers synaptophysin and PSD95 (Figure S3C). Confocal microscopy showed that the S830 HTT antibody detects huntingtin aggregates in R6/2 and HdhQ150 brains and that HDAC4 co-localised with some but not all cytoplasmic inclusions (Figures 3E and S3B) in both cases. Taken together, our data indicate that HDAC4 associates with soluble mutant HTT in vivo and co-localizes with cytoplasmic inclusions in all brain regions studied.
Transcriptional dysregulation is a well-documented molecular characteristic of HD pathogenesis. A comparative study of the striatal transcription profiles of seven mouse models and HD post mortem brains showed that the dysregulated signature in R6/2 and HdhQ150 models was highy comparable and in both cases more closely replicated that observed in human HD tissue than that of the other mouse models [34]. HDAC inhibitors were initially pursued as a therapy for HD because of their potential for reversing these transcriptional changes. Therefore, we performed Affymetrix microarray profiling of 9- and 15-wk cortex for WT, Hdac4Het, R6/2, and Dble::R6/2 mice to assess whether HDAC4 knock-down might rescue the global transcriptional dysregulation that occurs in R6/2 mice. As expected the cortical expression profile was profoundly changed between WT and R6/2 mice by 9 wk of age and was further dysregulated at 15 wk (Figure 4A). However, comparison of the R6/2 and Dble::R6/2 profiles indicated that only a very small number of probe sets were predicted to be differentially expressed at statistically significant levels (Figure 4A). This suggested that reduction in HDAC4 had not served to rescue transcriptional dyregulation. We used Taqman qPCR to validate the predicted changes in gene expression between R6/2 and Dble::R6/2 cortex. We were only able to detect statistically different expression levels of small effect size for Secis and Casc4, and in both cases this was in the opposite direction of that predicted by the arrays (Figure 4B).
Dysregulation of Bdnf promoter transcripts is a well-characterised hallmark of disease progression in HD [35], and restoration of Bdnf levels has been shown to correlate with phenotypic improvements in HD mouse models. As Bdnf probe sets were not represented on the arrays, we used Taqman qPCR to measure the levels of multiple Bdnf promoter trancripts as well as the coding exon (Bdnf-b) in cortex at 15 wk of age. We found that HDAC4 reduction increased Bdnf-b levels in WT cortex and almost restored the R6/2 dysregulated Bdnf transcripts to WT levels in Dble::R6/2 mice (Figure 4C).
MSNs in symptomatic R6/2 mice show pronounced morphological abnormalities, including dendritic shrinkage and spine loss. Largely consistent with these anatomical changes, at a behaviorally symptomatic age, R6/2 MSNs display a marked increase in membrane resistance, depolarization of the resting membrane potential (RMP), and an increased intrinsic excitability in response to current injection [36]. These phenotypes indicate that the R6/2 MSNs are hyperexcitable relative to MSNs in WT animals. In addition, symptomatic R6/2 mice display a progressive impairment in corticostriatal connectivity [37]. In combination, this reduction of cortical input, coupled with the abnormal excitability of the MSNs within the R6/2 striatum, will have serious consequences for appropriate striatal information processing and resultant basal ganglia output. These features likely contribute and in part underlie the impaired behavioral function exhibited by these mice. We determined the extent of functional improvement in MSNs from Dble::R6/2 as compared to R6/2 mice at 7–8 and 12 wk of age. As previously published, R6/2 MSNs exhibited a higher membrane resistance than those from WT and Hdac4HET mice at both ages, which was restored to WT levels in Dble::R6/2 mice (Figure 5A,E).
While R6/2 MSNs at 7–8 wk of age were not significantly depolarized (Figure 5B), by 12 wk the RMP was depolarized by approximately 3 mV relative to WT MSNs, and this was rescued in the Dble::R6/2 mice (Figure 5F). Despite previous reports of reduced rheobasic current (minimum current injection required to elicit an action potential) in R6/2 mice [36], this was a modest phenomenon in our hands and the reduction of HDAC4 in the Dble::R6/2 mice had no effect (Figure 5C,G). Action potential amplitude was, however, significantly reduced in R6/2 compared to WT and Hdac4HET at both ages and was fully restored in the Dble::R6/2 mice (Figure 5D,H). To assess an improvement in corticostriatal transmission, glutamatergic excitatory postsynaptic currents (EPSCs) were evoked by stimulating layer V cortical afferents innervating the striatum. In both 7–8 and 12 wk age groups, R6/2 mice showed significant reduction in EPSC amplitude for any given stimulus intensity applied. There was no difference in evoked EPSCs per genotype between the two age groups, and the data were therefore pooled (n = 28–32 neurons per genotype; Figure 5I). In the Dble::R6/2 mice, a significant restoration of evoked EPSC amplitude compared to R6/2 was noted. To further delineate the locus of this improvement, a paired-pulse stimulation paradigm (20 ms interstimulus interval) was employed to specifically assess glutamate release probability. R6/2 corticostriatal synapses displayed higher paired-pulse ratios than WT synapses, and this was significant in the 12-wk dataset, indicative of impaired glutamate release in this age group. This was fully restored in the Dble::R6/2 mice (Figure 5J). In agreement with the reduction in evoked glutamate release in R6/2 corticostriatal synapses, quantal glutamate release within the striatum, which can arise from release at both thalamostriatal and corticostriatal presynaptic terminals, was severely depressed in the 7–8-wk-old mice [mean frequency of mEPSCs in MSNs from WT = 2.3±0.25 Hz (n = 12), Hdac4HET = 1.87±0.38 Hz (n = 11), R6/2 mice = 0.13±0.03 Hz (n = 11)]. The Dble::R6/2 mice showed a profound rescue in this phenotype, with a mean frequency of mEPSCs of 1.1±0.14 Hz (n = 12) (significance against R6/2 frequency assessed by Kolmogirov–Smirnoff analysis p<0.0001; Figure 5K,L). There was no change in the amplitude of mEPSCs in R6/2 compared to WT mice, suggesting that postsynaptic 2-amino-3-(5-methyl-3-oxo-1,2-oxazole-4-yl) propanoic acid (AMPA) receptor function was not impaired in the HD model, and did not contribute to the impairment in glutamatergic transmission. mEPSC amplitude was unchanged in the Hdac4HET or Dble:R6/2 animals relative to WT or R6/2 mice (Figure 5M).
In conclusion, the combined restoration of the membrane properties of MSNs from the Dble:R6/2 mice alongside the improvement in glutamatergic cortical input to the striatum would be expected to significantly improve striatal information processing, normalize aberrant basal ganglia output, and result in an improvement in behavioral function.
In order to evaluate whether the molecular and electrophysiological changes that had been detected in the Dble::R6/2 mice might ameliorate HD-related behavioural and physiological phenotypes, we employed a set of quantitative, well-characterised tests. Mice were well matched for CAG repeat length (Table S1), and phenotypic parameters were measured from 4 to 15 wk of age in WT, Hdac4HET, R6/2, and Dble::R6/2 mice. All analyses were performed blind to genotype, and in all cases, the progression of R6/2 phenotypes was consistent with previous reports.
Rotarod performance is a sensitive indicator of balance and motor coordination that has been reliably shown to decline in R6/2 mice. In line with previous results, R6/2 rotarod performance was impaired by 8 wk (p<0.001) and deteriorated further with age (Figure 6A). HDAC4 knock-down had no impact on the performance of WT mice. However, it significantly improved R6/2 rotarod performance at all ages and delayed the progression of rotarod impairment by 1 mo to the extent that 12-wk-old Dble::R6/2 mice performed as well as 8-wk-old R6/2 mice (Figure 6A).
At 14 wk of age, close to end stage disease, the deterioration in the appearance of the Dble::R6/2 mice was considerably less marked than that of R6/2. To document this, we performed a modified SHIRPA analysis [38] and measured 16 phenotypic parameters (Table S2), for which nine gave a positive score in R6/2 as compared to WT mice. In all cases, the appearance of these phenotyes was improved in the Dble::R6/2 mice when compared to R6/2 (Figure 6B). In particular, piloerection and tremor could not be detected in Dble::R6/2 mice, and hunched back and body tone were vastly improved (Figure 6B). The phenotypic improvements are evident in Videos S1 and S2, and in the light of these, it was surprising that we were unabe to detect an any marked decrease in body weight loss (Figure 6D). The dramatic improvement in the appearance of the mice led us to assess whether HDAC4 knock-down had pro-survival effects, and we found that it extended the lifespan of R6/2 mice by approximately 20%, p = 0.0004 (log-rank test) (Figure 6C).
Brain weight was measured for cohorts of mice that were culled at 8, 12, and 15 wk of age, and as previously described, there was a progressive decrease in the weight of R6/2 brains as compared to WT. HDAC4 knock-down resulted in a very slight but statistically significant increase in brain weight at 9 and 12 but not at 15 wk of age (Figure 6E).
There is mounting support for the use of HDAC inhibitors in the treatment of a wide range of brain disorders, predominantly aimed at pathological alterations in the brain transcriptome. HDAC4 is a transcriptional repressor that is normally retained in the cytoplasm, but localises to the nucleus upon de-phosphorylation [39],[40]. In line with previous reports [41], we found that HDAC4 was located in the cytoplasm in mouse brain and showed that it did not relocate to the nucleus during disease progression. We demonstrated that HDAC4 associates with mutant HTT in a polyQ-dependent manner in vivo, consistent with an association that occurs between the polyQ stretch in HTT and the Q-rich domain of HDAC4. In line with these observations, we found that HDAC4 co-localised with cytoplasmic inclusions in both R6/2 and knock-in HD mouse models. The genetic knock-down of HDAC4 led to a significant delay in the cytoplasmic aggregation process and led to a significant restoration of synaptic function. At a physiological level, knock-down of HDAC4 extended lifespan and partially restored motor coordination and other neurological phenotypes. This suggests that a cytoplasm-based pathophysiological mechanism contributes to key aspects of neurodegenerative phenotype observed in HD.
A cytoplasmic mechanism of action for HDAC4-mediated beneficial effects was unexpected. In general, HDACs have been pursued as therapeutic targets because of their impact on epigenetics and transcription. HDAC4 is known to repress MEF2 in muscle [40], and it has been proposed that HDAC4 binds to HDAC3 to activate its deacetylase domain [24]. However, we found no impact on global transcriptional dysregulation upon HDAC4 knock-down in R6/2 mice, consistent with a predominantly cytoplasmic localisation of HDAC4. Surprisingly, the absence of HDAC4 in knock-out postnatal brain tissue had little effect on the brain transcriptome [42]. In fact, it has been shown that HDAC4 does not function as a lysine-deacetylase [43], and consistent with this, we found that HDAC4 knock-out had no effect on global acetylation in brain in vivo [42]. Our demonstration that the dysregulation of Bdnf promoter transcripts was alleviated in the double mutant mice is consistent with a cytoplasmic-based mechanism of action. Bdnf expression is repressed by RE1 silencing transcription factor (REST), which is retained in the cytoplasm in a complex containing HTT and which translocates to the nucleus in response to HD pathology [44]. The cytoplasmic aggregation process initiates prior to a reduction in Bdnf transcripts in R6/2 and other HD mouse models, and therefore it is not surprising that a delay in this cytoplasmic pathology might in turn result in a delay in Bdnf dysregulation.
Our demonstration that a delay in the cytoplasmic aggregation process has beneficial consequences is supported by the published correlation between the appearance of neuropil aggregates and disease progression [45] and is consistent with previous predictions [46]. In the human HD post mortem brain, neuropil aggregates are much more common than nuclear inclusions, present in a large numbers, and potentially associated with onset of clinical symptoms [47]. Our data suggest that HDAC4 might modulate HTT aggregation through a direct interaction serving to template or nucleate soluble HTT. Alternatively, given that HDAC4 may self-aggregate, it could influence HTT aggregation indirectly through perturbation of proteostasis networks [48]. The delay in HTT aggregation afforded by a reduction in HDAC4 was most pronounced in presymptomatic and early-stage disease and diminished with disease progression, presumably reflecting changes in aggregation kinetics. This association between phenotypic improvements and a shift from aggregated to soluble mutant exon 1 HTT is consistent with our previously published in vivo studies [49],[50]. Our data provide no information as to the aggregate species that is toxic, or as to whether all species of aggregates have detrimental consequences, but only indicate that shifting the equilibrium toward soluble HTT is beneficial in vivo.
The improvement in synaptic function as a consequence of HDAC4 knock-down was not related to a restoration in the expression level of dysregulated synaptic transcripts. Instead, it may act through reducing cytoplasmic aggregation as neuropil aggregates have been shown to inhibit axonal transport, synaptic function, and glutamate release in HD fly models [51]. The restoration of corticostriatal synaptic function demonstrated that the reduction of HDAC4 has functional consequences in the brain. However, in this study, HDAC4 was ubiquitously knocked down, and as HD has a peripheral component to its pathophysiology [52], it is conceivable that the reduction of HDAC4 also had beneficial consequences in tissues other than the brain. Given that HDAC4 has well-established functions in muscle, that muscle atrophy is a major symptom of HD, and that HDAC4 has been linked to disease progression in an ALS mouse model [53], we are currently investigating the extent to which HDAC4 reduction in muscle might contribute to the improved HD phenotypes.
The administration of HDAC inhibitors has been shown to improve disease phenotypes in a wide range of HD models [5]. To better understand which HDACs, when inhibited, are most responsible for these beneficial consequences, we embarked on a series of genetic manipulations in HD mouse models. In this article we show that the genetic knock-down of HDAC4 delayed cytoplasmic aggregation, improved synaptic function, and improved disease phenotypes. In contrast, the genetic knock-down of HDAC3 [27] and the class IIa members HDAC7 [10], HDAC5, and HDAC9 (our unpublished data) had no effect on R6/2 phenotypes. Strikingly, we recently showed that administration of SAHA caused a reduction in HDAC2 and HDAC4 at the protein but not RNA level in some R6/2 brain regions and that this correlated with a reduction in aggregation and a restoration of cortical Bdnf transcripts [11]. Therefore, we speculate that the beneficial effects of SAHA were at least in part transmitted through the down-regulation of HDAC4 via a mechanism not related to its enzyme activity.
Perhaps the best validated therapeutic target for HD is the HTT protein, and the reduction of HTT through gene silencing is being developed using both antisense oligonucleotides and RNAi [54]–[56]. However, delivery to the brain is a major challenge for these approaches, and as HTT has many essential functions, the potential liability of decreasing HTT to a detrimental level cannot be ignored. We have shown that reduction of HDAC4 shifts the ratio from aggregated to soluble HTT and therefore acts directly on the HD mutation. Our demonstration that HDAC4 levels can be decreased through the administration of a small brain-penetrant molecule (SAHA) is extremely promising as more selective inhibitors (e.g., specific to HDAC4 or class IIa enzymes) may have similar effects, making it possible to target HTT aggregation with a small molecule. Finally, these findings may have wider implications as HDAC4 is a component of Lewy Bodies in Parkinson's disease brains [57] and administration of SAHA improved the synaptic plasticity and learning behaviour in an Alzheimer disease model [58].
All animal work was approved by the local ethics committees and was performed in accordance with UK Home Office regulations or the Swiss Law (Kantonales Veterinäramt Basel-Stadt).
Hemizygous R6/2 mice were bred by backcrossing R6/2 males to (CBA×C57BL/6)F1 females (B6CBAF1/OlaHsd, Harlan Olac, Bicester, UK). Similarly, the Hdac4 knock-out colony [21] was maintained by backcrossing heterozygous males to B6CBAF1/OlaHsd females. HdhQ150 homozygous mice on a (CBA×C57BL/6)F1 background were obtained by intercrossing HdhQ150 heterozygous CBA/Ca and C57BL/6J congenic lines as described previously [19]. The HdhQ20 and HdhQ80 mice were from CHDI colonies at The Jackson Laboratory (Bar Harbor, ME) and maintained on a C57BL/6/J background. 129S2/SvHsd females were from Harlan Olac. The cross between HdhQ150 and Hdac4HET mice, both on a C57BL/6 background, was performed at Novartis. All animals had unlimited access to food and water, were subject to a 12-h light/dark cycle, and housing conditions and environmental enrichment were as previously described [59]. The R6/2 colony was kept on breeding chow (Special Diet Services, Witham, UK).
Genomic DNA was isolated from an ear-punch. R6/2 and HdhQ150 mice were genotyped by PCR, and the CAG repeat length was measured as previously described [25]. PCR conditions for genotyping Hdac4 knock-out mice were for WT band: Fw: CTTGTTGAGAACAAACTCCTGCAGCT, Rw: AGCCCTACACTAGTGTGTGTTACACA; for Hdac4 mutant band: Fw: AGCCCTACACTAGTGTGTGTTACACA, Neo Rw: CCATGGATCCTGAGACTGGGG.
Cycling conditions were 4 min at 95°C, 35× (30 s at 95°C; 45 s at 60°C; 2 min at 72°C), 10 min at 72°C using Taq polymerase and buffer from Promega. The HdhQ20, HdhQ80 mice [60] were genotyped as described [61] using the Hotstart polymerase (Thermoscientific).
Dissected tissues were snap frozen in liquid nitrogen and stored at −80°C until further analysis.
All behavioural tests were performed as previously described, and the data were analysed by repeated measures general linear model ANOVA with the Greenhouse Geisser post hoc test using SPSS software [59]. Pair-wise statistical comparisons were corrected for multiple comparisons using Bonferroni post hoc test in SPSS. Survival was assessed blind to genotype, and mice were euthanized when they reached end-stage disease. The data are presented as Kaplan–Meier cumulative survival functions and statistically analysed by the log-rank test.
Total RNA was extracted with the mini-RNA kit accordingly to manufacturer instructions (Qiagen). Reverse transcription (RT) was performed using MMLV superscript reverse transcriptase (Invitrogen) and random hexamers (Operon), and all Taqman-qPCR reactions were performed using the Chromo4 Real-Time PCR Detector (BioRad) as described [35]. Expression level of the gene-of-interest was normalised to the geometric mean of three endogenous housekeeping genes (Primer Design) as described [35]. The primer and probe sequences are detailed in Table S3.
For the Affymetrix arrays, biotinylated cRNAs were prepared from 200 ng total RNA using the GeneChip 3′ IVT Express Kit (Affymetrix) following the manufacturer's instructions. cRNA (15 µg) was hybridized to GeneChip Mouse Genome 430 version 2.0 Arrays (Affymetrix) and processed, stained, and scanned according to the manufacturer's recommendations. The quality of input RNAs and cRNAs was verified with the Bioanalyzer 2100 (Agilent Technologies) before use. Microarray quality control was performed using the software package provided on RACE [62]. Chips with a median normalized unscaled standard error greater than 1.05 were excluded. Affymetrix annotations (version 3.0) were used for probeset-to-gene assignments. Two-tailed t test was performed to assess the differences in gene expression between groups for each genotype (WT n = 8; R6/2 n = 9; Hdac4HET n = 8; Dble n = 9). Corrections for multiple testing were performed using the false discovery rate (FDR) according to Benjamini and Hochberg [63] with a significance threshold of p<0.05. The array datasets can be found at NCBI GEO accession number GSE38237.
All primary and secondary antibodies used in this study are presented in Table S4. Preparation of protein lysates and western blotting were as described previously [11]. In general, 20 µg protein lysate was fractionated on 10% SDS-PAGE gels. Aggregates were captured in Seprion ligand-coated plates (Microsens) and detected using the S830 sheep polyclonal or MW8 mouse monoclonal antibodies as described [25].
Sample preparation and the TR-FRET assay were performed as previously described [26].
Nuclear and cytoplasmic fractions were prepared as previously described [64], and their purity was determined by immunoblotting with antibodies to anti-histone H3 and α-tubulin. The AGERA assay was performed as described [65]. In general 100 µg of nuclear or cytoplasmic fractions, isolated from whole snap frozen brains, were loaded in nonreducing Laemmli buffer onto 1.5% agarose gels supplemented with 0.1% SDS and run at 3 V/cm followed by western blotting and immunodetection with anti-HTT (S830) antibodies. The high molecular weight protein marker was from Invitrogen.
Co-immunoprecipitation was performed as previously described [66]. Briefly, protein lysates were prepared from whole brains in HEPES buffer and incubated with protein-G Dynabeads (Invitrogen) overnight at 4°C on a rotating platform.
The full-length mouse Hdac4 gene in pCMV6 was obtained from Origene. This was amplified by PCR using high fidelity polymerase (Roche) and subsequently cloned into pCR2-Topo (Invitrogen) accordingly to the manufacturer's instructions. The detailed protocol used for GST pull-downs is available in the Text S1.
For immunohistochemical studies, brains were frozen in isopentane at −50°C and stored at −80°C until further analysis. We cut 10–15 µm sections using a cryostat (Bright instruments), air dried and immersed them in 4% PFA in PBS for 15 min, and washed them for 3×5 min in 0.1% PBS-Triton X-100. Blocking was achieved by incubation with 5% BSA-C (Aurion) in 0.1% PBS-Triton X-100 for at least 30 min at RT. Immunolabelling with primary antibodies in 0.1% PBS-Triton X-100, 1% BSA-C (HDAC4, S830) was completed by overnight incubation in a humidity box at 4°C. Sections were washed 3× in PBS, incubated for 60 min at RT in a dark box with the appropriate combinations of secondary antibodies diluted in PBS, washed 3× in PBS, and counterstained with DAPI (Invitrogen). Sections were mounted in Vectashield mounting medium (Vector Laboratories). Sections were examined using the Leica TCS SP4 laser scanning confocal microscope and analysed with Leica Application Suite (LAS) v5 (Leica Microsystems, Heidelberg, Germany).
Detailed procedures for acute striatal slice preparation, patch-clamp recordings, and the isolation of miniature excitatory postsynaptic currents (mEPSCs), along with the appropriate statistical analysis, can be found in Text S1.
All data were analysed with Microsoft Office Excel and two-tailed Student's t test or as otherwise stated.
|
10.1371/journal.pgen.1002826 | Reversal of PCNA Ubiquitylation by Ubp10 in Saccharomyces cerevisiae | Regulation of PCNA ubiquitylation plays a key role in the tolerance to DNA damage in eukaryotes. Although the evolutionary conserved mechanism of PCNA ubiquitylation is well understood, the deubiquitylation of ubPCNA remains poorly characterized. Here, we show that the histone H2BK123 ubiquitin protease Ubp10 also deubiquitylates ubPCNA in Saccharomyces cerevisiae. Our results sustain that Ubp10-dependent deubiquitylation of the sliding clamp PCNA normally takes place during S phase, likely in response to the simple presence of ubPCNA. In agreement with this, we show that Ubp10 forms a complex with PCNA in vivo. Interestingly, we also show that deletion of UBP10 alters in different ways the interaction of PCNA with DNA polymerase ζ–associated protein Rev1 and with accessory subunit Rev7. While deletion of UBP10 enhances PCNA–Rev1 interaction, it decreases significantly Rev7 binding to the sliding clamp. Finally, we report that Ubp10 counteracts Rad18 E3-ubiquitin ligase activity on PCNA at lysine 164 in such a manner that deregulation of Ubp10 expression causes tolerance impairment and MMS hypersensitivity.
| DNA damage is a major source of genome instability and cancer. A universal mechanism of DNA damage tolerance is based on translesion synthesis (TLS) by specialized low-fidelity DNA polymerases capable of replicating over DNA lesions during replication. Translesion synthesis requires the switch between replicative and TLS DNA polymerases, and this switching is controlled through the ubiquitylation of the proliferating-cell nuclear antigen (PCNA), a processivity factor for DNA synthesis. It is thought that DNA polymerase switching is a reversible process that has a favorable outcome for cells in the prevention of irreversible DNA replication forks collapse. However, the low-fidelity nature of TLS polymerases has unfavorable consequences like the increased risk of mutations opposite to DNA lesions. Here we identify Ubp10 as an enzyme controlling PCNA deubiquitylation in the model yeast S. cerevisiae. The identification of Ubp10 is a first step that will allow us to understand its biological significance and its potential role as part of a safeguard mechanism limiting the residence time of TLS DNA polymerases on replicating chromatin in eukaryotes.
| In living cells, tolerance mechanisms ensure that DNA can be replicated when it is damaged. These mechanisms prevent irreversible DNA replication fork collapse when the replisome encounters bulky lesions at damaged sites that block progression of replicative DNA polymerases [1], [2]. DNA lesions are bypassed either by a mechanism involving low stringency DNA polymerases called translesion synthesis (TLS) polymerases or by promoting template-switching between nascent chains within the same replication fork [2]–[6]. It is thought that both mechanisms efficiently prevent replisome stalling at damaged sites. The use of TLS polymerases may be mutagenic because they induce an error-prone process that causes damaged-dependent mutations. However, it has been shown that in yeast ultraviolet-radiation-induced DNA lesions are predominantly bypassed via translesion synthesis [7].
Eukaryotes ubiquitylate proliferating-cell nuclear antigen (PCNA) to signal damaged DNA and regulate the choice of alternative pathways to bypass DNA lesions during S-phase, therefore, to tolerate DNA damage [2]–[6]. The sliding clamp PCNA is monoubiquitylated at Lys164 by the Rad6-Rad18 (E2–E3) ubiquitin ligase complex in response to endogenous or exogenous damage causing disruptive covalent modifications of DNA interfering with high-fidelity replicative polymerases during S phase. Mono-ubiquitylated PCNA (ubPCNA) enhances the affinity of error-prone TLS DNA polymerases which facilitate translesion synthesis bypass. Then, the Mms2-Ubc13-Rad5 ubiquitin ligase complex may further ubiquitylate Lys164-mono-ubiquitylated PCNA to promote template switching, the error-free component of the bypass that involves sister-strand pairing [5], [8] and references therein). This regulatory mechanism based on covalent modifications of the Lys164 of the sliding clamp PCNA is a solidly established model conserved in all eukaryotes [2]–[5], [9].
Ubiquitylation of Lys164-PCNA (ubPCNA) greatly enhances binding of the sliding clamp with TLS polymerases [10]. In contrast with replicative enzymes, TLS polymerases are low fidelity DNA polymerases, non-processive enzymes that lack any proofreading activity but capable of replicating over DNA lesions [11] (and references there in). Indeed, TLS polymerases are DNA damage-tolerant enzymes but also mutagenic because they may incorporate mispaired deoxynucleotides opposite to lesions (damaged template) in an error-prone process [12], [13] (and references there in). Because of their low fidelity and low processivity when incorporating deoxynucleotides across from damaged and undamaged base pairs [12], [14]–[17], cells need to keep TLS DNA polymerases from sampling replicative DNA more that strictly required and/or to prevent them from extended interaction with replication forks. Therefore, cells may need a control mechanism to deubiquitylate ubPCNA as soon as TLS DNA polymerases have been able to replicate over the damaged site.
Human Usp1 has been identified as a protease that deubiquitylates mono-ubPCNA [18]. Upon UV-light induced DNA damage, Usp1 is degraded so that PCNA becomes ubiquitylated [18], [19], suggesting that Usp1 deubiquitylates PCNA continuously in the absence of DNA damage [18]. However, accumulation of ubPCNA does not correlate with Usp1 proteolysis when the progression of replication forks is stalled with HU [20], suggesting either a complex regulation of Usp1 activity (or its subcellular localization) when cells are exposed to other DNA damaging agents or the existence of at least one another PCNA deubiquitylating enzyme in mammals acting in response to other DNA damaging agents. Despite the identification of Usp1, little is known about the deubiquitylation of ubPCNA in any other organism.
In Saccharomyces cerevisiae, the protease (or proteases) that deubiquitylates ubPCNA remains unknown. Potential candidates in budding yeast are 17 genes that codify for different ubiquitin-specific proteases. Few of them have been extensively studied while others remain poorly characterized [21]–[23]. These genes are named UBPs (from UBP1 to UBP17), where UBP stands for ubiquitin protease. Among the ubiquitin-specifc proteases characterized, Ubp10/Dot4 is remarkable; this is a deubiquitylating enzyme related to gene-silencing that regulates histone ubH2B deubiquitylation and helps to localise the histone deacetylase Sir2 complex at telomeres, cryptic mating type loci (HML and HMR) and rDNA loci [24], [25]. Here we describe a new role for Ubp10 in deubiquitylating the sliding clamp ubPCNA. We performed a biochemical screening with yeast UBPs single mutants to identify ubiquitin proteases that might play a role in the reversal of PCNA ubiquitylation and found that UBP10 mutants accumulate ubiquitylated forms of PCNA. Consistent with a direct role in ubPCNA deubiquitylation, we found that catalyticaly active Ubp10 reverts PCNA ubiquitylation.
In yeast, the ubiquitylation of PCNA might be a reversible process catalyzed by deubiquitylating enzymes (or DUBs). Sequence and functional analyses have revealed that in budding yeast there are 17 genes (from UBP1 to UBP17) encoding different ubiquitin-specific processing proteases and thus potential candidates to deubiquitylate PCNA. To identify ubiquitin proteases that might play a role in the reversal of PCNA ubiquitylation, we examined PCNA ubiquitylation patterns of Saccharomyces cerevisiae strains lacking individual ubiquitin proteases. To detect modified forms of this sliding clamp we used a polyclonal rabbit antibody that specifically detects PCNA in S.cerevisiae cell extracts (Figure 1A). As shown in Figure 1B, ubp10Δ mutant cells accumulated di-ubiquitylated PCNA forms, a phenotype consistent with defects in deubiquitylation of this sliding clamp. This phenotype (the accumulation of ubiquitylated PCNA) was also observed in cells expressing a version of Ubp10 that lacks catalytic activity (ubp10C371S) (see later), a catalytic inactive form previously described [25]. We also found that the ubiquitylated PCNA forms accumulated in ubp10Δ mutant cells were covalent modifications on Lysine 164 of the sliding clamp (Figure 1C and Figure S1).
Ubp10 and Ubp8 are the ubiquitin proteases that remove monoubiquitin from histone H2B [24], [25]. Although these H2B-deubiquitylating enzymes have distintc functions [26], deletion of both UBP8 and UBP10 results in a synergistic increase in H2B ubiquitylation levels suggesting that they regulate the global balance of that histone modification [24], [25]. Thus, even though we detected normal levels of PCNA modifications in ubp8Δ mutant cells, we tested whether or not deletion of UBP8 in a ubp10Δ mutant further increased PCNA ubiquitylation levels. We found that the accumulation of ubPCNA was specific to ubp10Δ (Figure S2).
It has been shown that the ubiquitylation of PCNA is restricted to, although separable from, S-phase [7], [27], [28]. Under physiological circumstances active DNA replication forks are required for PCNA ubiquitylation [27]. In fact, PCNA ubiquitylation is induced by chemicals that cause disruptive covalent modifications of DNA, blocking replication and that involve the accumulation of single-stranded DNA. Thus, in S. cerevisiae, PCNA is ubiquitylated during S-phase in response to the detection of DNA lesions caused by methyl methane sulfonate (MMS), hydroxyurea (HU), 4-nitroquinoline 1-oxide (4-NQO), UV light, hydrogen peroxide (H2O2) and ionizing radiation [27]. We therefore wondered whether ubp10Δ mutants accumulate more ubiquitylated PCNA than wild-type cells in response to all these types of inducers. As shown for MMS, HU, 4-NQO and UV light (Figure 1D and 1E), we found that ubp10Δ mutant cells accumulated increased levels of ubiquitylated PCNA as compared to control wild-type cells. This observation indicates that in vivo Ubp10 modulates the level of DNA damaged-induced PCNA ubiquitylation.
The increased levels of PCNA ubiquitylation observed in UBP10 mutant cells suggested that Ubp10 could be a potential candidate for the deubiquitylation of PCNA in vivo. We therefore analyzed the ability of Ubp10 to counteract MMS-induced ubiquitylation of PCNA when overproduced. We examined PCNA ubiquitylation in strains in which expression of UBP10 was regulated by the strong galactose-inducible GAL1,10 promoter. Exponentially growing cultures were treated with MMS. Then, the expression of UBP10 was either induced or repressed by adding galactose or glucose, respectively. Samples were taken at regular intervals and processed for Western analysis of PCNA ubiquitylation (Figure 2A). Overexpression of UBP10 resulted in rapid reversion of PCNA ubiquitylation, consistent with a role as an ubiquitin-specific processing protease for PCNA. Interestingly, both mono- and di-ubiquitylated PCNA forms rapidly disappeared in cells overexpressing UBP10, suggesting that Ubp10 also deubiquitylates di-ubPCNA forms. These deubiquitylation events depended on the protease activity of Ubp10 as a catalytically inactive Ubp10C371S mutant form was unable to deubiquitylate PCNA in vivo in similar conditions (Figure 2B). We have also observed that Ubp10 overproduction reverts ubiquitylation of PCNA induced by treatments with HU, 4-NQO and UV radiation. In summary, these experiments indicate that overexpression of catalytically active Ubp10 can deubiquitylate ubPCNA in vivo. Importantly, this in vivo reaction did not require any other UBP gene, as active Ubp10 did deubiquitylate ubPCNA in any single UBP1-17 deletion (Figure S3).
Yeast PCNA mutants lacking the ubiquitin/SUMO-conjugation site K164 or mutated in the PCNAK164-E3 ubiquitin ligase Rad18 are hypersensitive to MMS (and other DNA damaging agents) because the ubiquitylation of this K164 amino acid residue is critical to tolerate DNA damage [29]. It is then reasonable to predict that the overexpression of the K164-ubPCNA ubiquitin-specific protease will counteract Rad18 activity and induce MMS hypersensitivity. Therefore, we exposed UBP10-overexpressing cells to the chronic presence of the alkylating chemical and found, as predicted, that high levels of expression of the catalytically active form of this ubiquitin protease (but not the inactive Ubp10C371S form) induced hypersensitivity to MMS (Figure 2C). Significantly, this effect was specifically related to high levels of expression of Ubp10 because overexpression of any other UBP gene neither sensitize cells to MMS nor induce PCNA deubiquitylation in vivo (Figure S4). Regarding UBP10 overexpression, two additional and testable predictions can be made, first, the hypersensitivity to MMS should depend on the PCNA lysine 164 modification. To test this prediction we used a simple epistasis analysis to determine the order of function of the POL30 and GAL1,10: UBP10 (Figure S5). We have indeed found that POL30 is epistatic to GAL1,10: UBP10 indicating that the MMS-sensitivity of Ubp10 overproduction depends on the PCNA lysine 164 modification. Second, given that mono-ubiquitylation of the K164 residue of PCNA is in principle important to enhance its interaction with mutagenic TLS polymerases, it is plausible to predict that the mutagenesis frequency of cells overexpressing UBP10 should be reduced as compared to wild-type cells. We have found that this is the case (Figure S6).
The above observations correlated the enzymatic activity of Ubp10 with PCNA deubiquitylation in vivo. However, these effects may depend on deubiquitylation of histone H2B, as Ubp10 deubiquitylates K123 ubH2B [24], [25]. In order to understand whether deubiquitylation of H2B and PCNA were independent from each other, we repeated our overexpression analysis in a bre1Δ mutant background. Bre1 is the E3 ubiquitin ligase that ubiquitylates histone H2B in yeast cells, thus, deletion of the BRE1 gene impedes H2BK123 ubiquitylation [30], [31]. Importantly, BRE1 deleted cells are viable, providing a tool to answer the question. As shown in Figure S7, overproduction of catalytically active Ubp10 reverts PCNA ubiquitylation and hypersensitize cells to MMS similarly in wild-type and bre1 mutant cells. These results indicate that Ubp10-dependent PCNA deubiquitylation is functionally separable from ubiquitylation of histone H2B.
MMS modifies guanines and adenines to methyl derivatives causing DNA base mispairing, inducing DNA damage and slowing down progression of DNA replication forks during S-phase [32]–[35]. MMS also induces ubiquitylation of PCNA in all model organisms tested to date (reviewed in [4]). To further study the role of Ubp10 in the modulation of PCNA ubiquitylation in yeast, we analyzed by Western blot samples taken at regular intervals from wild-type cells treated for 60 minutes with the alkylating chemical and compare them to samples taken from UBP10 mutant cells in similar conditions (Figure S8). As observed in the Figure S8, wild-type cells ubiquitylate PCNA after the MMS treatment and then actively deubiquitylate the sliding clamp in such way that 45 minutes after the release from the drug treatment ubiquitylated PCNA was barely detectable. In contrast UBP10 deleted cells maintained steady state levels of ubiquitylated PCNA throughout the experiment, suggesting that these cells lack the appropriate enzyme involved in the deubiquitylation of the modified clamp.
Having observed that deletion and overexpression phenotypes of Ubp10 were consistent with the hypothesis that this ubiquitin-specific protease deubiquitylates PCNA in yeast, we next addressed whether Ubp10 and PCNA interact in vivo, as expected for an enzyme-substrate complex.
Addition of single ubiquitin residue to Lys164 of PCNA in yeast is controlled by the E2–E3 complex Rad6–Rad18 during S-phase [7], [28]. Accordingly, the Rad6–Rad18 enzyme complex and its substrate PCNA interact in vivo, as has been observed by yeast two-hybrid analyses [29]. We speculated that Ubp10 could form a complex with PCNA in a Rad18 dependent manner, as it has been described previously for other E3-ubiquitin ligases [36]–[38]. If this were true, it could be predicted that these interactions might be detected by co-immunoprecipitation analysis. In particular, we were interested in determining a possible in vivo PCNA-Ubp10 interaction at endogenous levels of both proteins. Since we used a C-terminally myc-tagged Ubp10 strain we carefully checked growth rate, gene expression levels, PCNA and histone H2B deubiquitylation and found no differences with untagged wild-type controls, as shown for PCNA (Figure S9). By Western and co-immunoprecipitation assays, we found that Ubp10-myc is stable upon exposure to DNA damage and that Ubp10 binds PCNA throughout the cell cycle and in response to MMS-induced DNA damaged (Figure S9). We then studied Ubp10-PCNA interaction in wild-type and rad18Δ mutant cells and observed that Ubp10 and PCNA interact in vivo in a Rad18 semi-dependent manner (Figure 3A and 3B). We next tested Ubp10 and Rad18 interaction and found that Rad18 can associate in vivo with Ubp10 both in undamaged and exogenously DNA-damaged cells (Figure S10). These results suggest that in yeast cells Ubp10, PCNA and Rad18 could form a complex. These findings, particularly those related to PCNA and Ubp10 interaction, strongly support the hypothesis that Ubp10 is an ubiquitin-specific protease that deubiquitylates PCNA in yeast cells.
In S.cerevisiae, REV1 encodes a deoxycytidyltransferase required for the bypass of abasic sites in damaged DNA. Rev1p forms a complex with the subunits of DNA polymerase ζ Rev3 and Rev7, which are involved in error-prone lesion bypass as yeast TLS DNA polymerases [14], [39]. Furthermore, it has been shown that yeast Rev1 interacts with, and its activity is stimulated by, PCNA [40], [41]. Therefore, we reasoned that the accumulation of mono-ubiquitylated PCNA observed in UBP10 mutant cells could lead to an increased interaction between PCNA and TLS DNA polymerases, including the TLS-interacting Rev1 protein. We tested this possibility by co-immunoprecipitation assays in vivo using strains carrying myc-tagged Rev1 and either wild-type or C-terminal FLAG-tagged PCNA. We detected the reported interaction between the sliding clamp PCNA and the deoxycytidyltransferase Rev1 in the wild-type strain and, importantly, it was increased in cells lacking a functional Ubp10, as predicted (Figure 4A and 4B). We also found that this increase observed in ubp10Δ mutant cells was dependent on the PCNA lysine 164 modification (Figure S11). Interestingly, we found that the sliding clamp co-immunoprecipitated Rev1 from asynchronous or MMS-damaged cell cultures (Figure 4C and Figure S12). If this enhacement (in Rev1-PCNA interaction) observed in ubp10Δ mutant cells is due to an increase in the ubiquitylation of PCNA, it would be expectable to detect ubiquitylated PCNA in undamaged cells. To our knowledge, detection of ubPCNA in undamaged budding yeast cells remains elusive. However, by immunoprecipitating the sliding clamp from POL30-FLAG tagged cells, although weakly, we detected ubiquitylated PCNA in asynchronous cultures of exponentially growing wild-type and UBP10 mutant cells and indeed found that the mutant accumulated ubiquitylated PCNA (Figure S13). This observation supports the correlation between the increase in ubPCNA and the enhancement of Rev1-PCNA interaction in undamaged cells. Finally, we did not observe PCNA-Rev1 interaction in G1 synchronized cells, even though the Rev1 protein was present in the cell extracts (Figure 4B, 4C and Figure S12).
We next analyzed chromatin-associated Rev1 foci and found that, in agreement with the co-immunoprecipitation results, ubp10Δ mutant cells had increased numbers of Rev1 foci (mean±s.d.: wild type, 16.64±8.42; ubp10Δ, 20.47±10.24). Remarkably, a detailed analysis revealed a significant increment in nuclei with high numbers of Rev1 foci in UBP10 mutant cells (Figure 4D). In theory, the observed increased interaction between PCNA and Rev1 in UBP10 deleted cells could be suggestive of a greater TLS activity on replicating chromatin that would result in increased mutagenic rate. Therefore, we next monitored the forward mutation rate to canavanine resistance [42] in undamaged or MMS-damaged ubp10Δ mutant cells. However, we found no statistically-significant differences in the mutagenic rate when compared to that of wild-type cells (Figure S14), indicating that increasing levels of PCNA ubiquitylation has no observable impact in the frequency of mutation.
The Rev1-Rev3/Rev7 complex formation has been succesfully tested in yeast [43], [44]. However, having shown that mutation of UBP10 enhances Rev1 interaction with PCNA but does not increase mutation frequency (and in order to explain this discrepancy), we wondered whether the Rev3/Rev7 (DNA polymerase ζ) interaction with PCNA was regulated in a different way than the observed for Rev1 in ubp10Δ yeast mutant cells. In order to test this hypothesis, we first analysed Rev3-PCNA interaction in wild-type and ubp10Δ cells (Figure 5A). By co-immunoprecipitation assays, we found that Rev3, the catalytic subunit of pol zeta, interacts with PCNA in wild-type and ubp10Δ mutant strains. We also observed that the amount of Rev3 co-immunoprecipitated with PCNA was similar in both strains either in asynchronous cultures or when cells were treated with MMS. We nex studied the interaction of PCNA with the accessory subunit of DNA polymerase ζ Rev7 (Figure 5B and 5C). Rev7 stimulates the activity of Rev3 [14] and is required for mutagenesis induced after DNA damage in such a manner that deletion of REV7 decreases mutagenesis frequency in yeast [45]. Significantly, in our co-immunoprecipitation assays we did observe that the interaction of PCNA with Rev7 was greatly reduced in cells deleted for UBP10 supporting an explanation for the wild-type-like mutagenesis frequency observed in them.
The evidence presented up to here indicate that the activity of Ubp10 is required for reverting PCNA ubiquitylation but does not addesss when Ubp10-mediated PCNA deubiquitylation takes place during the cell cycle. Therefore, we were next interested in understanding whether deubiquitylation of PCNA occurs during S-phase. Through the depletion of nucleotides, the drug hydroxyurea (HU), an effective ribonuclease reductase inhibitor, causes an early S-phase arrest in S.cerevisae cells [46] and induces ubiquitylation of PCNA [27], thus, providing a way to study the regulation of PCNA ubiquitylation in the presence of stalled DNA replication forks. In this scenario, we compared PCNA ubiquitylation in wild-type and ubp10Δ mutant cells (Figure 6). Cells in logarithmic growth at 30°C were synchronyzed with α-factor and then released in 0.2M HU at the same temperature and samples (taken at regular intervals) processed for Western analysis of PCNA. We used as S-phase markers PCNA SUMOylation [29], Rad53 activation [35] and Clb5 accumulation [47]–[50]. As recently described [7], [27], [28], we detected PCNA ubiquitylation as soon as cells entered S-phase, coincident with the appearance of PCNA SUMOylation, Rad53 activation (in response to HU) and Clb5 accumulation (Figure 6C). Under the chronic presence of HU, in wild-type cells PCNA ubiquitylation reached a maximum 40 minutes after the release from the pheromone arrest and then started to decline with stalled DNA replication forks as judged from all markers, including DNA content analysis by FACS. The timing of PCNA ubiquitylation observed here correlates well with the recently described timing of association of Rad18 with replicating chromatin in HU treated cells [27]. The decrease in ubPCNA observed in wild-type cells was somewhat surprising; however, it does indicate that yeast cells down-regulate the modification of the clamp during S-phase. In contrast, cells lacking Ubp10 activity, even though they progressed into S-phase later or more slowly than controls (Figure 6B and 6C), accumulated increased amounts of mono and di-ubiquitylated forms of the clamp that remained high all throughout the synchronous experiment (see bar plot for ubPCNA in Figure 6C). The analysis of ubp10Δ mutant cells is consistent with the idea that this ubiquitin-specific protease down-regulates PCNA ubiquitylation during S-phase and suggest that Ubp10 is a major deubiquitylating enzyme for ubPCNA in budding yeast cells (see model in Figure 7).
In this work we present clear evidence indicating that Ubp10 controls PCNA deubiquitylation in S. cerevisiae. Ubp10 has a well established role as an ubiquitin-specific protease of ubH2B, a role related to gene-silencing (at telomeres, rDNA and cryptic mating type loci), together with Ubp8, the SAGA-associated ubH2B deubiquitylase involved in gene expression [24], [25]. Thus, in combination Ubp8 and Ubp10 regulate the global balance of ubH2B [24], [25]. In addition to this role, here we present results supporting that Ubp10 is an important ubiquitin-specific protease also in removing ubiquitin from ubPCNA in budding yeast. Our observations that wild-type cells deubiquitylate ubPCNA in response to the alkylating chemical MMS or under the chronic presence of HU show that there exists an active control to revert PCNA ubiquitylation in S.cerevisiae yeast cells. Moreover, our experiments with ubp10C371S mutant strains indicate that such control depends on the catalytic activity of Ubp10/Dot4.
UBP10 deleted cells or cells carrying a catalytically inactive form of Ubp10 accumulate ubPCNA, a phenotype consistent with the idea that in vivo Ubp10 is the protease that removes ubiquitin from ubiquitylated PCNA. In agreement with this role, overexpression of active Ubp10 reverts PCNA ubiquitylation and hypersensitizes cells to MMS. Moreover, Ubp10 and the sliding clamp PCNA interact in vivo as expected from the formation of and enzyme-substrate complex. Importantly, the function of Ubp10 as ubPCNA ubiquitin-specific protease is separable from histone H2B ubiquitylation, as Ubp10 deubiquitylates ubPCNA in cells lacking Bre1, the E3 ubiquitin ligase that in complex with Rad6 monoubiquitylates histone H2BK123 [31], [51]. However, the ubPCNA and ubH2B deubiquitylation roles of Ubp10 might be functionally related. One interesting hypothesis is that Ubp10-dependent deubiquitylation of ubPCNA and ubH2B are inseparable functions. It is arguable that Ubp10 might modulate both replication bypass and histone modification in order lo leave the epigenetic marks unaltered during DNA replication. In fact, it has been inferred from DT40 chicken cells defective in Rev1 that this TLS-associated deoxycytidyl transferase is involved in replication of G4-structured DNA regions and, as a consequence of it, in leaving intact their histone methylation epigenetic marks [52]. Since here we report a functional link between Rev1, PCNA, Rad18 and Ubp10, it is reasonable to surmise that Ubp10 would modulate PCNA ubiquitylation and (the maintenance of) histone imprinting during replication. These modulatory roles are also consistent with the fact that the modulator (Ubp10) might form part of the complexes (PCNA, Rad6-Rad18, Rad6-Bre1) involved in both actions.
An important observation presented in this work is that Ubp10 is able to remove mono-ubiquitin as well as di-ubiquitin from PCNA in vivo, suggesting that this ubiquitin protease enzyme may be crucial for keeping TLS polymerases in check as well as for down-regulating the error-free bypass. Thus, a single deubiquitylating enzyme might downregulate both branches of the tolerance pathway to DNA damage in budding yeast.
Where does PCNA deubiquitylation take place? The answer to this simple question is not necessarily trivial, since the localization Ubp10 might be a point of interest for future analysis. Initial studies in formaldehyde-fixed cells suggested that Ubp10 localizes primarily at the nucleus [53]; however, using in vivo studies of Ubp10-GFP as well as immunofluorescence analysis of Ubp10-myc on nuclear spreads, we have found that Ubp10 localizes mainly in the rDNA-containing nucleolar region (our own unpublished observations). Thus, does Ubp10 localize permanently to the nucleolus? ChIP evidence has confirmed rDNA loci, telomeres and cryptic mating type loci localization [24], [25], [54] so that Ubp10-dependent deubiquitylation of ubH2B should take place there. Deubiquitylation of ubPCNA may follow a more dynamic pattern (as DNA replication forks move during ongoing replication). Alternatively, and more simply, an undetected fraction of Ubp10 might be permanently located out of the nucleolus or might be released from this nuclear compartment to control the deubiquitylating processes during S-phase and postreplication repair. Future studies will address these alternatives.
As in yeast cells, PCNA ubiquitylation is required for mammalian cell survival after UV irradiation, HU or MMS treatment [55]. In human cells Usp1 deubiquitylates PCNA as well as the Fanconi's anaemia protein FANCD2 [19], [56]–[58]. It has been shown that human Usp1 incessantly deubiquitylates ubPCNA in the absence of DNA damage [18]. Upon UV light-induced DNA damage, Usp1 is (auto)proteolysed, such that PCNA becomes ubiquitylated [18], [19]. Our work has uncovered several differences in the regulation of PCNA deubiquitylation between yeast and human cells. First, we observed that UBP10 deleted yeast cells accumulate ubiquitylated PCNA forms in response to MMS, HU, UV-light and 4-NQO, suggesting that a single DUB (Ubp10) may control PCNA deubiquitylation in budding yeast. Second, Ubp10 appears to deubiquitylate PCNA during S-phase (when the sliding clamp is modified). Finally, Ubp10 protein levels remained constant when cells are exposed to DNA damage. Thus, it is unlikely that a similar Usp1-like autoregulatory mechanism on yeast Ubp10 ubiquitin protease would exist.
The evidence presented here supported the hypothesis that Ubp10 deubiquitylates PCNA to limit the residence time of TLS polymerases on DNA replication forks during S-phase. We tested this hypothesis directly by studying Rev1-PCNA interaction because Rev1 serves as a scaffold for the polymerase ζ, encoded by REV3 and REV7, for efficient bypass of DNA lesions [59]–[61]. In agreement with this hypothesis, we found that deletion of UBP10 resulted in an increased interaction between PCNA and Rev1 in undamaged and DNA-damaged cells, and that, in turn, this enhanced interaction resulted in a net increase in Rev1 foci in chromatin. However, in contradiction with an increased number of Rev1 foci, we have also found that deletion of UBP10 does not increase the mutagenic frequency. A conceivable explanation for this contradiction would be that and additional level of control on TLS polymerases may exist to regulate their activity. In this context, one simple possibility is that DNA polymerase ζ interaction with replicating chromatin may be hindered in UBP10 deleted cells. Therefore, to explain the observed discrepancy we studied the interaction of DNA polymerase ζ subunits Rev3 and Rev7 with PCNA. Significantly, we have found that DNA polymerase ζ accessory subunit Rev7 requires Ubp10 to fully interact with the sliding clamp PCNA. This observation explains why ubp10Δ mutant cells have a wild-type-like mutagenic frequency and, more importantly, it opens the unexpected possibility that Rev1 and DNA polymerase ζ subunits may be regulated in quite distinct ways regarding their interaction with PCNA and, thus, with replicating chromatin. Further studies will be required to test this hypothesis and to study the potencial role of Ubp10 in modulating DNA polymerase ζ subunit Rev7 binding to the sliding clamp PCNA. In summary, our data support that Rev1 interaction with PCNA is modulated by ubiquitylation of PCNA and, thus, follows the classical regulatory model. Here, we propose that Ubp10 participates in this modulation through the deubiquitylation of ubPCNA. However, from the observations presented here we also deduced that Ubp10 may play a direct or indirect role in regulating Rev7 interaction with the sliding clamp apparently in a PCNA ubiquitylation independent manner.
It is proper to mention here that the activity TLS-DNA polymerases activity may be regulated by checkpoint kinases. For example, it has been shown in budding yeast that Rev1 is regulated during the cell cycle [62], and that it is phosphorylated by the Mec1-Ddc2 kinase in response to various types of DNA damages [63]–[65]. Thus, in response to DNA damage, yeast cells would have two different levels of control: first, in modulating the interaction of PCNA and TLS polymerases, and second, in regulating TLS polymerases activity and/or stability. A control mechanism that may be conserved as ATR-mediated phosphorylation of DNA polymerase η is involved in the proper response to UV-mediated DNA damage in human cells [66].
What might be the biological significance of Ubp10-mediated ubPCNA deubiquitylation in budding yeast? It is tempting to say that our results suggest that the biological significance of the control of PCNA deubiquitylation in S.cerevisiae is to prevent extended residence time of Rev1 in replicating chromatin. However, there is no unfavorable outcome for yeast cells deleted for UBP10 as they fail to support a full interaction of (DNA polymerase ζ subunit) Rev7 with PCNA and, consequently, they show a wild-type-like mutagenic frequency. It is true that these opposite effects on Rev1 and Rev7 suggest the hypothesis that Ubp10 has a complex role in modulating TLS subunits interaction with PCNA (and perhaps with replicating chromatin). However, additional studies will be required to test this hypothesis. Significantly, it has been reported the functionality in tolerance of a PCNA mutant form constitutively fused to mono-ubiquitin [67]. Thus, an alternative interpretation of our results is that Ubp10-driven deubiquitylation of ubPCNA may not be that important to tolerate DNA damage in yeast as deletion of UBP10 has no impact in MMS sensitivity nor leads to a mutator phenotype.
General experimental procedures of yeast Molecular and Cellular Biology were used as described previously [68]–[71].
All the budding yeast used in our studies are listed in Table S1. Yeast strains were grown in rich YPA medium (1% yeast extract, 2% peptone, 50 µg/ml adenine) containing 2% glucose. For block-and–release experiments, cells were grown in YPA with 2% glucose (except where indicated) at 25°C and synchronised with α-factor pheromone in G1 by adding 40 ng/ml (final concentration, 2.5 hours). Cells were then collected by centrifugation and released in fresh media in the absence or in the presence of MMS (or other drugs as indicated). Overexpression experiments with cells grown in YPA medium with 2% raffinose at 25°C were conducted by adding to the medium 2.5% galactose (to induce) or 2% glucose (to repress) and further incubating with/without MMS. For flow cytometry, 107 cells were collected by centrifugation, washed once with water, and fixed in 70% ethanol and processed as described previously [68], [72]. The DNA content of individual cells was measured using a Becton Dickinson FACScan. Cells were prepared for flow cytometry as described [72], [73].
Exponentially growing or stationary cells were counted and serially diluted in YPA media. Tenfold dilutions of equal numbers of cells were used. 10 µl of each dilution were spotted onto YPAD (2% glucose) or YPAGal (2.5% galactose) plates (always supplemented with 50 µg/ml adenine), YPAD or YPAGal plates containing different concentrations of MMS (Sigma), or HU (Sigma), incubated at 25°C and scanned. MMS plates were always freshly made.
Forward mutation analysis at the CAN1 locus was performed essentially as described previously [74]. Cells were grown in rich medium (YPAD or YPAGal) to log phase and MMS (at indicated concentrations) was added to the half of each culture, which were further incubated until the saturation point was reached (24 hours for wild-type, ubp10Δ and ubp10Δ rev3Δ strains in Figure S14 to 48 hours for wild-type, GAL1,10:UBP10, rev3Δ and GAL1,10:UBP10 rev3Δ strains in Figure S6). Then, cells were plated on solid medium without arginine but containing 60 µg/ml canavanine (Sigma) and also in control YPAD plates (for reference). After 4 days, colonies were counted and the mutagenesis frequency (canavanine resistant cells versus total population) was calculated for each culture. The frequencies provided are mean values of six or more independent cultures of each indicated genotype, in at least three independent experiments.
Tagged alleles were constructed using the single step PCR-based gene modification strategy [75]. A similar strategy was used to generate specific gene deletions. The selection markers used were KanMX6, which allows selection with geneticin, HphMX4, which allows selection with hygromicin or NatMX4, which allows selection with nourseothricin. We used also LEU2 and HIS3 markers (as indicated in Table S1). The resulting genomic constructions were confirmed by PCR and sequencing. In the case of tagged alleles, the presence of tagged proteins was confirmed by Western blot.
Immunofluorescence of nuclear spreads was performed essentially as described previously [71], [79]. The anti-myc tag antibody (clone 4A6, 05-724; Millipore) was used at 1∶500 dilution and the Alexa Fluor 594-conjugated anti-mouse secondary antibody (A11032; Molecular Probes) was used at 1∶200 dilution. Images were captured using a Nikon Eclipse 90i fluorescence microscope equipped with an Orca-AG (Hamamatsu) CCD camera and a PlanApo VC 100×/1.4 objective. Images were processed and analyzed with the MetaMorph software (Molecular Devices). Quantification of chromosome-associated Rev1 was performed by counting the number of Rev1-myc foci in the DAPI-stained area.
|
10.1371/journal.pmed.1002258 | Genetic assessment of age-associated Alzheimer disease risk: Development and validation of a polygenic hazard score | Identifying individuals at risk for developing Alzheimer disease (AD) is of utmost importance. Although genetic studies have identified AD-associated SNPs in APOE and other genes, genetic information has not been integrated into an epidemiological framework for risk prediction.
Using genotype data from 17,008 AD cases and 37,154 controls from the International Genomics of Alzheimer’s Project (IGAP Stage 1), we identified AD-associated SNPs (at p < 10−5). We then integrated these AD-associated SNPs into a Cox proportional hazard model using genotype data from a subset of 6,409 AD patients and 9,386 older controls from Phase 1 of the Alzheimer’s Disease Genetics Consortium (ADGC), providing a polygenic hazard score (PHS) for each participant. By combining population-based incidence rates and the genotype-derived PHS for each individual, we derived estimates of instantaneous risk for developing AD, based on genotype and age, and tested replication in multiple independent cohorts (ADGC Phase 2, National Institute on Aging Alzheimer’s Disease Center [NIA ADC], and Alzheimer’s Disease Neuroimaging Initiative [ADNI], total n = 20,680). Within the ADGC Phase 1 cohort, individuals in the highest PHS quartile developed AD at a considerably lower age and had the highest yearly AD incidence rate. Among APOE ε3/3 individuals, the PHS modified expected age of AD onset by more than 10 y between the lowest and highest deciles (hazard ratio 3.34, 95% CI 2.62–4.24, p = 1.0 × 10−22). In independent cohorts, the PHS strongly predicted empirical age of AD onset (ADGC Phase 2, r = 0.90, p = 1.1 × 10−26) and longitudinal progression from normal aging to AD (NIA ADC, Cochran–Armitage trend test, p = 1.5 × 10−10), and was associated with neuropathology (NIA ADC, Braak stage of neurofibrillary tangles, p = 3.9 × 10−6, and Consortium to Establish a Registry for Alzheimer’s Disease score for neuritic plaques, p = 6.8 × 10−6) and in vivo markers of AD neurodegeneration (ADNI, volume loss within the entorhinal cortex, p = 6.3 × 10−6, and hippocampus, p = 7.9 × 10−5). Additional prospective validation of these results in non-US, non-white, and prospective community-based cohorts is necessary before clinical use.
We have developed a PHS for quantifying individual differences in age-specific genetic risk for AD. Within the cohorts studied here, polygenic architecture plays an important role in modifying AD risk beyond APOE. With thorough validation, quantification of inherited genetic variation may prove useful for stratifying AD risk and as an enrichment strategy in therapeutic trials.
| Across the United States, late-onset Alzheimer’s disease (AD) is the most common form of dementia.
There is a strong need for in vivo markers for AD risk stratification and cohort enrichment in therapeutic trials.
Although numerous studies have identified several genetic risk factors, including the ε4 allele of apolipoprotein E (APOE), genetic variants have not been integrated with genetic epidemiology for quantifying age of AD onset.
Using genotype data from over 70,000 AD patients and normal elderly controls, we evaluated the feasibility of combining AD-associated SNPs and APOE status into a continuous measure—a polygenic hazard score (PHS)—for predicting the age-specific risk for developing AD.
Using a survival model framework, we integrated single nucleotide polymorphisms associated with increased risk for AD into a PHS for each participant. By combining population-based incidence rates and the genotype-derived PHS for each individual, we derived estimates of instantaneous risk for developing AD, based on genotype and age, and tested replication in two independent cohorts.
Individuals in the highest PHS quartile developed AD at a considerably lower age and had the highest yearly AD incidence rate.
In independent cohorts, we found that the PHS strongly predicted empirical age of AD onset and longitudinal progression from normal aging to AD, and associated strongly with neuropathology and in vivo markers of AD neurodegeneration.
Additional prospective validation of these results on non-US, non-white, and prospective community-based cohorts is necessary before clinical use.
Genetic variants can be integrated within an epidemiology framework to derive a polygenic score that can quantify individual differences in age-specific genetic risk for AD, beyond APOE.
Quantification of inherited genetic variation may prove useful for AD risk stratification and for therapeutic trials.
| Late-onset Alzheimer disease (AD), the most common form of dementia, places a large emotional and economic burden on patients and society. With increasing health care expenditures among cognitively impaired elderly individuals [1], identifying individuals at risk for developing AD is of utmost importance for potential preventative and therapeutic strategies. Inheritance of the ε4 allele of apolipoprotein E (APOE) on Chromosome 19q13 is the most significant risk factor for developing late-onset AD [2]. APOE ε4 has a dose-dependent effect on age of onset, increases AD risk 3-fold in heterozygotes and 15-fold in homozygotes, and is implicated in 20%–25% of AD cases [3].
In addition to the single nucleotide polymorphism (SNP) in APOE, recent genome-wide association studies (GWASs) have identified numerous AD-associated SNPs, most of which have a small effect on disease risk [4,5]. Although no single polymorphism may be informative clinically, a combination of APOE and non-APOE SNPs may help identify older individuals at increased risk for AD. Despite their detection of novel AD-associated genes, GWAS findings have not yet been incorporated into a genetic epidemiology framework for individualized risk prediction.
Building on a prior approach evaluating GWAS-detected genetic variants for disease prediction [6] and using a survival analysis framework, we tested the feasibility of combining AD-associated SNPs and APOE status into a continuous-measure polygenic hazard score (PHS) for predicting the age-specific risk for developing AD. We assessed replication of the PHS using several independent cohorts.
We followed three steps to derive the PHS for predicting age of AD onset: (1) we defined the set of associated SNPs, (2) we estimated hazard ratios for polygenic profiles, and (3) we calculated individualized absolute hazards (see S1 Appendix for a detailed description of these steps).
Using the IGAP Stage 1 sample, we first identified a list of SNPs associated with increased risk for AD, using a significance threshold of p < 10−5. Next, we evaluated all IGAP-detected AD-associated SNPs within the ADGC Phase 1 case–control dataset. Using a stepwise procedure in survival analysis, we delineated the “final” list of SNPs for constructing the PHS [14,15]. Specifically, using Cox proportional hazard models, we identified the top AD-associated SNPs within the ADGC Phase 1 cohort (excluding NIA ADC and ADNI samples), while controlling for the effects of gender, APOE variants, and the top five genetic principal components (to control for the effects of population stratification). We utilized age of AD onset and age of last clinical visit to estimate age-specific risks [16] and derived a PHS for each participant. In each step of the stepwise procedure, the algorithm selected the one SNP from the pool that most improved model prediction (i.e., minimizing the Martingale residuals); additional SNP inclusion that did not further minimize the residuals resulted in halting of the SNP selection process. To prevent overfitting in this training step, we used 1,000× bootstrapping for model averaging and estimating the hazard ratios for each selected SNP. We assessed the proportional hazard assumption in the final model using graphical comparisons.
To assess for replication, we first examined whether the predicted PHSs derived from the ADGC Phase 1 cohort could stratify individuals into different risk strata within the ADGC Phase 2 cohort. We next evaluated the relationship between predicted age of AD onset and the empirical (actual) age of AD onset using cases from ADGC Phase 2. We binned risk strata into percentile bins and calculated the mean of actual age of AD onset in that percentile as the empirical age of AD onset. In a similar fashion, we additionally tested replication within the NIA ADC subset classified at autopsy as having a high level of AD neuropathological change [13].
Because case–control samples cannot provide the proper baseline hazard [17], we used previously reported annualized incidence rates by age estimated from the general US population [18]. For each participant, by combining the overall population-derived incidence rates [18] and the genotype-derived PHS, we calculated the individual’s “instantaneous risk” for developing AD, based on their genotype and age (for additional details see S1 Appendix). To independently assess the predicted instantaneous risk, we evaluated longitudinal follow-up data from 2,724 cognitively normal older individuals from the NIA ADC with at least 2 y of clinical follow-up. We assessed the number of cognitively normal individuals progressing to AD as a function of the predicted PHS risk strata and examined whether the predicted PHS-derived incidence rate reflected the empirical progression rate using a Cochran–Armitage trend test.
We examined the association between our PHS and established in vivo and pathological markers of AD neurodegeneration. Using linear models, we assessed whether the PHS associated with Braak stage for NFTs and CERAD score for neuritic plaques, as well as CSF Aβ1–42 and CSF total tau. Using linear mixed effects models, we also investigated whether the PHS was associated with longitudinal CDR-SB score and volume loss within the entorhinal cortex and hippocampus. In all analyses, we co-varied for the effects of age and sex.
From the IGAP cohort, we found 1,854 SNPs associated with increased risk for AD at p < 10−5. Of these, using the Cox stepwise regression framework, we identified 31 SNPs, in addition to two APOE variants, within the ADGC cohort for constructing the polygenic model (Table 2). Fig 1 illustrates the relative risk for developing AD using the ADGC Phase 1 case–control cohort. The graphical comparisons among Kaplan–Meier estimations and Cox proportional hazard models indicate that the proportional hazard assumption holds for the final model (Fig 1).
To quantify the additional prediction provided by polygenic information beyond APOE, we evaluated how the PHS modulates age of AD onset in APOE ε3/3 individuals. Among these individuals, we found that age of AD onset can vary by more than 10 y, depending on polygenic risk. For example, for an APOE ε3/3 individual in the tenth decile (top 10%) of the PHS, at 50% risk for meeting clinical criteria for AD diagnosis, the expected age of developing AD is approximately 84 y (Fig 2); however, for an APOE ε3/3 individual in the first decile (bottom 10%) of the PHS, the expected age of developing AD is approximately 95 y (Fig 2). The hazard ratio comparing the tenth decile to the first decile is 3.34 (95% CI 2.62–4.24, log rank test p = 1.0 × 10−22). Similarly, we also evaluated the relationship between the PHS and the different APOE alleles (ε2/3/4) (first figure in S1 Appendix). These findings show that, beyond APOE, the polygenic architecture plays an integral role in affecting AD risk.
To assess replication, we applied the ADGC Phase 1–trained model to independent samples from ADGC Phase 2. Using the empirical distributions, we found that the PHS successfully stratified individuals from independent cohorts into different risk strata (Fig 3A). Among AD cases in the ADGC Phase 2 cohort, we found that the predicted age of onset was strongly associated with the empirical (actual) age of onset (binned in percentiles, r = 0.90, p = 1.1 × 10−26; Fig 3B). Similarly, within the NIA ADC subset with a high level of AD neuropathological change, we found that the PHS strongly predicted time to progression to neuropathologically defined AD (Cox proportional hazard model, z = 11.8723, p = 2.8 × 10−32).
To evaluate the risk for developing AD, combining the estimated hazard ratios from the ADGC cohort, allele frequencies for each of the AD-associated SNPs from the 1000 Genomes Project, and the disease incidence in the general US population [18], we generated population baseline-corrected survival curves given an individual’s genetic profile and age (panels A and B of second figure in S1 Appendix). We found that PHS status modifies both the risk for developing AD and the distribution of age of onset (panels A and B of second figure in S1 Appendix).
Given an individual’s genetic profile and age, the corrected survival proportion can be translated directly into incidence rates (Fig 4; Tables 3 and S1). As previously reported in a meta-analysis summarizing four studies from the US general population [18], the annualized incidence rate represents the proportion (in percent) of individuals in a given risk stratum and age who have not yet developed AD but will develop AD in the following year; thus, the annualized incidence rate represents the instantaneous risk for developing AD conditional on having survived up to that point in time. For example, for a cognitively normal 65-y-old individual in the 80th percentile of the PHS, the incidence rate (per 100 person-years) would be 0.29 at age 65 y, 1.22 at age 75 y, 5.03 at age 85 y, and 20.82 at age 95 y (Fig 4; Table 3); in contrast, for a cognitively normal 65-y-old in the 20th percentile of the PHS, the incidence rate would be 0.10 at age 65 y, 0.43 at age 75 y, 1.80 at age 85 y, and 7.43 at age 95 y (Fig 4; Table 3). As independent validation, we examined whether the PHS-predicted incidence rate reflects the empirical progression rate (from normal control to clinical AD) (Fig 5). We found that the PHS-predicted incidence was strongly associated with empirical progression rates (Cochran–Armitage trend test, p = 1.5 × 10−10).
We found that the PHS was significantly associated with Braak stage of NFTs (β-coefficient = 0.115, standard error [SE] = 0.024, p-value = 3.9 × 10−6) and CERAD score for neuritic plaques (β-coefficient = 0.105, SE = 0.023, p-value = 6.8 × 10−6). We additionally found that the PHS was associated with worsening CDR-SB score over time (β-coefficient = 2.49, SE = 0.38, p-value = 1.1 × 10−10), decreased CSF Aβ1–42 (reflecting increased intracranial Aβ plaque load) (β-coefficient = −0.07, SE = 0.01, p-value = 1.28 × 10−7), increased CSF total tau (β-coefficient = 0.03, SE = 0.01, p-value = 0.05), and greater volume loss within the entorhinal cortex (β-coefficient = −0.022, SE = 0.005, p-value = 6.30 × 10−6) and hippocampus (β-coefficient = −0.021, SE = 0.005, p-value = 7.86 × 10−5).
In this study, by integrating AD-associated SNPs from recent GWASs and disease incidence estimates from the US population into a genetic epidemiology framework, we have developed a novel PHS for quantifying individual differences in risk for developing AD, as a function of genotype and age. The PHS systematically modified age of AD onset, and was associated with known in vivo and pathological markers of AD neurodegeneration. In independent cohorts (including a neuropathologically confirmed dataset), the PHS successfully predicted empirical (actual) age of onset and longitudinal progression from normal aging to AD. Even among individuals who do not carry the ε4 allele of APOE (the majority of the US population), we found that polygenic information was useful for predicting age of AD onset.
Using a case–control design, prior work has combined GWAS-associated polymorphisms and disease prediction models to predict risk for AD [19–24]. Rather than representing a continuous process where non-demented individuals progress to AD over time, the case–control approach implicitly assumes that normal controls do not develop dementia and treats the disease process as a dichotomous variable where the goal is maximal discrimination between diseased “cases” and healthy “controls.” Given the striking age dependence of AD, this approach is clinically suboptimal for estimating the risk of AD. Building on prior genetic estimates from the general population [2,25], we employed a survival analysis framework to integrate AD-associated common variants with established population-based incidence [18] to derive a continuous measure, the PHS. We note that the PHS can estimate individual differences in AD risk across a lifetime and can quantify the yearly incidence rate for developing AD.
These findings indicate that the lifetime risk of age of AD onset varies by polygenic profile. For example, the annualized incidence rate (risk for developing AD in a given year) is considerably lower for an 80-y-old individual in the 20th percentile of the PHS than for an 80-y-old in the 99th percentile of the PHS (Fig 4; Table 3). Across the lifespan (panel B of second figure in S1 Appendix), our results indicate that even individuals with low genetic risk (low PHS) develop AD, but at a later peak age of onset. Certain loci (including APOE ε2) may “protect” against AD by delaying, rather than preventing, disease onset.
Our polygenic results provide important predictive information beyond APOE. Among APOE ε3/3 individuals, who constitute 70%–75% of all individuals diagnosed with late-onset AD, age of onset varies by more than 10 y, depending on polygenic risk profile (Fig 2). At 60% AD risk, APOE ε3/3 individuals in the first decile of the PHS have an expected age of onset of 85 y, whereas for individuals in the tenth decile of the PHS, the expected age of onset is greater than 95 y. These findings are directly relevant to the general population, where APOE ε4 accounts for only a fraction of AD risk [3], and are consistent with prior work [26] indicating that AD is a polygenic disease where non-APOE genetic variants contribute significantly to disease etiology.
We found that the PHS strongly predicted age of AD onset within the ADGC Phase 2 dataset and the NIA ADC neuropathology-confirmed subset, demonstrating independent replication of our polygenic score. Within the NIA ADC sample, the PHS robustly predicted longitudinal progression from normal aging to AD, illustrating that polygenic information can be used to identify the cognitively normal older individuals at highest risk for developing AD (preclinical AD). We found a strong relationship between the PHS and increased tau-associated NFTs and amyloid plaques, suggesting that elevated genetic risk may make individuals more susceptible to underlying AD pathology. Consistent with recent studies showing correlations between AD polygenic risk scores and markers of AD neurodegeneration [22,23], our PHS also demonstrated robust associations with CSF Aβ1–42 levels, longitudinal MRI measures of medial temporal lobe volume loss, and longitudinal CDR-SB scores, illustrating that increased genetic risk may increase the likelihood of clinical progression and developing neurodegeneration measured in vivo.
From a clinical perspective, our genetic risk score may serve as a “risk factor” for accurately identifying older individuals at greatest risk for developing AD, at a given age. Conceptually similar to other polygenic risk scores (for a review of this topic see [27]) for assessing coronary artery disease risk [28] and breast cancer risk [29], our PHS may help in predicting which individuals will test “positive” for clinical, CSF, or imaging markers of AD pathology. Importantly, a continuous polygenic measure of AD genetic risk may provide an enrichment strategy for prevention and therapeutic trials and could also be useful for predicting which individuals may respond to therapy. From a disease management perspective, by providing an accurate probabilistic assessment regarding the likelihood of AD neurodegeneration, determining a “genomic profile” of AD may help initiate a dialogue on future planning. Finally, a similar genetic epidemiology framework may be useful for quantifying the risk associated with numerous other common diseases.
There are several limitations to our study. We primarily focused on individuals of European descent. Given that AD incidence [30], genetic risk [25,31], and likely linkage disequilibrium in African-American and Latino individuals is different from in white individuals, additional work will be needed to develop a polygenic risk model in non-white (and non-US) populations. The majority of the participants evaluated in our study were recruited from specialized memory clinics or AD research centers and may not be representative of the general US population. In order to be clinically useful, we note that our PHS needs to be prospectively validated in large community-based cohorts, preferably consisting of individuals from a range of ethnicities. The previously reported population annualized incidence rates were not separately provided for males and females [18]. Therefore, we could not report PHS annualized incidence rates stratified by sex. We note that we primarily focused on genetic markers and thus did not evaluate how other variables, such as environmental or lifestyle factors, in combination with genetics impact age of AD onset. Another limitation is that our PHS may not be able to distinguish pure AD from a “mixed dementia” presentation since cerebral small vessel ischemic/hypertensive pathology often presents concomitantly with AD neurodegeneration, and additional work will be needed on cohorts with mixed dementia to determine the specificity of our polygenic score. Finally, we focused on APOE and GWAS-detected polymorphisms for disease prediction. Given the flexibility of our genetic epidemiology framework, it can be used to investigate whether a combination of common and rare genetic variants along with clinical, cognitive, and imaging biomarkers may prove useful for refining the prediction of age of AD onset.
In conclusion, by integrating population-based incidence proportion and genome-wide data into a genetic epidemiology framework, we have developed a PHS for quantifying the age-associated risk for developing AD. Measures of polygenic variation may prove useful for stratifying AD risk and as an enrichment strategy in clinical trials.
|
10.1371/journal.ppat.1000584 | Getting to Grips with Strangles: An Effective Multi-Component Recombinant Vaccine for the Protection of Horses from Streptococcus equi Infection | Streptococcus equi subspecies equi (S. equi) is a clonal, equine host-adapted pathogen of global importance that causes a suppurative lymphodendopathy of the head and neck, more commonly known as Strangles. The disease is highly prevalent, can be severe and is highly contagious. Antibiotic treatment is usually ineffective. Live attenuated vaccine strains of S. equi have shown adverse reactions and they suffer from a short duration of immunity. Thus, a safe and effective vaccine against S. equi is highly desirable. The bacterium shows only limited genetic diversity and an effective vaccine could confer broad protection to horses throughout the world. Welsh mountain ponies (n = 7) vaccinated with a combination of seven recombinant S. equi proteins were significantly protected from experimental infection by S. equi, resembling the spontaneous disease. Vaccinated horses had significantly reduced incidence of lymph node swelling (p = 0.0013) lymph node abscessation (p = 0.00001), fewer days of pyrexia (p = 0.0001), reduced pathology scoring (p = 0.005) and lower bacterial recovery from lymph nodes (p = 0.004) when compared with non-vaccinated horses (n = 7). Six of 7 vaccinated horses were protected whereas all 7 non-vaccinated became infected. The protective antigens consisted of five surface localized proteins and two IgG endopeptidases. A second vaccination trial (n = 7+7), in which the IgG endopeptidases were omitted, demonstrated only partial protection against S. equi, highlighting an important role for these vaccine components in establishing a protective immune response. S. equi shares >80% sequence identity with Streptococcus pyogenes. Several of the components utilized here have counterparts in S. pyogenes, suggesting that our findings have broader implications for the prevention of infection with this important human pathogen. This is one of only a few demonstrations of protection from streptococcal infection conferred by a recombinant multi-component subunit vaccine in a natural host.
| Numerous research groups have vaccinated, using recombinant antigens, against streptococcal infections in mouse model systems and shown protection. We have here demonstrated efficient protective vaccination of the natural host, the horse, using recombinant antigens. Streptococcus equi subspecies equi (S. equi) is an equine host-adapted and highly contagious pathogen of global importance. Six out of seven Welsh mountain ponies vaccinated with a combination of seven recombinant S. equi proteins were protected from experimental infection as assessed by clinical examination, pyrexia, lymph node swelling, inflammation, bacterial recovery, and post mortem examination. The protective antigens consisted of five surface localized proteins and two endopeptidases that are specific for IgG; the latter were shown to be of major importance for efficacy. Several of the antigens used here have similarities in Streptococcus pyogenes, implying that our findings are of importance for development of a vaccine against this important human pathogen.
| Access to the genome sequence data of bacterial pathogens permitting the identification of surface exposed and secreted proteins has long been anticipated to revolutionize vaccine design, referred to as reverse vaccinology [1],[2]. However, few vaccines have been taken beyond studies in mouse model systems and shown to confer protection against challenge infection in the natural host.
Strangles, caused by Streptococcus equi subsp. equi (S. equi), is characterized by abscessation of the lymph nodes of the head and neck of the horse and is of significant welfare and economic importance. The development of effective preventative vaccines has been slow. A non-encapsulated strain of S. equi (Pinnacle IN™) has been used as a nasal vaccine against strangles, but has not been licensed for sale in Europe due to safety concerns. A second live attenuated vaccine was marketed in Europe [3] (Equilis StrepE), but was withdrawn in 2007. Safety concerns have also been raised over the use of Equilis StrepE [4],[5]. A safe and effective vaccine against S. equi is thus highly desired.
S. equi evolved from an ancestral strain of S. equi subsp. zooepidemicus (S. zooepidemicus). The population of the S. zooepidemicus group is extremely diverse and consists of at least 218 sequence types, whereas isolates of S. equi from the USA, Canada, Australia and Europe are either ST-179 or a single locus variant, ST-151 [6] (http://pubmlst.org/szooepidemicus/). The limited genetic diversity of S. equi suggests that an effective vaccine could confer broad protection to horses throughout the world.
We have demonstrated previously that vaccination of Welsh mountain ponies with EAG [7],[8], SclC [9] and CNE [10] (Trivacc) conferred partial protection against challenge by S. equi [11]. The amount of nasal discharge, the number of bacteria recovered from nasal washes and the occurrence of abscess material (empyema) in guttural pouches, following rupture of abscesses formed in the retropharyngeal lymph nodes, differed significantly between the vaccinated group and a non-vaccinated control group. However, clinical scoring and mean rectal temperatures did not differ significantly. This experiment thus showed that parameters of importance for spreading disease between horses were reduced significantly, but that the level of protection in individual horses was limited [11]. Using a set of seven recombinant proteins (Septavacc), we demonstrate here that 85% protection was obtained.
Five of the antigens in the Septavacc composition are predicted to be localized on the surface of S. equi (EAG [7],[8], CNE [10], SclC [12], SEQ0256 and SEQ0402 [13]) through sortase-mediated attachment to the peptidoglycan cell wall. EAG binds to albumin, α-2 macroglobulin (A2M) and IgG [8],[14]. CNE binds to collagen [10], and is located within the FimI pilus locus of S. equi and S. zooepidemicus [13],[15]. SclC is a member of a collagen-like protein family, which in S. equi consists of seven members, each with a unique N-terminal domain of unknown function [12]. The proteins encoded by SEQ0256 and SEQ0402 contain features typical of cell surface anchored proteins and an N-terminal non-repetitive domain. The N-terminal domains were used in this study, the functions of which are unknown and neither shows homology to any characterized protein. The two additional antigens in Septavacc, IdeE and IdeE2 are IgG endopeptidases, where IdeE2 has greater activity towards horse IgG. Both IdeE and IdeE2 are predicted to be secreted [16],[17] and IdeE has an antiphagocytic activity by binding directly to neutrophils [17].
The antigens in Septavacc were selected from a larger antigen pool based on the level of protection conferred in an experimental mouse model of strangles. Mice were immunized with recombinant antigens either individually, or in combination, and subsequently experimentally infected with S. equi. Following challenge, mice were examined daily for loss of weight and for nasal colonization in comparison with non-vaccinated controls. The reproducibility of the model is sufficiently robust to allow comparison between different experiments. A ranking list based on the protective efficacy of the different antigens could therefore be generated (Table 1). Vaccination with FNE, SFS and FNEB did not result in protection although good antibody titers were obtained with these antigens [18]. Good immune responses were obtained to all other antigens with the exception of IdeE2 and all of these antigens conferred significant protection in mice (Table 1).
Seven Welsh mountain ponies were vaccinated with Septavacc and an additional seven ponies were given adjuvant only as control via both the subcutaneous and intranasal routes, followed by experimental infection with 1×108 colony forming units (cfu) of S. equi strain 4047. Antibody responses against the antigens in serum samples and nasal washes were analyzed by ELISA (Figures 1A, B and C). All ponies responded well and it was noted that IgG responses in nasal washes and in sera had low correlation in individual ponies (R2 from 0.01 to 0.28); a pony could respond well in sera but less so in nasal washing and v.v. This implies the generation of independent immune responses in mucosa and sera, possibly as a result of the two routes of immunization employed. Exudation of IgG from sera to mucosal surfaces presumably also contributes to mucosal IgG. A high background level of IgG (in pre immune sera) against IdeE is presumably due to non-immunologic binding to IdeE. IgA responses in nasal washes against SEQ0256, SEQ0402, IdeE and IdeE2 were moderate but significant for SEQ0402, IdeE and IdeE2 with p values 0.04, 0.02 and 0.05 respectively.
In a separate study, seven ponies were immunized with a Pentavacc formulation, containing the same five surface antigens as Septavacc, but with the omission of IdeE and IdeE2. Serum samples and nasal washes were taken every month. Significantly elevated IgG levels to the five surface proteins could be detected 6 months post V3 and an additional booster (V4) on day 270 led to a rapid increase of IgG in sera against all antigens (Figure 2A). IgG against the antigens in nasal washes also persisted for a long time as shown in Figure 2B.
Sera from Septavacc vaccinated ponies were also tested for the ability to inhibit the IgG cleaving activity of IdeE and IdeE2. Sera pooled from ponies immunised with Septavacc could be diluted 16 fold and still inhibit the cleavage of human IgG by IdeE. The inhibitory activity against IdeE2 cleavage was much less. At a 2 fold serum dilution the cleavage of horse IgG was inhibited (Figure 3A and B). Neither pre-immune sera from the Septavacc vaccinated ponies nor sera from Pentavacc vaccinated ponies had any effect on the IgG cleavage activity of IdeE or IdeE2 (data not shown).
The swelling and abscessation of submandibular lymph nodes is a typical clinical sign of infection with S. equi. The mean lymph node scores in Septavacc vaccinated ponies differed from the control group and the number of days where an individual pony's score exceeded 2, differed significantly (p = 0.0013) (Figure 4).
The normal rectal temperature of Welsh mountain ponies is 37–38°C and a pony with a rectal temperature of 39°C or higher is considered pyrexic. All ponies in the control group became pyrexic at some stage during challenge compared to only one pony in the Septavacc vaccinated group. The accumulated number of days that individual ponies in the vaccinated or control groups were pyrexic was 5 and 30 days, respectively (p = 0.0001) (Figure 5).
Infection with S. equi leads to a systemic inflammation, manifested as an increase in blood fibrinogen and neutrophil levels. As shown in Figure 6A and B, fibrinogen and neutrophil levels of ponies vaccinated with Septavacc remained normal, whereas the non-vaccinated group had significantly higher mean values.
Ponies vaccinated with Pentavacc, containing the same antigens as in Septavacc but omitting the IgG endopeptidases, were challenged 14 days after the last booster. These ponies differed from the corresponding control group in terms of elevated temperature, fibrinogen levels and nasal discharge, but not significantly so. One pony was protected fully.
To minimize suffering and in accordance with our strict ethical and welfare code, ponies were euthanized as soon as clinical signs of S. equi infection became apparent. All of the control ponies were euthanized between 8 to 12 days post challenge. Vaccinated ponies, however, had reduced clinical signs and all ponies reached the end of the study, 21 days post challenge. Following euthanasia, all of the ponies were subject to post mortem examination to quantify the level of pathology observed using a scoring system as described in Materials and Methods.
Figure 7 summarizes the individual post mortem scores of ponies vaccinated with Trivacc (EAG, SclC, CNE) [11], Pentavacc and Septavacc, containing three, five and seven antigens respectively. Increasing the number of antigens comprising each vaccine significantly reduced the post mortem score. Only one of the ponies vaccinated with Septavacc had lymph node abscesses, compared with abscesses in all seven non-vaccinated ponies. To confirm these gross pathological findings, samples from ponies vaccinated with Septavacc were examined histopathologically and scored using a system as described in Materials and Methods. Again, significant differences were seen between the Septavacc and control groups (p = 0.006) (Figure 8). Histopathological examination of the left and right retropharyngeal and submandibular lymph nodes identified 19 lymph node abscesses in the control ponies and 3 lymph node abscesses in a single pony in the Septavacc group (p = 0.00001). Seventy-eight and 10% of all lymph nodes were culture positive for S. equi in the control and vaccine groups respectively (p = 0.004).
Taking all of the results together, vaccination with Septavacc resulted in 85% protection from disease, with only one vaccinated pony out of seven developing lymph node abscesses.
This study is one of only a few demonstrations of protection in a natural host from streptococcal infection conferred by a recombinant multi-component subunit vaccine. No significant adverse effects were seen in any of the vaccinated ponies, demonstrating that both the recombinant antigens and the adjuvant were safe. It is also clear that the antiphagocytic capsule and other immune evasion mechanisms employed by S. equi were not sufficient to counter successful vaccination with the recombinant proteins used here. The approach taken here is significantly safer than currently available live attenuated strains of S. equi.
The large difference in efficacy between Pentavacc and Septavacc (p = 0.036 for post mortem scoring), suggests that the inclusion of one or both of the endopeptidases IdeE and IdeE2 is important for protection in the natural host. As antibodies in sera from the Septavacc vaccinated ponies can inhibit the IgG cleaving activities of IdeE and IdeE2, it is conceivable that these antibodies prevent the destruction of antibodies directed against the other five components in Septavacc .We have recently found that immunization of mice with IdeE and IdeE2 as single antigens confer protection in the mouse model of strangles [19]. We assume this effect might be due to opsonising antibodies targeting IdeE and IdeE2 during the secretion process. It cannot be ruled out that the difference in efficacy between Pentavacc and Septavacc might also be attributed to the different vaccination schedules. The ponies vaccinated with Pentavacc were infected 6.5 months later than those vaccinated with Septavacc but were therefore given an additional booster (V4) two weeks before challenge. After this booster, antibody titers in these ponies reached the same level as in ponies vaccinated with Septavacc. With the exception of IdeE and IdeE2, the contribution of each protein to the protective activity of Septavacc has not been addressed in this study. Neither have we attempted to analyze in depth the mechanism of action of the protective antibodies. The activity of antibodies against surface localized antigens might de dual: adherence blocking and opsonic. We have previously found that antibodies against CNE effectively block adherence of S. equi to collagen [20].
With the exception of the 5′ variable region of SeM [21] the genomic variation between isolates of S. equi is thought to be minimal and all strains analyzed to date by MLST are either ST-179 or the single locus variant ST-151. The antigens used in this study were all cloned from the Swedish S. equi strain 1866, which is SeM type-9, whilst the S. equi strain 4047 challenge strain is SeM type-3 and was isolated in the UK. Therefore, it is likely that the antigens present in Septavacc will confer broad protection against S. equi strains from around the world.
S. equi shares >80% sequence identity with Streptococcus pyogenes [13] and several components utilized in our studies share similarity with S. pyogenes antigens, either by homology or function. The S. pyogenes gene encoding the collagen binding protein Cpa is located in the variable FCT region (fibronectin- and collagen-binding T-antigen) and is part of a pilus-like structure [22]. Similarly, cne is located in a pilus locus (FimI) that includes genes encoding SrtC.1 and a putative backbone pilus subunit suggesting that CNE is also attached to a pilus-like structure [13]. EAG, like GRAB from S. pyogenes, binds the proteinase inhibitor A2M [8],[23],[24]. SclC is one of seven collagen-like surface proteins in S. equi, whilst S. pyogenes genomes contain two such putative proteins, SclA and SclB [9]. The IgG-specific endopeptidases used here, IdeE and IdeE2 [19], are similar both in function and amino acid sequence to IdeS/Mac/sib35 of S. pyogenes [16],[17],[25],[26]. Antibodies against IdeS in convalescent patients were able to neutralize its function [27]. Interestingly, Cpa (plus other pili components) and Sib35 have been identified as protective antigens in mouse models of S. pyogenes infection. The Cpa combination and Sib35 gave good protection in a mouse system whereas antibodies against GRAB could only opsonize capsule-deficient mutants of S. pyogenes [28]–[30]. Thus, it is conceivable that vaccination of humans with a combination of S. pyogenes antigens similar to the ones used in Septavacc could prove effective against this important human pathogen
In a study by Timoney et al., [31] two combinations of recombinant proteins derived from S. equi (SzPSe, CNE, Se51.9, Se44.2 and Se46.8 or SeM, Se44.2, Se75.3, Se42.0, Se110.0 and Se18.9) were tested as vaccines against strangles. However, neither combination protected horses from infection with S. equi [31]. Two of these proteins CNE and IdeE2 (Se44.2) are included in the Septavacc vaccine, suggesting that the additional components of Septavacc are important in generating a protective immune response. It was also suggested that an effective strangles vaccine should result in immune-mediated tonsillar clearance since tonsillar adherence is a crucial early step in the pathogenesis of strangles [31],[32]. If this is the case, the route of immunization and choice of adjuvant, which differ between these studies, might be of utmost importance. We have generally noted that mice immunized by the intranasal route are far better protected than those immunized subcutaneously. IgA responses in nasal washes were obtained in this study. It is likely that the ability to block bacterial adherence to mucosal surfaces is of importance to the protection observed in this study. A good mucosal immune response was obtained in the mouse by using nanospheres to which S. equi proteins, extracted from the bacterial wall, were adsorbed [33],[34]. The poorer efficacy observed with Trivacc could be due, not only to fewer antigens, but also to the fact that immunizations were only intra muscular and i.n. In this study, immunizations were both subcutaneous, near the retropharyngeal lymph node, and i.n. The necessity of both routes of immunizations will be determined in future studies.
In conclusion, we have demonstrated that using recombinant antigens, a protective immune response against a streptococcal infection can be obtained in the natural host, not just in a mouse model system. Immune protection by vaccination does not necessarily require an attenuated live vaccine or vaccination with killed bacteria, conventional strategies for vaccine design.
Approval for mice experiments was obtained from Swedish Animal Welfare Agency. The pony studies were conducted under a Home Office Project License. The Animal Health Trust Ethical Review Committee approved the Research Program Proposal for these studies.
S. equi strain 1866 SeM type-9 (obtained from Nordvacc Läkemedel AB, Sweden) was used as source DNA for cloning the antigens and used as the challenge strain in mouse infection experiments. The S. equi strain 4047, SeM type-3 was isolated from a case of strangles (Animal Health Trust), was used as the challenge strain in horse infection experiments. To clone and express the recombinant antigens, SEQ0256, SEQ0402, SEQ0944, SEQ0936, and IdeE2, the E. coli strains ER2566 or BL21(DE3) and plasmid vector pTYB4, belonging to the IMPACT™ Protein Purification System (New England Biolabs Inc., MA) were used. The E. coli strain BL21(DE3) and plasmid vector pGEX-6P-1 belonging to GST-glutathione affinity system (GE Healthcare) were used to clone and express IdeE. Plasmids and PCR amplified DNA fragments were purified using the QIAprep Spin Miniprep (Qiagen, Hilden, Germany).
Strains of S. equi were grown on blood agar plates or in Todd-Hewitt broth (Oxoid, Basingstoke, Hampshire, United Kingdom) and E. coli clones were cultured in Luria-Bertani (LB) broth, supplemented with ampicillin (50 µg/ml), or on LAA plates [LB-broth with ampicillin and agar (15 g/L)]. Incubations were at 37°C unless otherwise stated.
The S. equi strain 4047 genome database (www.sanger.ac.uk/) was analyzed to select open reading frames encoding a number of extra cellular proteins to be tested as potential antigens in the vaccination trials. To identify the predicted signal sequences, the computer program SignalP (http://www.cbs.dtu.dk/services/SignalP/) was used.
Chromosomal DNA from S. equi strain 1866 was used as a template to PCR amplify the genes (or part of the genes) encoding SEQ0256, SEQ0402, SEQ0944, SEQ0936, IdeE, and IdeE2. The sequences of primers used are listed in Table 2. Cleavage sites for restriction enzymes were included in the primer sequences to match the cloning sites in the plasmid vectors. The PCR amplifications were performed using the primers (20 pmol/µl) and ReadyToGo™ PCR beads (GE Healthcare) using a standard PCR programme (Step 1, pre-heat 1 minute at 95°C, DNA strand separation; Step 2, 30 seconds at 95°C; Step 3, annealing 15 seconds at usually 5 degree below the melting temperature; and Step 4, elongation for 2 minutes at 72°C, Steps 2–4 were run for 30 cycles.) The PCR products were analyzed on an agarose gel, and thereafter purified using the QIAquick PCR Purification Kit™ (Qiagen). After restriction enzyme cleavage the fragments were purified again using the kit mentioned above before ligation into the respective vector using the ReadyToGo T4DNA Ligase (GE Healthcare). After ligation, the samples were transformed into competent E. coli strain ER2566 or BL21(DE3) using electroporation, spread on LAA plates and incubated over night at 37°C. Next day, a number of colonies were analyzed by PCR for the presence of inserts, using the respective primer combination. Clones with correct inserts were further analyzed by DNA sequencing.
To produce and purify the recombinant proteins SEQ0256, SEQ0402, SEQ0936, SEQ0944, [13] and IdeE2, the expression and purification system IMPACT™ T7 (NEB) was used. Briefly, following the manufacturer's instructions the clones containing the recombinant plasmids were grown at 37°C in LB media supplemented with ampicillin (final conc. 50 µg/ml). At an optical density (OD600 nm)∼0.6, the growth media were supplemented with IPTG (final conc. 0.3 mM) and the growth temperature shifted to 20°C. After incubation over night the cells were harvested and resuspended in a buffer [20 mM Tris-HCl (pH 8.0), 500 mM NaCl, 0.1 mM EDTA, and 0.05% (v/v) TWEEN20] and lysed by freezing and thawing. After centrifugation, the supernatants were sterile filtrated and applied onto a chitin column. The columns were washed extensively using the same buffer and treated subsequently with cleavage buffer [20 mM Tris-HCl (pH 8.0), 50 mM NaCl, 0.1 mM EDTA, and 30 mM dithiothreitol (DTT)]. The eluted samples containing the antigens were dialysed against phosphate-buffered saline [PBS; 137 mM NaCl, 2.7 mM KCl, 10 mM Na2HPO4, 1.4 mM KH2PO4 (pH 7.4)].
Recombinant IdeE was produced using the GST-glutathione affinity system. Briefly, according to the procedure described above, after growth, induction and harvest, the E. coli cells were suspended in PBS supplemented with TWEEN20, final conc. 0.1% (v/v) (PBST) whereupon the cells were lysed by freezing and thawing. After centrifugation, the supernatant was sterile filtrated and batch purified with Glutathione-sepharose beads. After extensive washing using PBST the fusion protein was treated with scissor protease to release IdeE. Finally, the amounts of antigens obtained were determined using spectrophotometry and the quality analyzed by SDS-PAGE coomassie staining. The proteins were stored finally at −20°C. IdeE and IdeE2 were produced as full-length protein with a few extra amino acids from the vectors and both proteins showed IgG-cleaving activity. It should be noted that IdeE2 had a tendency to partially precipitate upon freezing.
The production of EAG, SclC and CNE has been described previously [11].
The cleaving activity of IdeE was assayed against human IgG (Bethyl Laboratories inc., Montgomery, TX) whereas the IdeE2 activity was measured against horse IgG (Bethyl Laboratories inc., Montgomery, TX). Before cleavage, DTT was added to the endopeptidases to a final concentration of 0.5 mM. Thereafter the enzymes were added to IgG (1 mg/ml) for 30–60 minutes and cleavage was visualised by coomassie blue staining of SDS-PAGE gels.
To test for inhibitory activity of sera against the endopeptidases the same type of assay as above was used. Briefly, Septavacc sera from six ponies were pooled and diluted in steps of two. As controls, pre-immune Septavacc sera and Pentavacc sera were also pooled and diluted. The assay was performed by mixing 2 µl of either endopeptidase (15 µg/ml) separately with 2 µl of diluted sera for 15 min in room temperature after which 8 µl of IgG (1 mg/ml) was added and incubated for 45 minutes at 37°C and the samples analysed on SDS-PAGE.
Prior to experimental challenge infection, female NMRI mice (n = 15) were vaccinated with recombinant proteins derived from S. equi as described earlier [20]: 12 µg of each antigen and 10 µg of adjuvant (Abisco 300, Isconova, Uppsala, Sweden [35]) were given intranasally on days 1, 14 and 21. Control mice (n = 15) were given adjuvant only. Blood samples were collected for assessment of antibody titers in ELISA [20].
Infection of mice with S. equi strain 1866 was performed 7 days after the final booster as described previously [20]. Briefly, 106 CFU of S. equi strain 1866 (cultivated in THB+10% horse serum+1% yeast extract for 4 hours in a 5% CO2 enriched atmosphere) were given intranasally to anesthetized mice. Weight loss and colonization of nostrils were followed daily. Blood agar plates containing gentiana violet, to select for streptococcal growth, were placed gently onto the nostrils. Bacteria were spread out on the plates, which were then incubated overnight in a 5% CO2 enriched atmosphere. A scoring system was used where 0–5 colonies = 0, 5–100 = 1, >100 colonies = 2 and confluent growth = 3.
Healthy Welsh Mountain Ponies (n = 7) were vaccinated with Septavacc via administration of 1 ml subcutaneous (s.c.) injections bilaterally close to the retropharyngeal lymph nodes and 2 ml intranasally (i.n.) by spraying into each nostril on days 4, 60, and 74 (V1, V2 and V3). The Septavacc vaccine doses contained 150 µg for i.n. and 50 µg for s.c. injections of each antigen (EAG, CNE, SclC, SEQ0256, SEQ0402, IdeE, and IdeE2). Abisco 300 (Isconova, Uppsala, Sweden [35]) (500 µg per i.n. dose) and Abisco 200 (375 µg per s.c. dose) were used as adjuvants. Septavacc ponies were challenged on day 88. Pentavacc vaccinated ponies (n = 7) followed the same vaccination protocol as above, but were given an additional booster vaccination (V4) on day 270 and challenged on day 284. Negative control ponies were given adjuvant only, mixed with PBS (n = 7). Sera and nasal washes were taken regularly to quantify IgG responses by ELISA [11]. IgA was determined in nasal washes at a 2-fold dilution, using mouse anti horse-IgA monoclonal antibody (Serotec, Oxford, UK) followed by rabbit anti mouse-IgG HRP conjugated antibodies for detection (Dako, Denmark).
Ponies were transferred to a containment unit three days before challenge. Two weeks after the final booster immunization, each pony was challenged with S. equi strain 4047 administered via the spraying of a 2 ml culture containing 5×107 cfu into each nostril. Bacteria were grown overnight in Todd Hewitt broth and 10% fetal calf serum (THBS) in a 5% carbon dioxide enriched atmosphere at 37°C, diluted 40-fold in fresh pre-warmed THBS, further cultivated and harvested at an OD = 0.3. This infection dose has been shown to optimize the infection rate, whilst avoiding overwhelming the host immune response, as determined in previous studies [11],[36].
Ponies were monitored for the onset of clinical signs of disease over a period of three weeks post challenge by daily physical examination, rectal temperature, lymph node swelling and nasal discharge scoring. Blood samples were taken for evaluation of fibrinogen concentration as described in [11] and neutrophil levels by total white blood count performed on Beckman-Coulter ACTdiff analyzer with a manual differential count to calculate % neutrophils.
The level of swelling of SMLNs was defined as 0 = normal, 1 = slight swelling, 2 = moderate swelling, 3 = severe swelling and 4 = abscessated [11]. Bilateral swelling of submandibular lymph nodes was scored separately
Post mortem examination was performed on all ponies following the onset of clinical signs of infection or on reaching the study endpoint at 3 weeks post challenge. The severity of disease pathology was quantified according to a scoring system described previously [11]. The severity of disease on histopathological examination was scored according to the following scoring system: empyema of guttural pouch 5, lymph node abscessation 5, pharyngitis 1, lymphadenitis 1, and rhinitis 1.
Fischer's exact test was used for comparison of values from arbitrary scoring using a cut-off value splitting the group into “low/negative” or “high/positive”. Cut-off values were for nasal colonization in mice 1.5; for lymph node scoring 2; for pyrexia 39°C. Mann Whitney test was used for post mortem and histopathology scoring in ponies. T-test was used to compare temperatures, fibrinogen and neutrophil levels in ponies and weight loss in mice.
|
10.1371/journal.pntd.0001885 | A Novel Procedure for Precise Quantification of Schistosoma japonicum Eggs in Bovine Feces | Schistosomiasis japonica is a zoonosis with a number of mammalian species acting as reservoir hosts, including water buffaloes which can contribute up to 75% to human transmission in the People's Republic of China. Determining prevalence and intensity of Schistosoma japonicum in mammalian hosts is important for calculating transmission rates and determining environmental contamination. A new procedure, the formalin–ethyl acetate sedimentation-digestion (FEA–SD) technique, for increased visualization of S. japonicum eggs in bovine feces, is described that is an effective technique for identifying and quantifying S. japonicum eggs in fecal samples from naturally infected Chinese water buffaloes and from carabao (water buffalo) in the Philippines. The procedure involves filtration, sedimentation, potassium hydroxide digestion and centrifugation steps prior to microscopy. Bulk debris, including the dense cellulosic material present in bovine feces, often obscures schistosome eggs with the result that prevalence and infection intensity based on direct visualization cannot be made accurately. This technique removes nearly 70% of debris from the fecal samples and renders the remaining debris translucent. It allows improved microscopic visualization of S. japonicum eggs and provides an accurate quantitative method for the estimation of infection in bovines and other ruminant reservoir hosts. We show that the FEA-SD technique could be of considerable value if applied as a surveillance tool for animal reservoirs of S. japonicum, particularly in areas with low to high infection intensity, or where, following control efforts, there is suspected elimination of schistosomiasis japonica.
| Schistosomiasis japonica, a chronic human parasitic disease in the People's Republic of China, the Philippines and areas of Indonesia, is a zoonosis with over 40 different mammals, including a number of ruminants, that can act as reservoir hosts for the infection. Precise identification of the major infection reservoirs is important for the control of Schistosoma japonicum as their targeted treatment can prevent environmental contamination and transmission of the parasite, thus reducing the risk to humans. Current copro-parasitological techniques are generally unsatisfactory for identifying and quantifying S. japonicum eggs in ruminant feces due to the large volume of cellulosic debris present. The new approach we describe here, the FEA–SD technique, removes much of this material by sieving and centrifugation with ethyl actate and renders any remaining debris transparent by use of a potassium hydroxide (KOH) digestion, providing much improved visualization of eggs, enabling the collection of more accurate data on S. japonicum infection in ruminants. This new tool will be of particular value for monitoring schistosome prevalence and intensity in animal reservoirs in areas of the People's Republic of China that are heading toward schistosomiasis elimination.
| Schistosoma japonicum, the causative agent of Asian schistosomiasis, is endemic to the People's Republic of China, the Philippines and small pockets of Indonesia [1]–[7]. Unlike African schistosomiasis, mainly caused by S. mansoni and S. haematobium, schistosomiasis japonica is a zoonosis and it is estimated that over 40 mammalian species, comprising 28 genera and 7 orders of wild and domestic animals, can act as reservoirs and harbour S. japonicum infection [8]. The range of mammalian hosts complicates schistosomiasis control efforts and, as well as the public health considerations, the disease adds to the economic burden of communities as S. japonicum infection debilitates domestic livestock that are used for food and as work animals [9], [10]. Bovines, particularly water buffaloes (Bubalus bubalis), have been shown to be major reservoir hosts for schistosomiasis japonica in the lake areas and marshlands of southern the People's Republic of China [11]–[16]. However, their role in schistosome transmission has yet to be fully determined in other endemic areas, notably the Philippines, due partly to inconsistent results obtained with the different methods used for identifying and quantifying S. japonicum eggs in mammalian hosts [17]–[20]. In particular, the presence of bulk debris, including cellulosic fibrous material, in the feces of ruminants often obscures the eggs and impairs their visualization across all current copro-parasitological methods that involve microscopy.
Here we describe a new copro-parasitological method, the formalin–ethyl acetate sedimentation-digestion (FEA–SD) technique, which eliminates much of this bulk debris and cellulose material and facilitates much improved microscopic examination of S. japonicum eggs in the feces of bovines and other ruminant hosts. We show the FEA-SD technique is an effective technique for identifying S. japonicum eggs using fecal samples from naturally infected Chinese water buffaloes (Bubalis bubalis) and confirm its reproducibility using parasite-positive samples obtained from carabao, a subspecies (Bubalus bubalis carabanesis), in the Philippines. We show that the FEA-SD method is as efficient as real-time PCR (qPCR) for determining schistosome prevalence in bovines but is less costly to implement.
The conducts and procedures involving animal experiments were approved by the Animals Ethics Committee of the Queensland Institute of Medical Research (project no. P288). This study was performed in accordance with the recommendations of the Australian code of practice for the care and use of animals for scientific purposes, 2004.
Stool samples were taken intra-rectally from 13 water buffaloes, collected from Jiangxi province, People's Republic of China, shown to be naturally infected with S. japonicum, by the miracidial hatching test (MHT) [21]. The FEA–SD technique was then used on the positive samples to identify S. japonicum eggs, calculating egg recovery rates and bulk debris reduction. Full details of the procedure are as follows: First, bovine stool samples are collected rectally from the animals (approximately 500 g each) by a veterinarian or trained personnel. Each stool sample is homogenized with an applicator stick and 50 g of the stool mixture is taken and mixed to a slurry in a beaker with 300 ml of water. The slurry is then sieved, by pouring the slurry onto a 60 copper mesh (Tyler scale with a pore opening size of 250 µm) and using water to flush the smaller sediment onto a 260 copper mesh (61 µm), held below the 40–60 mesh. Sediment caught on the 260 mesh is washed with water into a conical flask and allowed to sediment naturally for 30 minutes. The excess water is removed, leaving sediment which is poured into a 50 ml tube. Approximately 50 ml of water is added to the conical flask and naturally sedimented for 30 minutes, again removing excess water and pouring sediment into the same 50 ml tube. This is repeated once more, to ensure all sediment is in the 50 ml tube. The tube is topped up to 50 ml with 10% formalin (v/v) and mixed thoroughly by vortexing, before standing at room temperature for 30 minutes to fix the eggs. The Falcon tube is then vortexed and, using a pasteur pipette, 10 ml of the suspension (equivalent to 10 g feces) placed into two 15 ml tubes (5 ml in each; equivalent to 5 g feces) labelled A and B. Ten percent (v/v) formalin solution is added to tubes A and B to take the volume up to 8 ml and the tubes mixed thoroughly by vortexing, after which 4 ml of 100% (v/v) ethyl acetate is added using a glass pipette and vigorously vortexed once more for 30 sec. With the cap of the tubes slightly loosened the tubes are centrifuged at 500 g for 10 minutes, resulting in a four-layer separation (Figure 1). It is important for this step that the tubes are spun at 500 g for stable and efficient debris removal. The ethyl acetate is removed from both tubes by gently rimming the bulk debris layer with a thin applicator stick and decanting the top three layers which are then discarded. Ethyl acetate and 10% (v/v) formalin can be added and spun again if necessary to remove further bulk debris. If the middle layer of bulk debris is very thin the tube should be shaken vigorously and the re-spun for better efficiency. Water is added to the remaining pellet to take the volume up to 5 ml and an equal volume of 10% (w/v) potassium hydroxide (KOH) is added to each tube. The tubes are mixed gently by vortexing to resuspend the pellet and the sample is digested overnight at 37°C.
After digestion, the sample is vortexed vigorously and centrifuged at 900 g for 10 minutes. The pellet is washed once with 10–15 ml of water to remove any residual KOH by centrifuging the solution for 10 minutes at 900 g, the supernatant removed and the final pellet is resuspended in 4–6 ml of water (as the samples are now fixed) and stored at 4°C. The sample is now ready for counting the S. japonicum eggs by microscopy. Before counting, the suspension is mixed gently with a pipette and the total volume of tubes A and B counted for each sample, pipetting 200–300 µl onto each slide for microscopy. It is important that this is done as the sensitivity of the procedure increases with the amount of suspension examined. The microscopy was performed blind by two independent microscopists, although this is not essential to the completion of the procedure. Infection intensity (eggs per gram of feces, EPG) is calculated based on the total egg number in 10 g of feces (i.e., the contents of tube A plus tube B).
In order to determine the effectiveness and reproducibility of the FEA-SD technique in reducing the bulk debris and cellulosic material present in water buffalo feces, stool samples from the 13 Chinese animals were individually processed and the initial and final volumes of debris measured and compared (Table 1). Egg recovery was measured by microscopic examination of each of the normally discarded top 3 layers (Figure 1) for the presence of eggs (Table 2). Eggs in the final sediment were counted as per the protocol described above (Table 3).
Twenty-five carabao fecal samples were collected intra-rectally from S. japonicum-endemic barangays (villages) from Western Samar province, the Philippines. These were subjected to the FEA-SD procedure applied earlier in China with the debris reduction measured (Table 2) and the final egg counts from Tubes A and B for each of the 25 samples determined (Table 4). Full details of the sample collection and methods used can be found in Gordon et al. 2012 [22].
The FEA-SD was compared directly with other fecal examination techniques including Kato-Katz (KK), the MHT, a validated qPCR assay and conventional PCR on 44 fecal samples collected during the same survey in Western Samar province referred to above. Full details of the sample collection and methods used can be found in Gordon et al. 2012 [22].
The FEA-SD technique removed an average of 61.5% — 69.2% of the bulk cellulose debris from the Philippine and Chinese, respectively, bovine stool samples prior to microscopic examination (Tables 1, 2). Any remaining debris was rendered transparent by the potassium hydroxide digestion step, so that eggs were readily observed compared with previous copro-parasitological techniques employing sieving only (Figure 2). Few eggs were present in the discarded bulk debris and supernatant, with an average of 95.2% of the total eggs recovered found in the final sedimented pellet (Table 3). Prevalence determined by the FEA-SD in the study undertaken on water buffaloes in the Philippines showed the FEA-SD (93.2%, 95% CI 85.4–100) had a similar sensitivity (90.9%, 95% CI 82.1–99.8) as the qPCR assay. By contrast the conventional PCR (31.8%, 95% CI 17.5–46.1), KK (25%, 95% CI 11.7–38.3) and MHT (19.1%, 95% CI 0.9–41.2) (Figure 3) gave much lower prevalence [21].
The amount of faeces excreted in one defecation by a large animal such as a water buffalo can exceed 45 kg so it is important to determine the minimal amount of feces that can provide optimal and consistent results using the FEA-SD technique. It was found that a sample of 10 g of feces (divided into tubes A and B) was critical for accurate quantification; as a comparison, 5 g of feces resulted in very inconsistent egg counts between tubes A and B.
The direct microscopic identification of schistosome eggs is the ‘gold’ standard for the diagnosis of zoonotic schistosomiasis in both animals and humans. The current microscopic methods of choice for the identification of S. japonicum eggs in bovines and other ruminants are, however, limited in terms of sensitivity and include, among other procedures, the MHT followed by a sedimentation filtration method, and the Danish Bilharziasis Laboratory (DBL) technique [23] (Table 5). More recently developed techniques include FLOTAC and the use of magnetic beads (Helmintex test) which have been shown to detect helminth eggs in low intensity infections [24]–[29]. Immunological techniques have also been applied to diagnostics however cross reactivity and identification of past infections, rather than current infections, have been issues. A recent study looking at Thioredoxin Peroxidase-1 in an ELISA system for identification of S. japonicum in bovines has shown promising results and no cross reaction with a closely related species [30].
The MHT has been used extensively in the People's Republic of China for the identification of S. japonicum in bovine feces [11]. The MHT – a qualitative diagnostic test – involves the concentration of ova from saline using fresh feces through a nylon tissue bag and suspension in distilled water. Miracidia are visualized macroscopically and their presence is diagnostic of infection; three hatches (50 g feces per hatch) are routinely carried out. The MHT is preferred to the Kato Katz technique (KK) (recommended for diagnosis of intestinal schistosomiasis in humans) [26], [31], [32] and the other microscopic methods, due to the large volume of feces produced by bovines, and the fact that, as discussed, bovine feces contain considerable amounts of cellulosic material. This obscures the microscopic visualization of schistosome eggs making slide reading difficult and hindering diagnosis. A drawback of the MHT is that it has fairly rigid requirements for suitable pH, temperature and water quality, which cannot always be met under field conditions. Furthermore, it does not, on its own, provide infection intensity information and, like the KK, its sensitivity decreases as infection intensity decreases. In order to obtain intensity of infection estimates, additional microscopic visualization of eggs is performed on MHT-positive samples following a filtration sedimentation procedure whereby 50 g of feces are passed through 30 (595 µm) and 150 (90–105 µm) sieves and the flow through suspended in a nylon bag to capture the sediment which is then resuspended and the eggs present counted [33]. This is similar to the DBL technique [21] and the same problem of the presence of cellulosic material is common to both procedures.
The differences in sensitivity of these different techniques for examining ruminant feces for the presence of S. japonicum eggs makes it difficult to compare historical data for prevalence and incidence and infection intensity and to evaluate the involvement of potential reservoir hosts, particularly bovines in schistosomiasis transmission. Table 5 reviews the published studies of diagnostic procedures for the identification of S. japonicum in bovines and the inconsistent data obtained for prevalence and intensity. Telling examples are our recent pilot survey of S. japonicum infection in carabao from Western Samar [22] and the results of another recently published (2010) study on carabao from Leyte, the Philippines, which showed very low S. japonicum prevalence by KK (3.7%), the DBL technique (3.7%) and the MHT (0%) but a high prevalence (51.5%) using qPCR on the same fecal samples [17]. These two studies clearly highlighted the requirement for a more accurate microscopic technique, exemplified by the FEA-SD method, if only to validate diagnosis by qPCR.
The FEA-SD technique, including sieving, sedimentation, centrifugation and digestion, takes approximately 1.5 hours to complete. The length of time taken for subsequent slide reading depends on the skill and experience of the technicians involved, but two well trained and experienced microscopists are able to read one sample in 20 minutes. This procedure is relatively straight forward and only requires a centrifuge. Bovine feces comprise a large mass containing primarily cellulosic fibres and a direct count is the only way to get infection intensity but the debris obscure eggs to a large extent and the FEA-SD is the only currently available technique which clears a large proportion of the debris and renders remaining debris transparent, thereby increasing egg visualisation (Figure 2). Based on cost of reagents only, the FEA-SD technique is far less expensive ($US0.65) to perform than qPCR ($US9.2) although both approaches provide a very similar level of diagnostic accuracy [22]. We are currently using the FEA-SD method to determine the prevalence and intensity of S. japonicum in large animal cohorts as part of extensive epidemiological and surveillance studies we are undertaking in both the People's Republic of China and the Philippines.
In summary, the FEA-SD method is an improved tool that can be used to visualize schistosome eggs and to determine the prevalence and intensity of infection of S. japonicum in bovines. The increased visibility of eggs in the final sediment (Figure 2) compared with the DBL, MHT (+filtration) and KK techniques, makes the FEA-SD an important new technique applicable for epidemiological studies where bovines and other ruminants, such as goats, are potentially important reservoir hosts for S. japonicum [11]–[13], [34], [35]. In addition to S. japonicum the FEA-SD method can also be used to identify and quantify eggs of other helminths, such as Fasciola sp. in naturally infected animals. The FEA-SD also has the benefit of costing less than qPCR, which increases its potential as a surveillance tool for evaluating control programs, including in areas where control has led to the suspected elimination of schistosomiasis japonica.
|
10.1371/journal.pntd.0007437 | Clinical and epidemiological features of paracoccidioidomycosis due to Paracoccidioides lutzii | The fungus Paracoccidioides lutzii was recently included as a new causative species of paracoccidioidomycosis (PCM) and most cases have been reported from Brazil. According to available epidemiological information, P. lutzii is concentrated in the Middle-West region in Brazil, mainly in the state of Mato Grosso. However, clinical and laboratorial data available on patients infected with P. lutzii remain extremely limited.
This work describes the clinical manifestations of 34 patients suffering from PCM caused by P. lutzii, treated along 5 years (2011–2017) at a reference service center for systemic mycoses in Mato Grosso, Brazil. Adult rural workers (men), aged between 28 and 67 predominated. All patients had the chronic form of the disease, and the oral mucosa (n = 19; 55.9%), lymph nodes (n = 23; 67.7%), skin (n = 16; 47.1%) and lung (n = 28; 82.4%) were the most affected sites. Alcohol intake (n = 19; 55.9%) and smoking (n = 29; 85.3%) were frequent habits among the patients. No patient suffered from any other life-threatening disease, such as tuberculosis, cancer or other inflammatory or infectious parasitic diseases. The positivity in culture examination (97.1%) was higher than that found for the direct mycological examination (88.2%). Particularly, one patient presented fungemia at diagnosis, which lead to his death. The time elapsed between the initial symptoms and the initiation of treatment of PCM caused by P. lutzii was 19.7 (31.5) months, with most patients diagnosed 7 months after the symptoms’ onset.
Compared with the classical clinical-epidemiological profile of PCM caused by P. brasiliensis, the results of this descriptive study did not show significant clinical or epidemiological differences that could be attributed to the species P. lutzii. Future studies may confirm or refute the existence of clinical differences between the two fungal species.
| Paracoccidioidomycosis (PCM) is an endemic mycosis in Latin America with high incidence in Brazil. The fungi Paracoccidioides brasiliensis (including genetic groups S1, PS2, PS3 and PS4) and Paracoccidioides lutzii are the etiological agents, but little is known about the clinical manifestations of PCM caused by P. lutzii. Regarding eco-epidemiological aspects, the habitat is believed to be the soil due to the predominance of the disease among rural workers and other individuals who work in contact with the land. Paracoccidioides spp. has been isolated from aerosol samples, armadillos and dog food, but more data are needed to better understand the ecology of this fungus. The Middle-West region of Brazil presents the highest number of cases of P. lutzii infection. It is important to note that this species presents particularities regarding the serological diagnosis in patients. Thus, this study aims to verify possible clinical-epidemiological differences in 34 patients from this geographical region. Our results do not point out significant clinical or epidemiological differences between the two species causing PCM. In Brazil, the Ministry of Health has made an effort to include this disease in the list of compulsory notification diseases in order to implement a health policy aimed at an early detection, diagnosis and treatment.
| Paracoccidioidomycosis (PCM) is the most prevalent deep mycosis in Latin America, being endemic only in Brazil, Colombia and Venezuela. In Brazil, the state of Mato Grosso (Middle-West region), has a large number of cases, and the recently described new species P. lutzii [1] was recovered from the clinical isolates of patients from this geographic location. Paracoccidioides lutzii and P. brasiliensis are thermal dimorphic fungi, which grow at room temperature as mycelia and as yeasts with bipolar or multipolar buds at a temperature of 35 to 37°C (parasitic form). Estimates of annual incidence in Brazil vary from 0.71 cases to 3.70 cases per 100 thousand inhabitants [2]. According to information from the Ministry of Health, 3,181 cases of PCM deaths were recorded in Brazil between 1980 and 1995, resulting in a PCM mortality rate of 1.45 cases per million inhabitants (2.59 for the Southern region, 2.35 for the Central-West region, 1.81 for the Southeast, 1.08 for the North, and 0.20 for the Northeast) [3]. In Brazil, PCM is the 8th cause of mortality among the parasitic infectious diseases. Even so, it is still included in the group of neglected diseases, and there is no requirement for compulsory notification despite the severity of the disease and the fact that it is considered a public health problem [4]. The incidence of hospital admissions for PCM in Brazil is 7.99/1000 inhabitants, surpassing other endemic mycosis such as histoplasmosis and coccidioidomycosis [5].
The state of Mato Grosso is known as the country's granary, being the largest producer of soy, corn, cotton together with cattle breeding. This productivity is achieved due to the intense modernization of farming techniques. For this reason, most of the cases of patients affected by PCM are directly related to the agricultural activities carried out in rural properties of different territorial extensions. On the other hand, agricultural machine operators also constitute a target audience for PCM acquisition.
Recently, 65 isolates of P. brasiliensis were analyzed through nuclear and mitochondrial DNA, as well as the morphology of conidia and yeasts; in this study, the authors propose a new classification for the P. brasiliensis complex and the taxonomic recognition of the four genetic groups as P. brasiliensis (S1), P. americana (PS2) P. restrepiensis (PS3), P. venezuelensis (PS4), suggesting that they be considered as distinct species [6].
Humans and the nine-banded armadillo (Dasypus novemcinctus) are the accidental hosts of Paracoccidioides spp. and are usually infected in rural and peri-urban environments. Despite the consensus that the fungus’ habitat is the soil, few studies were able to demonstrate the isolation from this micro niche, existing many gaps concerning the knowledge on the still unresolved eco-epidemiology of PCM. Recently, P. brasiliensis and P. lutzii were detected in soil samples from three different locations in Brazil using molecular methods [7]; nevertheless little is known about the pathogenicity, virulence of strains, and more detailed aspects relating to the eco-epidemiology of the new species P. lutzii. In 2018, Hrycyk et al. [8] confirmed that while armadillos are highly infected by P. brasiliensis, including multiple infections by distinct genotypes or species (P. brasiliensis and P. americana) in the same animal, the same does not hold true for P. lutzii, which in turn seems to present less capacity for mycelial growth and conidial production, when developing in a soil-related condition, but this deserves further investigation.
Respiratory infection occurs via inhalation of conidia present in nature, which later reach the pulmonary alveoli. Usually, the infection is controlled by the cellular immune response, but scars can remain with latency of yeast cells. Thus, there is usually asymptomatic infection or nonspecific symptoms, or even some individuals showing the progression of infection to disease [4].
When the disease develops, the classical clinical forms are known as acute or subacute ("juvenile"), prevalent in children and young adults, in which there is inadequate Th2 cell type response to control the fungal infection. The chronic form represents 80 to 95% of the cases, affects individuals in the productive age (after the third decade of life), usually affecting the lungs, upper region including lesions in the oral mucosa, nasal mucosa, skin in places adjacent to the mouth and nose, and cervical lymph nodes. The incubation period of the disease is uncertain and may develop after many years after the individual's initial contact with the fungus [2, 4].
Paracoccidioides brasiliensis is composed of a cluster of molecular siblings recognized as S1 (S1a and S1b), PS2, PS3, and PS4 [9, 10]. The phylogenetic species S1a and S1b are widespread and predominantly found in lower South America, especially in the southeast and South of Brazil, Argentina, and Paraguay [10]. PS2 has a sporadic distribution and has been less frequently reported, with human cases only being reported thus far in Venezuela and the southeast of Brazil. The PS3 and PS4 species are, to date, exclusively endemic to Colombia and Venezuela, respectively [11].
Phylogenetic analyses demonstrated that P. lutzii represents a highly divergent lineage monophyletically separated from P. brasiliensis. Paracoccidioides lutzii is often found in the Middle-West region [1] and North [12] of Brazil, and most of the genetically evaluated clinical isolates were from the state of Mato Grosso. Regarding morphology, conidia of P. lutzii are elongated (2–22 μm), while that of P. brasiliensis measure from 2 up to 5 μm [13, 14]. To date, the main difference related to P. brasiliensis and P. lutzii lies in the serological diagnosis, where there is a need to employ local antigenic preparations in serological techniques such as ELISA, immunodiffusion and latex [15, 16].
The taxonomic description of a new species has raised the curiosity of physicians due to the possible clinical implications. Furthermore, characteristics of the in vivo susceptibility of P. lutzii to drugs conventionally used in the history of PCM also raise the interest of the professionals that manage patients affected by PCM.
The objective of this work was to describe the first results concerning the epidemiological and clinical characteristics of patients affected by P. lutzii from the Middle-West (Mato Grosso) of Brazil and reflect on possible similarities or differences between these characteristics and the classical profile of the disease caused by P. brasiliensis.
This study was submitted to and approved (CAAE: 17177613.6.0000.5541) by the Federal University of Mato Grosso (UFMT) and protocol number 1796–10 by the Federal University of São Paulo (UNIFESP). All adult subjects provided informed written consent and the study was approved by ethical committee under number 288.250/CEP/HUJM/UFMT.
A descriptive study was carried out on 34 confirmed PCM cases caused by P. lutzii (Fig 1), that is, those with compatible clinical manifestations and positive fungal culture for Paracoccidioides spp. from different clinical materials and which were later confirmed by genotyping as P. lutzii. The patients in the study were enrolled at a reference service center of systemic mycoses of the Júlio Muller University Hospital–Federal University of Mato Grosso (UFMT / HUJM), Cuiabá, Mato Grosso—Central-West region of Brazil.
The Paracoccidioides spp. isolates were obtained from various clinical material (sputum, cervical lymph aspiration, blood, oral mucosa scraping, scraping of the larynx, scraping of the nasal mucosa, fragment of skin biopsy). The clinical materials were cultivated in Sabouraud Dextrose Agar (DIFCO) and incubated at a temperature of 35º C in a BOD incubator (Eletrolab) for a period of up to 20 days. Colonies with cerebriform appearance and creamy color, typical of the yeast forms, were isolated with subsequent confirmation of micromorphological characteristics of Paracoccidioides spp.
Isolates morphologically identified as Paracoccidioides spp. were subjected to molecular characterization using either HSP70 amplification [1] or via TUB1-RFLP [17]. DNA was extracted and purified from fungal colonies with the Fast DNA kit protocol (MP Biomedicals). The primer pair HSPMMT1 (5’-AAC CAA CCC CCT CTG TCT TG-3’) and PLMMT1 (5’-GAA ATG GGT GGC AGT ATG GG-3’) targeting an exclusive indel region of P. lutzii were used for PCR [1]. Isolates Pb01 and B339 were used as controls of P. lutzii (positive) and P. brasiliensis (negative) respectively. In addition, for TUB1-RFLP, the protocol described by Roberto et al. [17] was used. TUB1 fragments were amplified using the primer pair α-TubF (5′-CTG GGA GGT ATG ATA ACA CTG C-3′) and α-TubR (5′-CGT CGG GCT ATT CAG ATT TAA G-3′) [18] following a double digestion with BclI and MspI restriction endonucleases. The reaction contained 13 μL H2O, 3 μL TUB1-PCR product, 2 μL 10× fast digest buffer, and 1 μL each of the BclI (10 U/μL; Thermo Scientific) and MspI (10 U/μL; Thermo Scientific) restriction endonucleases. The digestion mixture was incubated at 37°C for 2 hours. The digested products were electrophoresed on 2.5% (w/v) agarose gels for 120 min at 100V in the presence of GelRedTM (Biotium, USA). We included a lane loaded with 50bp DNA Step Ladder (Promega, USA). Molecular identification was performed at the Medical and Molecular Mycology Laboratory (UNIFESP/EPM). The bands generated by PCR or TUB1-RFLP were visualized using the L-Pix Touch (Loccus Biotecnologia, São Paulo, Brazil) imaging system under UV illumination.
Epidemiological and clinical data were collected from medical records of P. lutzii PCM treated patients between 2011 and 2017. The categorical variables were summarized by percentages and 95% confidence interval, and the numeric variables by mean and standard deviations. All analyses were performed by Stata Statistical Software version 12.0 (College Station, Texas, USA).
Altogether 34 patients with confirmed diagnosis of PCM were evaluated, 33 men (n = 33; 97.1%) and only one woman (2.9%), with a mean (SD) age of 46.7 (9.3) years of age. Most of the patients (75.7%) resided in the north and central regions of the state of Mato Grosso (Fig 1); 73.5% in rural areas and 26.5% in urban areas. The occupations of farmer (53.6%) and rural truck driver (32.1%) were the most frequent. Smoking (85.3%) and alcohol intake (55.9%) were very frequent among patients (Table 1). None of them suffered from other life-threatening diseases.
The species P. lutzii was identified by TUB1-RFLP in all 34 patients described, 30 (88.2%) being new cases and 4 (11.8%) relapsed cases of the disease. The PCM in this series of cases was multifocal in 88.2% (n = 30) and unifocal in 11.8% (n = 4). All patients had the chronic clinical form of the disease, with pulmonary involvement in 82.4%, lymph nodes (Fig 2A and 2B) in 67.7%, oral (Fig 2C) in 55.9%, cutaneous in 47.1%, laryngeal in 32.8%, nasal in 11.8%, bone (Fig 2D) in 11.8% and 2.9% in adrenal glands. One of these patients had symptoms of fungemia by P. lutzii. No patient presented central nervous system or genital involvement. The average time (SD) elapsed between the initial symptoms and the initiation of treatment of PCM by P. lutzii was 19.7 (31.5) months, with most patients diagnosed 7 months after the symptoms’ onset (Table 2).
The diagnosis of PCM was initially confirmed by culture in 97.1% (n = 33) of cases, direct mycological examination (DME) in 88.2% (n = 30), histopathological examination in 35.3% (n = 12). Clinical specimens used for the mycological exams were ganglionic secretion (n = 14), scraped oral mucosa lesion (n = 13), sputum samples (n = 3), skin biopsy (n = 3) and blood (n = 1) (Table 1).
The decision on the treatment of patients was based on the II Brazilian Consensus of Paracoccidioidomycosis [4], using sulfamethoxazole + trimethoprim in 88.2% of the patients. Out of these, 7 (23.3%) also used itraconazole and another 2 (6.9%) amphotericin B deoxycholate. The initial hematological and biochemical evaluation of the patients did not present any relevant changes (Table 2).
In the present study on 34 patients with confirmed infection by P. lutzii, there were no clinical or epidemiological differences that could be attributed to the P. lutzii species.
The epidemiological profile of PCM has been revealing remarkable changes in frequency, demographic characteristics and geographical distribution. More than a decade ago studies published by our research group showed differences between isolates of Paracoccidioides spp. Initial investigations were conducted looking for correlations between clinical forms of the disease, geographical origin of same, susceptibility to antifungal drugs and epidemiological findings [19, 20]. In 2009, Batista et al. [21], showed significant differences in serological test results using double radial immunodiffusion for diagnosis of PCM when sera from patients from the Middle-West and Southeast regions of Brazil were evaluated. The exoantigens obtained from isolates from patients from these geographical regions affected by PCM presented strong evidence of antigenic variation among the isolates [15, 22]. It was also observed through the RAPD technique that clinical isolates from different anatomical sites (arm and face) of a same patient presented genetic differences [23]. All evidence collected related to possible antigenic differences whenever exoantigens from different geographic locations [21] were used by different researchers who obtained results from the use of different molecular techniques seeking correlation with virulence of isolates [24] and clinical forms of the disease [20], was important for the proposal of a new species: P. lutzii [1].
However, the vast literature related to clinical, demographic and epidemiological aspects of P. brasiliensis as a single etiologic agent of the disease until 2009, highlights classical presentations fairly known by medical professionals. For the physician, it is important to assess the epidemiological, clinical, diagnostic and therapeutic impact on the disease of different species of Paracoccidioides, i.e., whether there are indeed differences regarding clinical manifestations between the two species: P. lutzii and P. brasiliensis, possibly attributed to the antigenic differences of clinical isolates [15], or even to the virulence of the strains [24, 25].
Taking into account the acute/subacute forms according to Ferreira [26], a multisystemic involvement of the disease is observed; the presence of lymphadenomegaly, cutaneous lesions, hepatosplenomegaly or abdominal masses. Jaundice, ascites, and peripheral edema may also be present. The latter justify the investigation of hypoalbuminemia. Signs of adrenal involvement, as well as neurological involvement, are rare in this clinical form. Digestive complaints, such as abdominal pain, chronic malabsorptive diarrhea and vomiting, are also quite frequent. Fever and weight loss complete the clinical picture, presence of growth or pain in the bone region requires the identification of bone lesions. According to Mendes [27] and Valle et al. [28], the chronic form is assessed through signs and symptoms related to the pulmonary, tegumentary and laryngeal involvement (cough, dyspnea, mucopurulent expectoration, ulcerated skin lesions and nasopharyngeal mucosa, odynophagia, dysphagia and dysphonia); lymphatic (adenomegaly); adrenal [29, 30] (asthenia, weight loss, hypotension, darkening of skin, abdominal pain). Relating to the central nervous system, according to Pereira et al. [31] and Almeida et al. [32] the following may be observed: headache, motor deficit, convulsive syndrome, changes in behavior and/or level of consciousness. Regarding the digestive impairment, diarrhea and sometimes malabsorption syndrome are reported [33]. In this study, all patients evaluated presented the chronic form of the disease, where pathognomonic signs and symptoms of this form were recognized, mainly showing pulmonary, lymphatic, oral and cutaneous impairment. There was no clinical evidence in this sample of patients evaluated (n = 34) that could be highlighted, considering the etiology of PCM caused by P. lutzii. One case of fungemia was observed [34], but it is not possible to infer that P. lutzii is more virulent than P. brasiliensis because of this finding. Moreover, out of the 34 cases evaluated with etiology of PCM by P. lutzii, only two were classified as severe chronic form, the majority (n = 32) being classified as moderate clinical form.
Considering the proposed species (P. brasiliensis S1a, S1b, PS2, PS3, PS4 and P. lutzii), Macedo et al. [35] described an autochthonous clinical case in the southeast of Brazil (Rio de Janeiro), classified as P. brasiliensis PS2. These authors reported that few cases with this molecular taxonomy have been recorded in the literature when compared with S1 and PS3, and that among the cases registered pointing PS2 as the etiologic agent a higher frequency of the chronic form of the disease was observed. This finding was also observed in 34 patients affected by PCM caused by P. lutzii assessed in this study. Associated habits (smoking and drinking) were also frequent, as well as the frequency in male individuals in the productive age. These characteristics coincide with those described in the literature for classical PCM caused by P. brasiliensis (smoking (>20 cigarettes/day for >20 years) and alcohol intake (>50g/day). They are also often associated with the mycosis) [36]. Regarding the duration of symptoms in months for patients affected by P. lutzii in this series of cases, two groups were the most frequent: 13 patients allocated in the range of 1 to 6 months, and 12 patients ranged higher than 12 months, corroborating the classical data already published for P. brasiliensis.
In terms of distribution by regions in the state of Mato Grosso, the North (n = 14) and Central South (n = 11) regions were responsible for the largest number of cases. The concentration of the highest number of cases in the Northern region can be explained by environmental factors due to the opening of new agricultural frontiers with forest felling, especially in the Amazon—Mato Grosso region [37]. In addition, the occurrence of different species of Paracoccidioides may also be contributing to the change in the epidemiological pattern [37].
The suspected diagnosis of PCM occurs through clinical and epidemiological data, but the confirmation is done primarily by the identification of the etiologic agent in fresh tissue examinations, cultures and histopathologic preparations, which are considered the gold standard in the definition of the disease, being known as direct techniques in the diagnosis of PCM. Indirect techniques are represented by the presence of antibodies and circulating antigens in the serum of patients with PCM. A very interesting result was found in this study, with higher positivity for the culture identification (97.1%) when compared with that found by direct mycological examination (88.2%).
Generally speaking, it is not possible, so far, to establish important clinical differences that can be attributed to P. lutzii or P. brasiliensis complex. In 2017, our research group evaluated, in another study, a total of 554 patients who were treated at the same hospital during the study period (1998 to 2014), 527 had confirmed PCM diagnosis. Out of 527 patients, 244 (46.3%) patients (mean age, 48.4 [10.9] years; range, 14–83 years), classified as the chronic form of PCM. All patients were living in rural areas, and most performed activities related to agriculture [38]. These data show that the acute form of PCM is less frequent in the state of Mato Grosso, central region of Brazil, a geographical region where a higher frequency of P. lutzii has been observed so far.
This is the first study that presents a series of cases of P. lutzii, identified by molecular methods and correlating them with the clinical and epidemiological profile of affected patients. The actual incidence of each phylogenetic species and its involvement in clinical practice should include other studies in different regions of Brazil and Latin America to compare the forms of PCM and clinical manifestations with the genetic profile of these entities.
Only a few studies are found to date in the literature offering the molecular identification of clinical isolates and their association with clinical characteristics of patients affected by PCM. For comparison purposes, considering clinical findings and molecular characterization, Macedo et al. [39] recently carried out phylogenetic analysis of 54 Paracoccidioides spp. clinical strains from Rio de Janeiro, Brazil where P. brasiliensis (n = 48) and P. americana (n = 6) were identified as the causative agents of PCM. Considering the clinical classification, the authors reported that 41 strains were identified as P. brasiliensis, 23 corresponded to the chronic form, and 16 were acute. In Mato Grosso, all 34 clinical cases infected by P. lutzii corresponded to the chronic form of PCM. In relation to the affected organs, for both P. lutzii (Table 2) and P. brasiliensis [39], lungs and lymph nodes were the most affected. Regarding the severity of the disease, 7 were classified as mild, 18 (moderate), and 16 (severe) in the case of P. brasiliensis [39], in contrast to 32 cases (moderate form) PCM caused by P. lutzii, and only 2 cases were classified as severe.
The number of clinical cases evaluated by Macedo et al [39] concerning P. americana is very small. Thus, it is difficult to make any inference or comparison considering clinical manifestations. This has proven to be a limitation for the study [39].
Based on the clinical findings regarding P. lutzii and P. brasiliensis complex, there is no evidence that allows us to point out significant clinical differences between species. For this reason, we believe that studies with a greater number of isolates should be conducted to confirm or refute the hypothesis that there are clinical differences related to the different species. However, the genetic susceptibility of the host should always be an important parameter to be considered, as well as the virulence of the strain, regardless of the species that causes PCM.
|
10.1371/journal.pntd.0002742 | A Mixed Methods Study of a Health Worker Training Intervention to Increase Syndromic Referral for Gambiense Human African Trypanosomiasis in South Sudan | Active screening by mobile teams is considered the most effective method for detecting gambiense-type human African trypanosomiasis (HAT) but constrained funding in many post-conflict countries limits this approach. Non-specialist health care workers (HCWs) in peripheral health facilities could be trained to identify potential cases for testing based on symptoms. We tested a training intervention for HCWs in peripheral facilities in Nimule, South Sudan to increase knowledge of HAT symptomatology and the rate of syndromic referrals to a central screening and treatment centre.
We trained 108 HCWs from 61/74 of the public, private and military peripheral health facilities in the county during six one-day workshops and assessed behaviour change using quantitative and qualitative methods. In four months prior to training, only 2/562 people passively screened for HAT were referred from a peripheral HCW (0 cases detected) compared to 13/352 (2 cases detected) in the four months after, a 6.5-fold increase in the referral rate observed by the hospital. Modest increases in absolute referrals received, however, concealed higher levels of referral activity in the periphery. HCWs in 71.4% of facilities followed-up had made referrals, incorporating new and pre-existing ideas about HAT case detection into referral practice. HCW knowledge scores of HAT symptoms improved across all demographic sub-groups. Of 71 HAT referrals made, two-thirds were from new referrers. Only 11 patients completed the referral, largely because of difficulties patients in remote areas faced accessing transportation.
The training increased knowledge and this led to more widespread appropriate HAT referrals from a low base. Many referrals were not completed, however. Increasing access to screening and/or diagnostic tests in the periphery will be needed for greater impact on case-detection in this context. These data suggest it may be possible for peripheral HCWs to target the use of rapid diagnostic tests for HAT.
| Human African trypanosomiasis (HAT or sleeping sickness) is a fatal but treatable disease affecting poor people in sub-Saharan Africa. Most HAT diagnostic equipment, infrastructure and expertise is located in hospitals. The expense of expanding testing services to remote areas using mobile teams severely restricts their use. Non-specialist healthcare workers (HCWs) in first-line (primary) health care facilities can contribute to control by identifying patients in need of testing based on their symptoms. We therefore trained first-line HCWs to recognise potential syndromic cases of HAT and refer them to a hospital screening service. Against a low baseline of HCW HAT referral experience, four months after the intervention, HCW knowledge of HAT symptoms increased and HCWs in 71.4% of facilities across the county had made referrals, incorporating new and pre-existing ideas about HAT case detection into referral practice. There was only a modest increase in numbers of referred patients received at the hospital for screening, however, largely because of distance. In an era where approaches to HAT case detection and control must increasingly be integrated into health referral systems, it is vital to understand the opportunities and challenges associated with syndromic case detection in first line facilities to design effective interventions.
| Found in remote sub-Saharan areas where health systems are often weak and/or destabilised by armed conflict, human African trypanosomiasis (HAT, or sleeping sickness) is one of the world's most neglected tropical diseases (NTDs). It is caused by infection with trypanosome parasites that are transmitted primarily by tsetse flies (Glossina) and is nearly always fatal if untreated. HAT caused by Trypanosoma brucei gambiense represents more than 90% of global HAT burden and is endemic in geographically limited foci in west and central Africa [1]. In these areas, humans are assumed to be the main reservoir of infection.
HAT can be asymptomatic or involve non-specific symptoms in the first stage of disease. Characteristic symptoms involving mental and physical deterioration progressing to death are more likely to appear in the second stage, once parasites have entered the brain [2], [3]. The natural duration of gambiense HAT is thought to be a median of almost a year and a half for each stage [4].
Systematic active screening (AS) and treatment of at-risk populations using laboratory-equipped mobile teams identifies both symptomatic and non-symptomatic cases and has been a key method used to control recent large epidemics in central Africa [1], [5]. It is also, however, resource-intensive and, for this reason, has not been sustained in many areas as caseloads have decreased [5]–[8], despite calls for an intensification of control activities to achieve elimination [9]. This has been especially problematic in South Sudan where, since the signing of the north-south peace agreement in 2005, the country has seen a critical decrease in funding for HAT control activities as international non-governmental organisations have disengaged from HAT service delivery [7]. Approaches to screening which are sustainable at low incidence rates and low funding rates will need to be used to maintain case detection and control activities in this type of context. Building these around existing health services is likely to be the most pragmatic and sustainable approach.
One option that has been tried in the past is to split the work involved in active case-finding into parts, using community-based healthcare workers (HCWs) to collect blood samples on foot for subsequent screening at hospital laboratories, where most HAT diagnostic equipment and expertise is located [10]–[13]. Newly-introduced rapid diagnostic tests (RDTs) for HAT could similarly be used [14]. Fewer resources can then theoretically be spent on following-up a smaller number of people for confirmatory testing.
Another option is systematically to screen populations when they come into contact with health facilities where screening resources are already available [13], [15]–[17], either comprehensively or under programmes targeting specific populations as has been suggested for antenatal care [18]. Targeting of children at routine child health visits or via school health programmes is not typically prioritised since adults are more likely to be infected [19]. Currently, neither of these approaches to case-detection is frequently used.
A third cost-effective alternative is to target screening resources to symptomatic patients who have a higher probability of being HAT cases than the general population. This is the principle behind referral-based passive screening (PS), the least resource-intensive and most commonly used alternative to AS. Under this approach, patients may be referred for HAT screening by a HCW or by a lay-person (family member, neighbour, church leader, etc) knowledgeable about HAT symptoms and HAT test availability (Figure 1). Recognition of symptoms to prompt HAT referral, however, is commonly thought to be delayed by lengthy and/or complicated preceding treatment-seeking events including syndromic misdiagnosis of malaria, typhoid or HIV [20]–[24]. The specificity of RDT- and microscopy-based diagnosis for malaria, HIV and HAT are also each lowered in the presence of co-infections [25]–[27]. In a context of extreme material poverty, as in most highly HAT-endemic areas, transportation, opportunity and direct costs of recurrent treatment-seeking may severely diminish patient motivation to keep seeking treatment or complete referrals [22], [28], [29]. HAT referral completion may also potentially be influenced by patient perceptions of severity and treatability of symptoms [22], knowledge of treatment requirements or associated cultural prohibitions [30], [31], or HAT-related stigma [30].
While PS is often the approach responsible for detecting the majority of HAT cases over the length of a program [21], [22], [32]–[35], providing much of the information on incidence used to quantify the burden of HAT locally and globally, surprisingly little attention is paid to optimising its use. Syndromic algorithms that can be used to guide HCW referral have recently been proposed by us [36] but most hospital programmes employ no systematic approach to identify potential cases. Additionally, outside of a few outbreak situations when AS is not feasible [37], [38], typically little emphasis is placed on education or sensitisation of HCWs about the signs and symptoms of HAT which should trigger referral. Furthermore, the contribution that HCWs in peripheral facilities could make to syndromic case detection is habitually over-looked in vertically-organised HAT control programmes [37], [39].
One notable exception to this trend was a five year programme in Niaki focus, Republic of Congo, in the 1980s, which trained HCWs in all first-level public facilities to screen syndromically and detect cases via basic microscopy [15]. Patients could also be syndromically referred for serological and full confirmatory screening at a central hospital. This strategy appears to have greatly improved participation of peripheral facilities in case-detection. By its final year, these facilities were diagnosing more new cases from rural areas than the referral hospital or mobile teams (and responsible for >30% of the total). A simpler integrated case detection strategy based on syndromic detection in peripheral facilities has recently been suggested for the Democratic Republic of Congo [22]. Under it, referrals could be made to a central facility or specialised supervisors could perform screening tests on visits. The deployment of RDTs for HAT to first or second-line facilities in a passive screening approach could also incorporate syndromic referral rationale to conserve test resources. Others have called for specific syndromic training on HAT for peripheral HCWs [20], [24], [37] but, to our knowledge, no such intervention has been conducted and systematically evaluated.
In this study, we present the results of a syndromic HAT training intervention which targeted HCWs in peripheral facilities in the Nimule focus of South Sudan with the objective of increasing the rate of syndromic referrals to a central screening and treatment centre.
This study was approved by the London School of Hygiene & Tropical Medicine's ethical review committee and was registered with the Ministry of Health, Government (now Republic) of South Sudan. All HCW participants provided written informed consent and patients provided verbal informed consent. Verbal consent was approved by the LSHTM committee for use with patients because of low literacy rates in the study area. Receipt of consent was documented on data collection forms.
In South Sudan, training on HAT is offered through the national control programme by World Health Organisation staff but is currently targeted to hospital-based inpatient HCWs on the clinical management of patients and to laboratory staff on diagnostic protocols. HCWs of all levels who have received formal training in South Sudan are briefly introduced to HAT symptomatology, diagnosis and treatment in curriculum. Many informally trained HCWs who are nevertheless popular providers of patient care in this context have never received any training on syndromic HAT recognition.
HAT screening and diagnosis is currently only available at six hospitals in the country [7]. In the Nimule focus of Magwi County (MC), Eastern Equatoria State, HAT testing and treatment services are available at a single site, Nimule Hospital, supported by the non-governmental organisation, Merlin (Medical Emergency Relief International) [7]. Services in this historic focus were re-introduced at the end of the Sudanese civil war, in 2005. Transmission is thought to have increased in recent years due to population movements of internally-displaced and returning refugee populations from neighbouring endemic areas. Small-scale (<30% population coverage) AS surveys conducted in 2005, 2006 and 2008 revealed a low-to-moderate estimated HAT prevalence of around 1% with the highest prevalence in any village estimated at 5.8% [40]. No surveys have been conducted in 2/6 payams (districts) furthest from the hospital in the east of the county but cases have been identified from these areas via passive screening.
Preliminary formative research by us suggested a need to improve knowledge of HAT symptoms and screening service availability among Nimule area peripheral HCWs as a necessary (although possibly not sufficient) step to increase screening [40]. Excessive sleeping, mental confusion, ‘swollen’ body parts due to oedema or weight gain, pain, fever and enlarged lymph nodes appeared to be the key symptoms which signalled HAT to referring HCWs. Along with these symptoms, excluding common diagnoses such as malaria and typhoid was perceived by HCWs as a valuable strategy which helped them identify likely HAT cases. Few HCWs appeared routinely to make HAT referrals, however, either because making HAT diagnoses was perceived as something that should be done by hospital-based mobile teams or because they did not recognise the local magnitude of the problem (and hence would not consider HAT as a possible diagnosis) without the cue of a large community campaign.
HCWs from peripheral healthcare facilities were trained to recognise potential syndromic cases of HAT during routine outpatient practice and refer them for screening to Nimule hospital. The rate of patient referrals received at the hospital from a peripheral healthcare facility for HAT screening before versus after the intervention was the primary measure of intervention effectiveness. Secondary (intermediate) measures were changes in the number and distribution of HAT referrals made by trained HCWs in the county and changes in HCW knowledge of HAT symptoms.
Based on interviews with peripheral HCWs about their existing HAT referral practices before the intervention, we theorised that the training intervention, if effective, could increase the rate of HAT referrals made in two main ways: by expanding the range of symptoms that prompt HCWs to make HAT referrals or by encouraging HCWs to think of HAT as an appropriate diagnosis to make at the peripheral level, or both (Figure 1). Common patient case-finding techniques and clinical reasons for referrals after the training were therefore identified qualitatively with HCWs to explore which type of syndromic knowledge best explained changes in referral volumes. The appropriateness of referrals made by peripheral HCWs was also appraised a posteriori by a panel of four clinicians with experience of HAT, based on reported signs and symptoms.
Finally, barriers that HCWs faced to support completion of referrals by patients were also explored.
Hospital-based data collection on pre- and post-intervention referrals took place over the 10 month period, July 2009–April 2010. HCW HAT training took place during the month of November 2009. Facility-level referral data were collected during a four-week evaluation period in March 2010.
This training intervention was designed according to three principles. First, it should achieve high coverage of HCWs in all types of provider setting from which patients with sleeping sickness might seek care. Second, it should be realistic in terms of what could be delivered by government or non-governmental organisation programmes at scale. Third, it should specifically recognise existing HAT referral strategies and discuss local barriers to syndromic case detection as identified from preliminary qualitative research ([40] and others).
The training was delivered through a series of six one-day workshops in each payam of the county. Invitations were distributed to facilities via local authorities and health organizations one month before the training. During briefing visits, local authorities were asked to identify the location of all public and military health facilities in the payam as well as any private facilities and traditional herbalists or witch doctors (henceforth referred to as traditional practitioners) they were aware of. Additional private facilities were added to the invitation list by driving or walking around payams and noting health structures. Chiefs and HCWs were also interviewed to identify popular local traditional practitioners. Facilities were invited to send as many HCW-diagnosticians to the workshops as could be spared without disrupting essential services. HCWs from small drug shops were invited if they reported that they examined and diagnosed patients on their premises. Nimule hospital HCWs were excluded from the HAT training workshops unless they also operated peripheral private facilities; outpatient staff had already been sensitised to HAT symptomatology during HAT patient management training earlier in the year. Participants were offered no training incentives other than lunch, a training manual and a certificate of attendance.
The training workshops were conducted by the Nimule hospital HAT programme manager (ES) and an expatriate HAT researcher (JP) in English and either of the two local languages, Madi or Acholi. Arabic simultaneous translations were also offered as needed. Training was didactic as well as participative and covered several topics: HAT transmission, distribution, control, clinical signs & symptoms, diagnosis and treatment, as well as a discussion on taking patient histories, recording referrals and counselling patients to complete referrals. No specific syndromic referral algorithm was introduced.
Key messages discussed about local barriers to syndromic HAT case-detection included: the problematic expectation among HCWs and lay people that HAT patients present more often with an increased rather than decreased appetite, the complication that frequent drunkenness adds to identifying HAT symptoms, the particular difficulty that imprisonment for mental symptoms poses for soldiers to seek treatment, and management of differential biomedical diagnoses that can be confused with HAT (especially malaria; magical poisoning and witchcraft were not discussed). The problematic history of local HAT programme funding was presented with the aim of empowering HCWs to play a greater role in case-detection outside of hospital-directed HAT activities (see File S1 for training manual). Workshops finished with a group memory game in which participants were asked to guess which HAT symptoms were being acted out by other participants.
We trained 108 HCWs from 61 public, private and military health facilities, as well as 12 traditional practitioners and 3 local government health authorities (123 HCWs total, Table 1). Participants came from almost all (97.4%) public facilities in the county, all military facilities and from more than half (62.5%) of private facilities (Table 2).
Participants were predominantly male and from the two indigenous tribes of the county: Madi and Acholi (Table 1). 13% of HCWs came from the payam where the hospital is located (Nimule), 29.2% from neighbouring payams (Pageri and Mugali) and 57.7% from payams distant from the hospital. Most HCWs were recently returned refugees (median number of years worked in the county: 3, range 0–42, 55% of participants >30 years of age) and had a low level of formal clinical education. Clinical officers (CO), who receive 3–4 years of clinical training, certificated nurses who receive three and enrolled nurses who receive two, made up only 13.9% of participants (training category A). Community health workers (CHWs) who receive nine months of clinical training made up 41.5% (category B) and participants working in health facilities with no formal clinical training (mostly nursing assistants trained on the job; category C) made-up the largest proportion (44.7%). HCWs who attended HAT training collectively estimated to see between 1400–1900 patients/day, equivalent to approximately 1% of the county population. As a group, CHWs engaged in the majority of patient consultations (462–550/day), followed by informally trained HCWs (378–450/day) and COs/nurses (189–225/day). Public health facilities were responsible for more consultations (861–1025/day) than private facilities (25–125/day), military facilities (25–75/day) and traditional practitioners (22–30/day); however, the under-representation of private facilities and traditional practitioners in this training workshop should be noted.
Private clinics were most likely to have any testing service available (78.6% vs 58.8% in public and 66.7% in military facilities, data not shown) and public facilities most likely to have only malaria rapid-diagnostic tests (RDTs) available. One facility (a private clinic) reported having the capacity to diagnose HAT microscopically. Stock-out of tests, reagents and equipment were, however, a problem in all but 12 facilities, and stock-out of drugs was a common reported reason why HCWs made referrals. Infectious diseases were the most common reason (34.6%) for the last general referral participants made. Participants collectively ranked a list of common illnesses/conditions for which treatment is commonly sought by patients in MC according to how common they are in the county, in the following order: malaria, drunkenness, syphilis, typhoid, HIV, mental illness, tuberculosis and magical poisoning, with HAT ranked last (only 9.9% rated HAT as “very common”; 63.1% rated HAT as “not common”).
This study evaluated the effect that a simple intervention, training of peripheral HCWs to recognise and refer potential syndromic cases of HAT, could have on improving PS rates in rural patient populations in MC. We evaluated this intervention in several ways which suggest that it had a positive effect on HCW referral practices, but a quantitatively marginal effect on passive case-detection largely because of failure to complete referrals by patients.
Against a low baseline of HCW HAT referral experience (only a quarter of HCWs trained in a third of facilities had ever made a HAT referral before), the intervention appeared to have increased the numbers of peripheral HCWs and their distribution in facilities across the county who might now regularly consider HAT in their differential diagnoses of patients. HCWs in almost three-quarters of facilities surveyed in the county had made at least one HAT referral in the four months since training. Two-thirds of the 71 referrals made after the intervention came from new referrers. This increase was especially notable in HCWs working in areas most distant from the screening facility who probably had less exposure to the HAT programme's active and passive screening activities.
In tests, HCW knowledge of the signs and symptoms associated with HAT disease was significantly improved across all demographic sub-groups of participants such that, after the intervention, any differences in HAT symptom knowledge before the training intervention became non-significant. This improvement in knowledge suggested that there was need for HAT training on symptom recognition in all groups of HCWs in the county and that the HAT training package was delivered at a level that was accessible to all participants, regardless of their level of formal clinical training and experience with the disease. Knowledge scores for HAT symptom associations conceptualised in preliminary research as ‘pre-existing’ in HCW HAT case detection narratives were generally higher than for associations conceptualised as ‘new’ or introduced by the training but improvements in both types of knowledge indicated both consolidated learning and expanded knowledge across a range of HAT symptoms.
When syndromic knowledge was translated into referral practice, the referrals made after the intervention appeared to be syndromically appropriate, based on expert review. Qualitative interviews suggested that in these, HCWs used both pre-existing and new ideas about how to identify likely cases of HAT syndromically. Numerically, pre-existing ideas tended to predominate. Considering the relative frequency with which particular signs and symptoms might be expected to occur in a typical, mainly adult patient population, less use of some of the ‘new’ ideas about HAT introduced in workshops might naturally be expected (e.g., convulsions and neurological problems vs. mental confusion and headaches). Fertility problems, however, were probably an example of a new concept that was under-utilised by HCWs in referrals; furthermore, almost 2/5 HCWs tested in the evaluation period did not correctly associate this symptom with HAT, suggesting need for further exploration and discussion of the significance of HAT-fertility concepts with HCWs.
When HCWs were approached by patients at facilities or after community meetings, differentially ruling out malaria/typhoid was a very common but complex HAT referral behaviour. Often, HCWs' desire to resolve failed malaria/typhoid diagnoses appeared to be the driving reason behind HAT referral, prompting HCWs to specifically look for HAT-supportive signs such as enlarged lymph nodes. In severe cases, malaria/typhoid was always ruled out but we could not determine in this study whether this potentially acted as a barrier to earlier consideration of HAT [20], [24]; preliminary research by us suggested that peripheral HCWs considered this highly appropriate. Outside of formal consultations, HCWs less commonly relied on malaria/typhoid differential diagnoses to make HAT referral decisions, suggesting that HCWs may gain confidence in making syndromic referrals with an extended period of observation.
Along with expanding and consolidating existing syndromic HAT detection knowledge, we believe that the training intervention also empowered HCWs to think of HAT as more common in MC and as an appropriate syndromic diagnosis to make at the peripheral level, supporting translation of knowledge into practice (Figure 1). Evidence for this comes mainly from the increased numbers of HCWs who made a referral but who never had before, coupled with supportive statements by HCWs about what they learned from the training. Many HCWs proactively adapted their working practices to compensate for reduced patient attendance, for instance, when stock-outs of medicines were frequent or prolonged, such that identification of patients occurred outside of formal healthcare facilities as often as it did inside of them. HAT referral also appeared to be particularly accommodated as a useful differential syndromic diagnosis that could be referred on when malaria or typhoid was ruled out, as already discussed.
The training intervention largely, however, failed to empower patients to complete HAT referrals made in the periphery (or HCWs to meaningfully support the process), under the existing referral system's constraints. Despite a 6.5-fold relative increase in this type of referral received at Nimule Hospital attributable to the HCW training intervention, in terms of the absolute numbers of patients screened, this increase was modest. A third of patients referred for HAT from areas close to the hospital eventually presented for screening, but in areas more than one day's walk away (half of the county), only a very small minority of patients who were referred in the intervention presented.
Long distances and the lack of affordable transportation opportunities for patients were probably the most important barriers preventing patients referred for HAT from reaching the hospital for screening, as in other studies of centralised health service use in South Sudan and rural African settings [44]–[50]. Relatedly, the fact that patients referred from private and military facilities were more likely to complete referrals suggested that relative wealth may have been important, since private facility services (tests, drugs) cost money and soldiers receive government salaries. Treatment costs and other factors have been used to explain decisions by households who ‘wait to go’ in malaria studies [31] and transportation subsidies with community-based counselling have successfully increased referral completion in HIV programmes in Africa [51]. Although not measured in this study, higher household wealth and maternal education (of patients or primary care-givers) have been associated with increased maternal and child health care-seeking in South Sudan [52] and may have been important in patient referral completion here. Referring HCWs were rarely able to successfully support the referrals process (for instance, by assisting patients to secure transportation or advocating for soldiers to obtain departure orders) but quantitative evidence suggested that individual HCWs' level of clinical education may have influenced this process, as elsewhere [49], [53]. Specific ideas introduced in the training about considering HAT as a diagnosis for patients who commonly appear drunk or are imprisoned for strange/violent behaviour were used by HCWs who commonly interact with military populations. Given the apparent difficulty that military HCWs faced in advocating patient referrals, however, further intervention might be needed to support HCWs in discussions with commanders to resolve some of the uncertainty that leads to potential disagreement in this context [54]. Further study of referral completion from the perspective of patients might also generate other critical barriers and opportunities to support completion not identified here.
The planned introduction of HAT RDTs [14], improved microscopy techniques [55] and potentially safe oral medicines [56] appropriate for use in primary and secondary facilities offer a promising new approach to bring screening, diagnosis and treatment services closer to patients and HCWs in peripheral areas like MC in South Sudan. Offering syndromic training or implementing syndromic algorithms may help guide the rational use of RDTs as the first step in case detection algorithms and this study suggests that educational training can be used to augment peripheral HCW HAT syndromic referral practice. The implementation of either peripheral HCW-controlled syndromic or RDT technologies might conceivably also synergistically empower HCWs to make more syndromic referrals in the first place. In an elimination scenario, involving health workers in the periphery, close to the last patients, will also be essential to monitor the absence of new syndromic cases. As this study suggests, substantial infrastructural challenges related to the ruling out of other common diagnoses and management of HAT RDT-positive patients who need access to hospital-based confirmatory tests will need to be addressed.
We could not follow-up all HCWs (or any traditional practitioners) to identify all peripheral referrals made. The opportunity for patients to complete their referrals was limited to within four months of the training, regardless of how long after training that referral occurred, and completion could be associated with season. Furthermore, we had to rely on referring HCWs' accounts to understand reasons behind those uncompleted referrals.
A greater overall effect on passive case detection might have been observed in a higher prevalence scenario, where more syndromic referrals might have been made, even if referral completion rates remained the same. This suggests an even greater potential benefit of intervention for more highly endemic areas such as the Democratic Republic of Congo. Under-recognition of HAT symptoms in practice likely continued to be a problem in this study but we could not directly assess syndromic case detection performance during clinical encounters without changing clinical practice. Determining missed opportunities for referral in this circumstance would have been difficult, as no ‘gold standard’ tool (such as a highly discriminating syndromic algorithm [36], [57]) exists against which to evaluate HCW syndromic recognition of patients in need of HAT screening.
In this environment where peripheral HAT referral was uncommon, identifying a specific new type of referral behaviour (HAT referral ever made, baseline measure 1) appeared more informative and valid than retrospective estimation of monthly referral rates, using HCW recall without reference to records or a key timeline of events (baseline measure 2). If we take as true that two-thirds of referrals made after the intervention were by new referrers, three explanations are possible to explain the difference in effect the intervention appeared to have on baseline HCW referral behaviour: the monthly rate of referrals made by experienced referrers substantially decreased after the intervention, the estimated pre-intervention rate consisted mainly of ‘old’ referrals, or the pre-intervention rate was substantially over-estimated on questionnaires. We believe the latter explanation to be most plausible. Despite these limitations, which are inevitable in this kind of evaluation in South Sudan, this study is the first to evaluate the effect of peripheral HCW training on HAT referrals to a PS service.
In an era when approaches to HAT case detection and control must increasingly be integrated into health referral systems, it is important to understand the opportunities and challenges associated with syndromic case detection in first line facilities to design effective interventions. This is important now, and will also be important when RDTs for HAT become available in peripheral areas to assess whether they can be targeted to the right patients. In this training intervention in South Sudan, we were able to increase the numbers of peripheral HCWs who now identify potential cases and refer for HAT by expanding and consolidating pre-existing syndromic HAT detection knowledge and by encouraging consideration of HAT as a diagnosis that is appropriate to make at the peripheral level. Barriers to patients reaching facilities for confirmatory testing remain. Additional interventions will be required to improve peripheral HCW and patient access to screening, diagnosis and treatment to contribute meaningfully to HAT control under a passive case-detection approach. The advent of rapid diagnostic tests, if they prove useable in peripheral settings, may provide part of this solution.
|
10.1371/journal.ppat.1007246 | Type I IFN signaling blockade by a PASylated antagonist during chronic SIV infection suppresses specific inflammatory pathways but does not alter T cell activation or virus replication | Chronic activation of the immune system in HIV infection is one of the strongest predictors of morbidity and mortality. As such, approaches that reduce immune activation have received considerable interest. Previously, we demonstrated that administration of a type I interferon receptor antagonist (IFN-1ant) during acute SIV infection of rhesus macaques results in increased virus replication and accelerated disease progression. Here, we administered a long half-life PASylated IFN-1ant to ART-treated and ART-naïve macaques during chronic SIV infection and measured expression of interferon stimulated genes (ISG) by RNA sequencing, plasma viremia, plasma cytokines, T cell activation and exhaustion as well as cell-associated virus in CD4 T cell subsets sorted from peripheral blood and lymph nodes. Our study shows that IFN-1ant administration in both ART-suppressed and ART-untreated chronically SIV-infected animals successfully results in reduction of IFN-I-mediated inflammation as defined by reduced expression of ISGs but had no effect on plasma levels of IL-1β, IL-1ra, IL-6 and IL-8. Unlike in acute SIV infection, we observed no significant increase in plasma viremia up to 25 weeks after IFN-1ant administration or up to 15 weeks after ART interruption. Likewise, cell-associated virus measured by SIV gag DNA copies was similar between IFN-1ant and placebo groups. In addition, evaluation of T cell activation and exhaustion by surface expression of CD38, HLA-DR, Ki67, LAG-3, PD-1 and TIGIT, as well as transcriptome analysis showed no effect of IFN-I blockade. Thus, our data show that blocking IFN-I signaling during chronic SIV infection suppresses IFN-I-related inflammatory pathways without increasing virus replication, and thus may constitute a safe therapeutic intervention in chronic HIV infection.
| Innate and adaptive immune activation is one of the strongest predictors of HIV disease progression to AIDS and non-AIDS morbidity and mortality. Type I interferon (IFN-I) signaling is a major driver of such activation even in ART-treated people. Therefore, manipulation of IFN-I signaling has been proposed as a therapeutic approach to improve HIV disease outcome. However, we previously reported that blockade of IFN-I signaling in rhesus macaques during acute SIV infection results in increased SIV reservoir size, accelerated CD4 T cell depletion and faster progression to AIDS. Our present study addresses the more clinically relevant effects of IFN-I signaling blockade in chronic SIV infection during treatment with antiretroviral therapy. We found that administration of a novel, long half-life, PASylated IFN-I receptor antagonist suppressed IFN-I-related inflammatory pathways but, in contrast to the case of acute SIV infection, did not result in loss of control of SIV replication. Thus, our findings provide a rationale for the safe and effective use of such interventions to reduce inflammation in HIV-infected people during ART.
| Persistent inflammation during chronic HIV infection is a central contributing factor to immune exhaustion, CD4 T cell depletion and progression to AIDS [1–3]. Previous studies aimed at understanding the nature of this immune dysfunction have revealed a key role for type I interferons (IFN-I). IFN-I has been shown to suppress HIV infection in vitro [4] and SIV infection in rhesus macaques in vivo [5]. Studies in non-human primates have demonstrated a link between type I IFN responses and pathogenic SIV infection [6–8]. While IFN-I signaling resulting from SIV infection waned during the transition from acute to chronic non-pathogenic infection in SIV natural hosts African green monkeys and sooty mangabeys, a persistent response was associated with pathogenic infection and progression to AIDS in experimental SIV infection of rhesus and pigtail macaques. In humans, plasma levels of IFN-I have been shown to correlate directly with plasma HIV RNA and inversely with CD4 T cell count [9]. Moreover, administration of IFN-I to HIV-infected persons resulted in lower CD4 T cell counts [10] and increased CD8 T cell activation [11]. These findings attributed, at least in part, the severity of infection and exacerbation of the disease to type I IFN signaling and raised considerable interest in the potential therapeutic benefits of blocking IFN-I during infection. Associations between IFN-I and chronic viral infections have led to numerous studies where IFN-I signaling was manipulated [12]. Blockade of IFN-I signaling with anti-type I IFN receptor (IFNAR) antibody in murine LCMV infection resulted in reduced immune activation and improved viral clearance [13, 14]. Recently, two independent studies showed that administration of anti-IFNAR antibodies to ART-suppressed, HIV-infected humanized mice resulted in reduced immune activation and lowered reservoir size [15, 16]. The efficacy of IFN-blockade in the mouse models has provided a rationale for testing in the SIV-infected non-human primates. In our prior study, we found that blocking IFN-I during acute SIV infection in rhesus macaques resulted in reduced expression of antiviral genes, increased size of the SIV reservoir and accelerated CD4 T cell depletion and progression to AIDS [5]. Quite reasonably, this adverse outcome of IFN-I blockade during acute infection raised major concerns for the safety of IFN-I blockade during chronic HIV infection. As treatment with antiviral drugs reduces but does not completely normalize inflammation and IFN-I signaling [2, 17, 18], it is important to assess the effects of manipulation of IFN-signaling during chronic infection under ART. Thus, our primary objective in the present study was to test the effect of IFN-blockade on both inflammatory status and the control of viral replication during ART-treated and untreated chronic SIV infection in monkeys, and thus to establish the safety profile of this experimental therapy for clinical use in ART-treated HIV infection.
Eight weeks after rectal SIVMAC251 challenge, 25 rhesus macaques initiated daily ART while 10 remained untreated (Fig 1). To block type I IFN signaling, infected animals were treated with a PASylated type I interferon receptor antagonist (IFN-1ant) [5] which was obtained by fusion of the native IFN-1 antagonist with a 600 residue conformationally disordered chain of Pro/Ala/Ser amino acids [19], leading to a significantly prolonged plasma half-life of 19.4 ± 2.6 h in rhesus macaques and a 70-fold enhanced area under the curve compared with the unmodified IFN-1 antagonist (S1 Fig), while maintaining high receptor binding activity with KD = 451 ± 3 pM. Antagonist or saline placebo were administered either twice or three times per week from weeks 16 to 24 post-infection, with 3x weekly doses resulting in higher plasma concentrations compared to 2x weekly injections. Among the animals that did not receive ART, 4 placebo (group 1) and 6 IFN-1ant-treated animals (group 2) received 3x weekly injections from week 16 to week 24 and had no significant difference in plasma virus load (VL) before administration of the antagonist (S2 Fig). In the ART-treated groups, 9 placebo-treated animals (group 4) and 11 IFN-1ant-treated animals (group 4) received twice weekly injections while a set of 5 animals with a higher initial median plasma virus load (group 5) received 3 times weekly IFN-1ant. Despite significantly higher plasma VL from week 4 to 5 post-infection in the 3 times weekly IFN-1ant group, Plasma VL at start of ART (week 8) and at start of antagonist administration (week 16) showed no statistical difference between the groups.
We first assessed the longitudinal effects of IFN-I antagonist administration on the expression levels of ISGs at several timepoints: Baseline, post-infection/pre-ART (week 6), pre-IFN-1ant (week 14), during IFN-1ant (weeks 17, 18, 19, 22) and post-IFN-1ant (week 26). For this purpose, we chose a set of 19 ISGs previously demonstrated to be reduced by IFN-1ant treatment during acute SIV infection [5]. As measured by mRNA-seq, the expression levels of several ISGs were increased as expected in all groups following SIV infection (Fig 2A). In the absence of ART, animals treated with IFN-1ant also showed reduction of ISG expression levels compared to placebo animals; but these remained at higher levels in all ART-untreated animals compared to ART-treated animals. In the animals that received ART, there was a significant reduction of ISG expression by week 14, after 6 weeks of ART administration (P ≤ 0.0001; S3 Fig). With levels already markedly reduced by ART, administration of the antagonist to these animals from week 16 to week 24 resulted in further reduction of the ISG expression by week 18 and 19 in the 2x weekly treatment group, but were modest in the 3x weekly group (Fig 2A–2C). Quantification of the overall effect of IFN-1ant treatment on ISG expression by comparison of the pre-blockade and post-blockade (wk14 to wk19 post-infection (p.i.) absolute expression levels of the ISG set showed consistent reduction after IFN-1ant treatment in both the 2x weekly ART-treated and ART-untreated animals (Fig 2B). The overall expression of the ISG set showed a 1.9-fold reduction in ART-untreated IFN-1ant animals as compared to placebo (P ≤ 0.001; Fig 2C). Among ART-treated animals, 2x weekly administration of the IFN-1ant resulted in a 1.24-fold lower expression of these ISGs (P = 0.03) whereas a 3x per week regimen did not result in significant changes (Fig 2C). Overall, the observations of lowered ISG expression after administration of IFN-1ant in the 2x weekly ART group and in the ART-naïve group, demonstrated the pharmacological efficacy of the PASylated IFN-1ant during chronic ART-treated and ART-naïve SIV infection, building on our prior results in which antagonist efficacy was achieved in acute SIV infection. In order to assess the extent of the antagonist-mediated suppression, plasma levels of the cytokines IL-1β, IL-1ra, IL-6 and IL-8 were measured before and after antagonist administration. Despite a significant reduction in ISG expression levels, antagonist treatment in ART-untreated and ART-treated SIV-infected macaques had no effect on circulating levels of these cytokines (Fig 3), suggesting a selective suppression of inflammatory pathways. Collectively, these data demonstrate that IFN-1ant suppressed type I IFN associated inflammatory pathways. Importantly, ART status and viremia impacted on the extent to which antagonism of type I IFN signaling affected expression levels of ISGs. In the absence of ART, ISG expression was reduced but remained higher than when ART alone was given. By itself, ART significantly reduced ISG expression levels and further reduction upon antagonist treatment was more apparent with low viremia.
To assess the risk that blocking type I IFN signaling could pose for the control of SIV replication, we measured plasma VL during and after antagonist treatment. In the absence of ART, there was no significant difference between the plasma VL of animals receiving antagonist or placebo from antagonist administration at week 16 p.i. up to week 50 post-infection (Fig 4A). In the ART-treated arm, all groups had measurable plasma VL upon initiation of the antagonist or placebo at week 16 (median plasma VL ranging from 190 to 950 RNA copies/ml) and showed intermittent rebound in plasma VL up to week 49 (Fig 4B). However, from antagonist administration at week 16 until 34 weeks later (i.e. week 50 p.i), pairwise tests of VL or area under the curve showed no significant difference between IFN-1ant-treated animals compared to the placebo group. Regression models adjusting for week 8 (ART start) and week 14 (pre-IFN-1ant) plasma VL also showed no significant effect of the antagonist. Of note, elevated median plasma VL observed from week 20 to 42 in the animals treated 3x per week with antagonist are likely due to the higher VL in these animals prior to initiation of the antagonist treatment. Although statistical significance was not reached, median plasma VL of animals treated 3x per week were consistently higher compared to other ART-treated groups from weeks 10–15 (before antagonist; Fig 4B). To confirm that high plasma VL were unlikely to occur when administering the same IFN-1ant dose regimen to animals having more suppressed viremia, we extended ART treatment and delayed start of the IFN-1ant administration to week 35 instead of 16 in a separate group of animals. In these animals, the virus was further suppressed by week 35 and administration of 3x-weekly antagonist with follow-up for an additional 33 weeks (week 68 p.i) was not associated with plasma VL increases (Fig 4C).
Thus, administration of PASylated IFN-1ant to chronically SIV-infected macaques did not significantly impact plasma VL and appeared to be safe both in ART-treated and untreated infection. Individual plasma VL curves from baseline to week 50 p.i. for all groups are presented in S4 Fig. Given that type I IFN are also important for the control of other viral infections such as CMV, we measured plasma CMV load before and after antagonist administration and observed no increase in both the proportion of CMV+ animals or the plasma virus loads of animals that were CMV infected prior to antagonist administration (S1 Table). In addition, administration of the antagonist had no effect on blood chemistry (S2 Table) which is routinely used to gauge major bodily functions. Together, these data reinforce our findings that treatment of SIV-infected macaques with PASylated antagonist, during untreated or ART-suppressed chronic infection, was well tolerated and did not induce overt immunodeficiency.
We next set out to address whether administration of PASylated type I IFN antagonist impacts the size of the SIV reservoir. Virus DNA was measured in CD4 T cell subsets sorted from PBMC and LN (according to gating strategies shown in S5 Fig). In the PBMC, CD4 T cells were sorted into total CCR5+, central memory (CM) and effector memory (EM) cells. While some EM cells express CCR5 [20], all CCR5+ cells were gated prior to EM in order to measure virus DNA in the total fraction of cells expressing CCR5 given its role SIV infection. As expected, ART initiation at week 8 significantly reduced the amount of SIV gag DNA copies measured in all cell subsets in all animals in each experimental group (Fig 5A). By week 24, treatment with 3x weekly IFN-1ant in ART-treated animals marginally reduced SIV gag DNA copies in CCR5+ T cells only (P = 0.02). However, administration of 2x or 3x weekly IFN-1ant did not show significant effects on CM, EM or total SIV gag copies. In the absence of ART, SIV gag DNA copies remained high from week 8 through 24 in PB cell subsets, with no difference between antagonist and placebo groups. In the LN, CD4 T cells were sorted into CM, EM, germinal center follicular helper (GC Tfh) and non-germinal center Tfh (non-GC Tfh) subsets. After ART initiation, the amount of SIV gag DNA copies continuously decreased up to week 50 p.i. in all LN subsets in all animals, with no effect of the antagonist treatment given from week 16 to week 24 (Fig 5B). In the absence of ART, the cell-associated SIV gag levels remained high across all timepoints and unaffected by IFN-1ant administration in all LN subsets. While cell-associated virus measures are commonly reported using the same numerical denominator, adjusting for the actual frequencies of each subset within a sample gives a more accurate insight into the contribution of each cell subset to the total cell-associated virus DNA level. After adjusting for cell subsets frequencies in our samples, the distribution of SIV gag DNA copies across the various PB and LN subsets remained similar between IFN-1ant and placebo treated animals (Fig 5C and 5D). In the PBMC, infected cells belonged almost entirely to the central memory CD4 T cell type (94–100% in ART-treated and 86–91% in ART-untreated). In the LN, infected cells were in majority CD4 CM in ART-treated (up to 71%) and GC Tfh in ART-untreated (up to 55%) animals. We previously observed that type I IFN blockade during untreated, acute SIV infection results in increased reservoir size and accelerated progression to AIDS [5], and raised legitimate concerns regarding safety. In the current data, blocking type I IFN signaling during chronic SIV infection did not increase the size of the SIV virus reservoir irrespective of ART treatment status, even in animals in which residual SIV viremia persisted during ART.
Even though type I IFN antagonist treatment did not significantly affect cell-associated or plasma virus loads, we nevertheless explored its potential effect on modulating T cell function during SIV infection. T cell exhaustion is associated with persistent antigenic stimulation; many reports have previously highlighted that chronic type I IFN signaling during viral infections results in CD8 T cell exhaustion [21–23] and have shown that blocking IFN-I signaling restores T cell function in LCMV-infected mice [13, 14, 24]. Therefore, we assessed T cell activation and exhaustion by measuring frequencies of CD8 memory T cells expressing CD38, HLA-DR, Ki67, LAG-3, PD-1 and TIGIT before and after antagonist administration. None of these markers of T cell activation and exhaustion were affected by antagonist treatment in either ART-untreated or ART-treated SIV infection (Fig 6A and 6B). From our RNA-seq data, comparison of the expression levels of T cell activation and exhaustion genes (S6 Fig) as well as GSEA of T cell activation and exhaustion pathways (S3 Table) showed no significant effect of antagonist treatment. Furthermore, administration of the antagonist did not affect CD4:CD8 ratio and frequencies of CCR5+ CD4 T cells in the PBMC (Fig 6C). Therefore, blocking type I IFN signaling during ART-treated or untreated chronic SIV infection showed no significant impact on T cell activation or exhaustion.
Finally, we explored whether antagonist administration under ART would influence recrudescence of viremia upon ART interruption. Within a week after ART stop, there was a resurgence of viremia that reached pre-ART levels by week 3 and remained high up to 14 weeks after ART interruption in all groups (Fig 7). Comparison of individual antagonist groups to placebo showed that after adjustment for VL at ART initiation (week 8) there was no significant effect of the antagonist on the mean log10 plasma VL set point defined as 6 consecutive weeks starting 4 weeks after ART cessation. Thus, despite significantly reducing expression of ISGs, blocking type I IFN signaling in ART-treated chronic SIV infection did not result in increased plasma VL compared to placebo even after ART interruption.
Our findings are primarily relevant to the implementation of IFN-I blockade strategies in clinical HIV studies. Type I IFNs are important mediators of antiviral immunity but their permanent engagement in chronic HIV infection also contributes to a persistent inflammatory state that promotes pathology. As is being advocated for inflammatory diseases such as systemic lupus erythematous or systemic sclerosis [12], therapeutic blockade of IFN-I signaling could reduce inflammation and improve control of HIV infection. However, adverse outcomes observed in acute SIV infection [5] raised major safety concerns for the potential use of similar approaches during the chronic stage. Here, we assessed the effect of blocking IFN-I signaling during ART-treated and ART-untreated chronic SIV infection. Our principal findings were: (1) administration of an IFN-I receptor antagonist with prolonged half-life to ART-treated and ART-untreated SIV-infected rhesus macaques showed a therapeutic benefit in terms of lowering inflammation in part as observed by amelioration of ISG expression despite unaltered levels of measured pro-inflammatory cytokines; and (2) in contrast to observations made during acute SIV infection [5], blockade of IFN-I signaling during chronic SIV infection did not lead to loss of control of viral replication. In this regard, our study shows that in chronic SIV infection, even in situations with residual viral replication, blocking type I IFN signaling did not lead to loss of control of the infection; and supports the rationale that the use of an IFN-I antagonist during chronic HIV infection is safe.
The difference in outcome of IFN-I signaling blockade between acute [5] and chronic SIV infection is most likely due to the timing of IFN-I signaling for control of the infection. Significant increase in plasma VL and SIV reservoir size upon blockade of IFN-I signaling in the acute phase implies a critical role for IFN-I at the onset of infection. In contrast, our data show that once persistent infection has been established, IFN-I signaling plays a less prominent role in the control of virus replication. Recently, administration of exogenous IFN-I along with ART to chronically SIV infected animals was shown to increase ISGs expression with no effect on virus control [25]. Similar differences in the role of IFN-I between acute and chronic viral infection have been reported in LCMV studies. For instance, addition of exogenous IFN-I increases control of the infection in the acute stage but does not decrease virus titers in the chronic stage [22, 26]. While such findings support the notion that the importance of IFN-I signaling for viral control changes over the course of the infection, the underlying reasons remain unclear.
It remains uncertain what clinical benefits would emerge as a result of blocking IFN-I signaling during chronic HIV infection. In ART-treated HIV infection, IFN-I signatures remain elevated despite effective HIV suppression by cART [27]. Targeting IFN-I could further suppress residual inflammation and rescue T cells from exhaustion. Despite reduced expression of some of the most prominent antiviral ISGs downstream of IFNAR such as Mx1 and OAS2 [28], blocking IFN-I signaling in our study had no measurable impact on T cell activation or exhaustion. Of note, the higher antagonist dose (3x weekly) in our study was less efficacious in reducing ISG expression, which may be due to partial agonist activity of the antagonist to induce minimal ISG expression, which has been reported at high concentrations in vitro [29]. In addition, because the 3x weekly ART group had a generally higher VL compared to animals that received twice weekly IFN-1ant (Fig 4), there is likely a threshold where any measurable effect of the antagonist added to the effect of ART is influenced by the viremia at the time antagonist treatment is initiated. Our findings possibly reinforces the importance of timing on the outcome of IFN-I signaling blockade as administration of IFN-I antagonist [5] and more recently anti-IFN antibody treatment [30] in SIV infected macaques was shown to reduce T cell activation. The observations in our study also differ from observations made in prior studies on IFN-I blockade in chronically HIV-infected humanized mice which showed reduced expression of ISGs along with enhanced viral suppression and reduced T cell activation [15, 16]. There are notable differences in experimental design that could account for this discrepancy. With a maximum of 4 weeks antibody treatment in the mice as opposed to 8 weeks of a long half-life IFN-I antagonist in our study, it is possible that both timing and duration could influence the outcome of blocking IFN-I signaling. Blockade for a limited time may rescue immune responses but viral clearance mechanisms may require a contribution from the IFN-I signaling pathway. Another possibility may be related to the animal models used. While the studies in humanized mice indicate that a reduction of cellular makers of inflammation was associated with improved control of virus replication, our data in non-human primates suggest that a lengthy blockade of IFN-I signaling reduces ISG expression but has no effect on other inflammatory pathways or T cell activation and thus may not affect control of virus replication. The important point, however, is that blockade of IFN-I signaling in chronically SIV-infected non-human primates did not lead to an increase in SIV replication as is the case in acute SIV infection. Dissection of experimental variables such as timing, duration, dose and possibly the evaluation of combination interventions in future studies will delineate optimal conditions for the therapeutic blockade of IFN-I signaling. A recent study in LCMV demonstrated that virus control and T cell exhaustion are mediated by different type I IFNs despite their use of the same receptor [24] and HIV studies on administration of various IFN-I subtypes to humanized mice revealed differences in their ability to suppress the virus [31, 32]. Thus, an intriguing hypothesis is that by manipulating specific type I IFNs and their subtypes (i.e. IFNα vs. IFNβ) in HIV/SIV infection, it may be possible to decouple the various biological activities of the IFN-system and selectively target deleterious activities while maintaining beneficial ones.
Administration of antiretrovirals in our study successfully suppressed SIV but most animals had virus loads above the limit of detection even after 40 weeks of ART. Consequently, the antagonist was administered to animals with detectable plasma VL. The presence of this residual viremia allowed evaluation of the safety aspect of IFN blockade in chronic SIV infection, which was a concern in light of previous findings that blockade during acute infection accelerated mortality. While our study concluded that IFN-I blockade in chronic SIV infection did not impair control of the infection, it remains to be seen whether the pre-intervention plasma virus load influences the outcome of IFN-I blockade. The use of, for example, a macrophage-tropic virus to assess how IFN-blockade influences infection of myeloid cells will help gain a full understanding of the clinical implications of blocking IFN-I in chronically HIV-infected persons.
Despite reductions in risk of death with ART, high rates of serious non-AIDS events associated with inflammation [3] continue to reduce quality and expectancy of life in HIV-infected people [33, 34]. Thus, therapeutic blockade of IFN-I signaling which plays a key role in the persistence of inflammation, even during suppressive ART, has the potential to safely improve clinical outcome in HIV-infected persons.
40 Mamu A01- B08- B17- adult Indian origin Macaca mulatta were challenged by two intrarectal inoculations within five days of 1ml SIVMAC251 (1ml of 1:25 dilution, stock 8.91 x 108 SIV RNA copies ml-1). At week 8 post-infection, 30 animals were started on antiretrovirals while 10 animals were left untreated. The ART regimen was subcutaneous injections of 20mg/kg/day Tenofovir and 30mg/kg/day Emtricitabine (Gilead), as well as orally administered drugs including 100mg BID Raltegravir (Merck), 800mg BID Darunavir (Janssen Pharmaceuticals) and 100mg BID Ritonavir (Abbvie), all given mixed with food. To investigate the effect of blocking IFN-I signaling, infected animals received i.m. injection of 3.5mg per injection given 2 or 3 times per week with a type I interferon receptor antagonist (IFN-1ant) used previously [5] whose plasma half-life was significantly increased by PASylation [19] as assessed in healthy macaques. PAS-IFN-1ant was produced by fermentation in E. coli according to a published procedure [19] where a human IFN-α2b carrying the amino acid substitutions R120E, E159K, S160R and R162K (in the mature protein) was equipped with a structurally disordered N-terminal PAS#1 sequence of 600 residues and secreted into the bacterial periplasm to facilitate formation of the structural disulfide bonds. Purification was achieved by substractive anion exchange chromatography and ammonium sulfate precipitation followed by a cation exchange and a finishing anion exchange chromatography, resulting in a homogenous protein preparation with ≤5.5 IU endotoxin per mg protein. Based on plasma concentrations measured after in vivo administration of the antagonist to healthy macaques, the chosen regimen of 2 and 3 times per week should result in minimum plasma concentrations of 5 nM and 10–20 nM respectively (S1 Fig). Detailed experimental set-up and IFN-1ant treatment is presented in Fig 1.
Animal use and all study procedures (protocol VRC-13-453, renewed once as VRC-16-678) were approved by the Vaccine Research Center (VRC) Animal Care and Use Committee (ACUC), meeting National Institutes of Health standards; and in accordance with the American Association for Accreditation of Laboratory Animal Care guidelines, all federal, state, and local regulations, and in compliance with The Guide for the Care and use of Laboratory Animals. All animals, Indian origin rhesus macaques (Macaca mulatta) were socially housed at the National Institutes of Health with oversight from facility behavioral management staff. Primary enclosures consisted of stainless steel primate caging provided by a commercial vendor. Animal body weights and cage dimensions were regularly monitored. Overall dimensions of primary enclosures (floor area and height) met the specifications of The Guide for the Care and Use of Laboratory Animals, and the Animal Welfare Regulations (AWR's). All primary enclosures were sanitized every 14 days at a minimum, in compliance with AWRs. Secondary enclosures (room level) met specifications. All animals were housed under controlled conditions of humidity, temperature and light (12-hour light/12-hour dark cycles). Animals were fed commercial monkey chow, twice daily, with supplemental fruit or other produce at least three times per week. Filtered water was available ad libitum. Animals were observed at least twice daily by trained personnel, including behavioral assessments. Environmental enrichment included provision of species appropriate manipulatives, and foraging opportunities, as well music and video watching opportunities multiple times per week. No adverse events have been associated with study interventions. For procedures requiring chemical immobilization and sedation, different anesthetics were used at the discretion of the attending veterinarian according to the IACUC approved protocol. Prior to immunization, drug treatments or blood draws, anesthetics included Ketamine Hydrochloride 5.0–25.0 mg/kg IM with xylazine 0.5–1.0 mg/kg. For technical procedures, Buprenorphine Hydrochloride 0.015 mg/kg was administered. For euthanasia according to endpoints specified in the IACUC approved protocol, animals were initially sedated with ketamine (10–25 mg/kg, IM) followed by lethal overdose of sodium pentobarbital to effect.
Plasma was separated from EDTA blood by centrifugation and PBMCs were isolated by density centrifugation using Ficoll-Paque Plus (GE Healthcare) and Leucosep Centrifuge Tubes (Grenier Bio-One). Lymph nodes (LN) were collected into RPMI 1640 (Gibco) supplemented with 10% fetal bovine serum (Gibco) and 1% Penicillin-Streptomycin-Glutamine (Gibco) and cell suspensions were passed through a 70μm filter to remove debris.
For CD4 T cell subsets sort, PBMC were stained with fluorochrome-labelled mAbs anti-CD28-CY5PE, anti-CCR5-PE, anti-CD3-CY7APC, anti-CD4-BV605, anti-CD8-Pacific blue (BD Biosciences) and anti-CD95-BV785 (in house conjugated, BD Biosciences). LN cells were stained with anti-CD28-CY5PE, anti-CD3-CY7APC, anti-CD4-BV605 (BD Biosciences), anti-CD8-BV570, anti-CXCR5-PE (eBioscience) and anti-CD95-BV785 (in house conjugated, BD Biosciences). PBMC and LN CD4 subsets of interest were sorted and lyzed in proteinase K (100ug mL-1, Sigma Aldrich) for SIV gag qPCR. For assessment of T cell activation and exhaustion, cryopreserved PBMC were thawed and stained with fluorochrome-labelled mAbs anti-CD38-FITC (Stem Cell), anti-Ki67-CY7PE, anti-CD28-CY5PE, anti-CD3-CY7APC (BD Biosciences), anti-HLA-DR-TRPE (Life Technologies), anti-PD-1-BV711, anti-CD95-BV785 (Biolegend), anti-TIGIT-APC (ThermoFisher) and anti-LAG3-PE (R&D). All samples were stained with Aqua LIVE/DEAD Fixable Dead Cell Stain.
Plasma SIVgag RNA was assayed as described previously [35]. For cell associated virus, SIVgag and rhesus albumin DNA were simultaneously quantified in cell lysates by qPCR using plasmid standards for absolute quantification of gag and albumin copy numbers with the following primers and probes used at final concentrations of 625nM and 250nM, respectively:
SIV-Gag-F GTCTGCGTCATpTGGTGCATTC
SIV-Gag-R CACTAGkTGTCTCTGCACTATpTGTTTTG
SIV-Gag-P CTTCpTCAGTkTGTTTCACTTTCTCTTCTGCG
Rh-Alb-F TGCATGAGAAAACGCCAGTAA
Rh-Alb-R ATGGTCGCCTGTTCACCAA
Rh-Alb-P AGAAAGTCACCAAATGCTGCACGGAATC
Plasma concentrations of IL-1β, IL-1ra, IL-6 and IL-8 were measured by bioplex assay using a premixed non-human primate kit from Millipore according to the manufacturer recommendations.
DNA was isolated from plasma using the QIAamp DNA Blood Mini Kit (QIAgen 51106) according to manufacturer’s instructions. Isolated DNA was then analyzed using an Applied Biosystems QuantStudio Real-Time PCR system (12K Flex) with Cytomegalovirus (CMV) specific primers (Forward: ATC CGC GTT CCA ATG CA, Reverse: CGG AGG AGC ACC ATA GAA GGT) and a TaqMan Probe (6FAM CCT TCC CGG CTA TGG MGBNFQ). Each sample was run in triplicate for 40 cycles along with positive controls. Copy number was calculated by comparison to a standard curve and the viral load was reported as CMV copies/mL of plasma.
RNA was extracted from cryopreserved PBMCs using RNAzolRT (Molecular Research Center) according to the manufacturers’ instructions. Purified RNA was used for transcriptome analysis. Briefly, polyadenylated transcripts were purified on oligo-dT magnetic beads, fragmented, reverse transcribed using random hexamers and incorporated into barcoded cDNA libraries based on the Illumina TruSeq platform. Libraries validated by microelectrophoresis were sequenced on an Illumina HiSeq 4000 in 100-base single-read reactions. RNA-seq analysis was conducted at the Yerkes Nonhuman Primate Genomics Core Laboratory. Estimates of gene-wise and isoform-wise expression levels for individual genes were performed using DESeq2 [36]. The RNA-seq data were submitted to the Gene Expression Omnibus repository at the National Center for Biotechnology Information database (GSE112148).
To identify pathways differentially modulated by IFN-1ant, Gene Set Enrichment Analysis was performed as follows. For each contrast, transcripts were ranked by differential expression using the Signal2Noise metric. GSEA was performed using the desktop module available from the Broad Institute (www.broadinstitute.org/gsea/). GSEA was performed on the ranked transcript lists using 1,000 phenotype permutations, and random seeding. Gene sets used included the MSigDB (http://software.broadinstitute.org/gsea/msigdb/collections.jsp) H (hallmark), C5 (GO), C2 (curated), C7 (immunologic) gene sets ([37]), and additional custom gene sets.
|
10.1371/journal.pntd.0004024 | The Diversity and Geographical Structure of Orientia tsutsugamushi Strains from Scrub Typhus Patients in Laos | Orientia tsutsugamushi is the causative agent of scrub typhus, a disease transmitted by Leptotrombidium mites which is responsible for a severe and under-reported public health burden throughout Southeast Asia. Here we use multilocus sequence typing (MLST) to characterize 74 clinical isolates from three geographic locations in the Lao PDR (Laos), and compare them with isolates described from Udon Thani, northeast Thailand. The data confirm high levels of diversity and recombination within the natural O. tsutsugamushi population, and a rate of mixed infection of ~8%. We compared the relationships and geographical structuring of the strains and populations using allele based approaches (eBURST), phylogenetic approaches, and by calculating F-statistics (FST). These analyses all point towards low levels of population differentiation between isolates from Vientiane and Udon Thani, cities which straddle the Mekong River which defines the Lao/Thai border, but with a very distinct population in Salavan, southern Laos. These data highlight how land use, as well as the movement of hosts and vectors, may impact on the epidemiology of zoonotic infections.
| Scrub typhus, caused by the pathogen Orientia tsutsugamushi, is endemic in Southeast Asia, including Laos, accounting for up to 15% of cases of undifferentiated fever in adult patients. Despite its public health importance, little is known about the genetics of the O. tsutsugamushi population in Laos—this information is important for optimizing diagnostics and epidemiological surveillance. We conducted a 4 year prospective study to examine the genetic diversity of O. tsutsugamushi causing scrub typhus in Lao patients and highlight the geographical differentiation that can occur even within a small country.
| Scrub typhus, caused by the Gram negative obligate intracellular coccobacillus Orientia tsutsugamushi, is an important cause of acute febrile illness in Asia responsible for up to 23% of cases of undifferentiated fever [1]. The infection represents a major disease burden throughout a region ranging from northern Japan to Pakistan, to Russia in the north and northern Australia in the south. Over 55% of the world’s population lives in this densely populated endemic area [2]. It can affect patients of all ages, with at least one billion people living in rural areas at risk, and perhaps approximately a million patients needing medical attention every year [3]. Scrub typhus is transmitted to humans through the bite of infected larval trombiculid mites [4]. The clinical manifestations range from fever, headache, muscle pain, cough, and gastrointestinal symptoms, to coma, multi-organ failure and death [5].
In this study, we present data on the strain diversity and population structure of O. tsutsugamushi in Lao PDR (Laos), a country where the incidence of scrub typhus is almost certainly under-reported. Two recent studies found that up to 15% of adult patients with undifferentiated fever had scrub typhus [6, 7]. Molecular typing studies, aimed at understanding and monitoring the distribution of this disease, have been most commonly based on the highly polymorphic 56 kDa-outer membrane protein. This approach has indicated that the genotypes causing infection in Vientiane are similar to those circulating elsewhere in Laos and in Taiwan [8]. However, the use of a single surface protein gene marker can result in low resolution, or may provide misleading evidence concerning strain relatedness due to the confounding effects of recombination or immune selection. To address these shortcomings, we use multilocus sequence typing (MLST), which utilises sequences of multiple housekeeping genes. These are under less diversifying selection than surface protein genes, and are better able to both determine the relationships between closely related genotypes and reveal the genetic structure and mode of evolution of the bacterial population (clonal vs panmictic). Two alternative MLST schemes have been developed for O. tsutsugamushi; The first scheme characterised isolates from 84 Thai patients with scrub typhus, and revealed a highly diverse O. tsutsugamushi population with a very high rate of recombination [9]. Moreover, the rate of mixed infection, as indicated by ambiguous sequence at 1 or more loci, was as high as 25%. This high rate of mixed infection was not found by the authors who developed the second MLST scheme, although their scheme was applied to a relatively small number of cultured Cambodian strains rather than directly to patient blood [10]. Here we report the results of a 4 year prospective study aimed at determining the temporal dynamics and geographical structure of O. tsutsugamushi from patients from three different regions of Laos. We used the original MLST scheme (9) to characterise 74 isolates from these three different regions. The data confirm a highly diverse, recombining population, and reveal evidence for geographical structuring and local clonal expansion within Laos.
A prospective study of patients presenting with acute fever to three hospitals in Laos from August 2008 to December 2012 was carried out. The hospitals, chosen to be in central, north and south Laos, were Mahosot Hospital in the capital Vientiane (17° 57ʹ N 102° 36ʹ E) [6], Luang Namtha Provincial Hospital in the northwest (21° 00ʹ N 101° 24ʹ E), and Salavan Provincial Hospital in the south (15° 43ʹ N 106° 25ʹ E) [7] (Fig 1). At Mahosot Hospital rickettsial blood culture was performed on all patients with suspected typhus. These were positive by point of care diagnostic test for either anti-O. tsutsugamushi IgM (CareStart assay; AccessBio, USA or Scrub Typhus IgM ICT, PanBio Inc., Australia) [11] or anti-R. typhi IgM (murine typhus Dip-Sticks IgM IBT, Panbio Inc., Australia) [12]. At Luang Namtha and Salavan patients with fever for ≤ 8 days, an admission tympanic temperature of ≥38°C, no obvious causes of fever (e.g. abscess, severe diarrhea, pneumonia), and negative malaria rapid diagnostic test had rickettsial culture performed.
This study was approved by the ethical review committee from the Lao National Ethics Committee for Health Research, Ministry of Health of Laos (No 25/NECHR), the Oxford Tropical Ethics Committee, UK and the Ethical Committee of Faculty of Tropical Medicine, Mahidol University, Thailand (Approval no. MUTM 2014-029-01). All patients subject in this study have provided written informed consent, and parents or legal guardians of any children participant provided written informed consent on their behalf.
O. tsutsugamushi was isolated from EDTA blood by in vitro isolation as previously described [13]. Briefly, 5 ml of blood was drawn from the patient and centrifuged at 3,000 rpm for 10 min. 200 μl of the buffy coat was collected and mixed with 1ml of RPMI 1640 medium containing 10 mM HEPES (PAA, Austria) supplemented with 10% (v/v) fetal calf serum and transferred to L929 (mouse fibroblast) cell culture. The mixture was incubated in the presence of 5% CO2 at 35°C for 2 hours, the supernatant was removed and 5 ml new culture media were added and for further incubation. Cell culture media was changed three times per week by removing 2.5 ml media and replacing this with an equal volume of fresh media. Rickettsia infected samples were identified using indirect immunofluorescence assays as previously described [13]. Briefly, the bacterial culture was coated onto microscope glass slides and monoclonal antibodies for scrub typhus (STG-100), typhus group monoclonal antibody (TG-100) and spotted fever group monoclonal antibody (SFG-100) (Australian Rickettsial Reference Laboratory, Geelong, Australia) were added, followed by secondary goat anti-human IgA/M/G labelled with FITC (Invitrogen, USA). Fluorescence microscopy was used to identify cells infected with Orientia /Rickettsia. To confirm the presence of O. tsutsugamushi or Rickettsia spp., quantitative real-time PCR assays based on the 47 kDa outer membrane protein gene for identification of O. tsutsugamushi, 17 kDa surface protein gene for genus Rickettsia and ompB gene for R. typhi were performed as previously described [14–16].
Genomic DNA of O. tsutsugamushi from the in vitro cell cultures was extracted using QIAamp DNA Mini kit 250 (QIAGEN, USA) and characterized using multilocus sequence typing (MLST) as previously described [9]. The housekeeping genes gpsA, mdh, nrdF, nuoF, ppdK, sucB, sucD were amplified by PCR and sequenced in both directions using nested primer pairs (Table 1). The sequence data were edited and analyzed using SeqMan from LaserGene software (DNASTAR Inc., Wisconsin, USA) and allele numbers assigned by reference to previous data [9]. The MLST scheme for O. tsutsugamushi is hosted on the PubMLST website (http://pubmlst.org/otsutsugamushi/) [17]. The relationships between the STs were visualized using two implementations of the BURST algorithm [18]; e-BURST v. 3 and goeBURST [19]. The dN/dS of the 7 housekeeping gene partial sequences were calculated using START2-Sequence Type Analysis and Recombinational Tests, (http://pubmlst.org/software/analysis/start2/) [20]. A neighbour-joining tree of the isolates was constructed based on the 2,700 bp concatenated sequence of all loci (gpsA-mdh-nrdF-nuoF-ppdK–sucB-sucD) using MEGA version 5 [21]. The data in the current study was supplemented with existing data from Thailand [9]. The amount of genetic differentiation between populations from different geographical locations was estimated using F-statistics (FST), which reflect the rates of migration, mutation and drift [22]. The estimation of the recombination per mutation ratio (r/m ratio) was calculated by comparing the sequences of non-identical alleles in all single locus MLST variants with their clonal founders. Multiple nucleotide changes (>1) were assumed to be caused by recombination while single nucleotide differences not found elsewhere in the database were assumed to be due to de novo mutation [23].
A total of 2,844 patients presenting with acute fevers were recruited from the three hospitals: 1,401 from Luang Namtha, 893 from Mahosot hospital and 550 from Salavan. A total of 195 (6.8%) of these patients were culture positive for O. tsutsugamushi, 58 from Luang Namtha (4.1% of all patients from this hospital), 118 from Vientiane (13.2% of all patients from this hospital), and 19 from Salavan (3.4% of all patients from this hospital). The relatively high percentage of patients from Vientiane confirmed as scrub typhus positive may reflect difficulties incurred during the transportation of samples from the other two areas. A total of 215 isolates were confirmed by both IFA and PCR, of which 195 (90.7%) were scrub typhus and 20 (9.3%) were R. typhi (Table 2). The first 81 out of 195 (41.5%) O. tsutsugamushi isolates were selected for MLST; 74 (91.3%) were successfully amplified and sequenced, while the remaining 7 isolates produced poor quality data, most likely as a result of mixed infection. Of the final 74 isolates, 51 originated from Vientiane, 11 from Luang Namtha and 12 from Salavan. These 74 isolates corresponded to 50 different sequence types (STs), 43 of which were novel to this study (STs 50 through ST 92). Simpson’s index of diversity was calculated as 0.98 (95% CI 0.97–0.99) confirming a highly diverse population [24]. The seven STs that were not novel had been previously reported by Sonthayanon et al. in a study of O. tsutsugamushi from patients presenting to a hospital in Udon Thani, Northeast Thailand [9]. In the current study, the isolates corresponding to these seven previously recorded STs all originated in Vientiane, which is on the Laos/Thai border. These shared STs largely correspond to the clonal complexes previously described in the Thai study; ST29 and ST30 correspond to CC29, ST37 and ST25 correspond to CC37, ST9 corresponds to CC10 and ST4 corresponds to CC13. The remaining ST common to both studies was ST1 which is a single locus variant (SLV) of ST2, the second most common ST noted in the Thai study [9]. This is consistent with the view that these clusters are commonly encountered in both Thailand and Vientiane, although they appear not to have made significant incursions to other regions of Laos.
Of the novel STs, 34 originated from Vientiane, 7 from North-Laos (Luang Namtha) and 9 from South-Laos (Salavan). The proportions of novel STs are therefore very similar in Vientiane (66%) and Luang Namtha (63%), but slightly higher in Salavan (75%). These novel STs did not simply reflect different combinations of previously described alleles, as might be expected in this highly recombining species, but also new allele sequences. 18 new alleles were noted for gpsA, 6 for mdh, 13 for nrdF, 16 for nuoF, 17 for ppdK, 8 for sucB and 14 for sucD. This again points to considerable population diversity, both in terms of the overall number of STs, but also in terms of the number of alleles per locus.
The most common sequence type was ST86, which was represented by 7 isolates, all from Vientiane (Table 3). ST37, ST58, and ST71 were each represented by 3 isolates, and were also recovered from a single origin (the three ST37 isolates were all from Vientiane, the three ST58 isolates from Luang Namtha and the three ST71 isolates from Salavan). Twelve STs represented by two isolates each were noted, nine of which originated exclusively from Vientiane (STs 1, 4, 9, 30, 51, 59, 67, 78), with one pair from Luang Namtha (ST65) and one pair from Salavan (ST75). There was a single occurrence of an ST being recovered from more than one region; ST69 corresponded to one isolate from Luang Namtha, and one from Vientiane. In summary, whilst the majority of the STs (34/50; 68%) are only represented by a single isolate, in those cases where a single ST is represented by multiple isolates those isolates exhibiting the same ST also originate from the same region (with the exception of a single isolate pair). Given that a pair of isolates drawn at random from the data would be expected to originate from the same region only 51.6% of the time (calculated by summing the probabilities that a random pair of isolates both correspond to one of the three regions), this observation is therefore strongly indicative of geographical clustering and the local clonal expansion of specific STs.
The MLST data for the 74 isolates from Laos were visualized using eBURST (Fig 2A) and goeBURST (Fig 3). The major difference between these two implementations of the BURST algorithm is that goeBURST provides the option to depict links between STs that differ at more than two loci, whilst eBURST will only show single locus variant (SLV) and double locus variant (DLV) links. Three clonal complexes (CCs) are resolved by eBURST. ST86, which is the most common ST, defines 4 single-locus variants (SLVs) (STs 58, 85, 87, 88) and 1 double locus variant (DLV) (ST84). ST29 is represented by two isolates and also defines two SLVs (ST30, ST55). ST25, which is represented by a single isolate, defines 2 SLVs (ST37, ST83). Three SLV pairs are noted (ST52 and ST53, ST56 and ST57, ST64 and ST65). Relaxing the linkage criteria to double locus variants reveals that the pair ST56 and ST57 are connected to CC86, ST54 is connected to CC25, ST62 and ST63 are connected to the ST64/65 pair, and ST66 and ST77 are joined by a DLV link. Twenty-seven of the 50 STs were not linked to any other STs on the basis of single or double locus variation, which further illustrates the high level of allelic diversity in this population. Analysis using goeBURST also pointed to ST86 as a likely founder (Fig 3), and shows several additional putative links of descent from this genotype by relaxing the criteria for joining STs.
Our data were then compared with previously published data using comparative eBURST (Fig 2B). Besides the seven STs that were noted previously in Thailand and in Vientiane (but not other regions in Laos), there is almost no overlap between the two countries, and only a single SLV link was noted between STs from Thailand (ST43) and Laos (ST59). Although there has been some O. tsutsugamushi migration between Thailand and Vientiane, as indicated by the 7 shared STs, Vientiane lies across the Mekong river from Udon Thani in Thailand and it is not surprising that STs present in Thailand are also present in patients in this city. However, there is no evidence from this analysis for overlap between the Thai isolates and isolates from elsewhere in Laos.
In order to examine further the phylogeny and geographical structuring of the O. tsutsugamushi isolates, we combined the data for the 74 isolates from Laos with data for 83 strains from Thailand and 6 reference strains (Karp, Kato, Gilliam, Sido, Boryong and Ikeda), giving a total of 163 strains. For each strain, we concatenated the individual allele data, resulting in fragments of 2,700 bp length. This alignment was then used to construct a neighbour-joining tree as implemented in MEGA version 5 (Fig 4).
The clusters resolved by eBURST from both the current and previous Thai study are annotated on the phylogenetic tree. The largest clusters of Thai isolates are CC29 and ST1/ST2, and the tree shows the identical, or very closely related isolates, from Vientiane that correspond to these clusters. The most common clonal complex in the current study, CC86, is confirmed to be composed of a mixture of isolates from Vientiane and Luang Namtha. A second cluster of five isolates from Luang Namtha (including ST65) and three from Vientiane is also noted. In contrast eight of the isolates from Salavan correspond to a single diverse cluster (incorporating ST71 and ST75), which also incorporates a single isolate from Thailand. The other four Salavan isolates are found elsewhere in the tree, but do not correspond to any of the major clusters. The relatively high level of diversity in the major Salavan cluster points to a local population of bacteria circulating in relative isolation in this region over a protracted period of time. In contrast, the more closely related clusters (those of common STs and close relatives that can be identified by eBURST, such as CC29) represent more recent introductions in to a region followed by rapid clonal expansion.
The phylogenetic analysis paints the following general picture regarding geographical structure and spread. Isolates from Vientiane overlap with isolates from Udon Thani and from Luang Namtha, pointing to a key role of the capital both as a focus for movement of isolates between Thailand and Laos and also potentially as a reservoir for spread to and from other parts of Laos. However, there is no evidence for migration between Udon Thani and Luang Namtha. In contrast, the isolates from Salavan appear distinct from all other regions and are less clustered. This indicates that the strains in this region have been diversifying in relative isolation.
In order to explore this picture further we computed pairwise FST values for each of the four populations corresponding to Vientiane, Luang Namtha, Salavan and the previously published strains from Udon Thani. These values are given in Fig 5 along with a dendrogram illustrating the level of differentiation between the four populations. Of the 6 pairwise comparisons, two show low levels of differentiation, two moderate and two high. Low levels of genetic differentiation (~0.05) are apparent between the Vientiane and the Thai populations, and between the Vientiane and Luang Namtha populations, consistent with the phylogenetic analysis. A moderate level of differentiation is noted between the Thai and Luang Namtha populations and the Vientiane and Salavan populations (0.17 and 0.18 respectively). Finally, a high level of differentiation is seen between the Salavan population and both the Thai and Luang Namtha populations (0.24 and 0.27 respectively). In summary, this analysis is consistent with the interpretation of the phylogenetic tree in confirming the relative isolation of the Salavan population, and pointing to the Vientiane population as being most “central” (i.e. showing the lowest average differentiation to all other populations).
The ratios of dN/dS of the 7 loci corresponded to a range of 0.002–0.310 (Table 4), confirming that all genes are evolving predominantly by purifying (stabilizing) selection. This also indicated that synonymous substitutions were more common than non-synonymous substitutions for all the genes tested for these bacteria. We note that the dN/dS ratio of sucD and nrdF is approximately an order of magnitude lower than the other genes. This was also apparent in the previous data in Thailand [9] and suggests particularly strong selective constraint on these two genes in the O. tsutsugamushi genome.
In order to understand how O. tsutsugamushi was diversified, an estimation of the ratio of recent recombination to mutation events (r/m) in clonal complexes was performed by comparing the sequences of mismatched alleles in clonal founders and single locus variants [23]. The estimated ratio of recombination to mutation of these two populations in Laos (n = 74) and Thailand (n = 89) was high at both the nucleotide level (95:1) and at allele level (17:1), suggesting that the diversification of O. tsutsugamushi is predominantly characterized by recombination rather than mutation at both nucleotide level and allele level.
This study represents the first investigation into the diversity and phylogeography of O. tsutsugamushi in Laos. As the isolates were characterized using the same MLST scheme, it was possible to combine these data with those from a previous study focusing on strains from Udon Thani in Northeast Thailand. Our results reveal a highly diverse and recombining population in Laos, as evidenced by the high diversity index (0.98) with high number of STs per isolate (50 STs in 74 isolates, 0.68 STs per isolate). This is consistent with the previous Thai study (0.95 STs per isolate) [9], although the diversity in the current Lao sample set might be expected to be slightly higher, as it represents three distinct geographic sources. The ecological implications of the very high rate of recombination are unclear, but this may reflect co-colonisation of either the mites or the rodents. It is possible that high rates of recombination might also reflect a mechanism for diversification and host adaptation in O. tsutsugamushi [25]. The O. tsutsugamushi genome displays a massive proliferation of mobile elements and repeat sequences [26] which are thought to facilitate horizontal gene transfer.
Approximately 8.6% (7/81) of the sequenced isolates appeared to represent mixed infections in patients. This is a lower frequency than in Thailand, where 25% of cultures from patients’ blood were noted to be probable mixed infections. It is possible that the relatively low frequency of mixed infection in the Laos data is an artifact resulting from the in vitro culture which may have acted to amplify single predominant strains at the expense of more rare variants over several passages. It is also possible that the predominant isolate was more highly adapted to the culture conditions.
It is not clear whether mixed infection primarily results from multiple mite bites, or from co-colonisation of multiple strains within individual mites. The latter possibility is supported by the detection of multiple antigenic strains of O. tsutsugamushi in both naturally infected and laboratory-reared chigger mites (Leptotrombidium spp.) [27, 28]. As expected, the MLST genes show low dN/dS ratios, which is indicative of stabilizing selection. This is particularly true for sucD and nrdF which may be under unusual levels of selective constraint. The sucD gene produces succinyl-CoA synthase which uses succinyl CoA as a substrate to produce succinate and generate GTP in the citrate pathway (TCA cycle) [29, 30]. Unlike other rickettsia O. tsutsugamushi has no active pyruvate dehydrogenase enzyme to convert pyruvate to acetyl-CoA (30) and has only a partial TCA cycle starting with α-ketoglutarate and ending with oxaloacetate. This ‘minimal’ citric acid cycle requires succinyl-CoA synthase and may explain why sucD is relatively conserved in Orientia. The ribonucleoside-diphosphate reductase beta subunit gene (nrdF) is involved in purine and pyrimidine biosynthesis in O. tsutsugmushi. The genetic and metabolic diversity in Rickettsia has been reported previously [30]. While the proteins unique to Rickettsia spp. represent a broad spectrum of functional categories (carbohydrate, lipid transport), more than 60% of the proteins unique to O. tsutsugamushi belong to the replication, recombination and repair process categories. The nrdF gene product is strongly associated with these 3 processes, perhaps explaining why it is relatively conserved in O. tsutsugamushi.
Limitations of the study include the fact that not all O. tsutsugamushi isolates underwent MLST and that the patient inclusion criteria for patients recruited in the north and south of the country differed from those recruited in the centre.
Our MLST data reveals a mixed picture concerning geographic structure and migration. First, there is clearly migration and overlap between the strains from Vientiane and from Udon Thani. This is evidenced by shared STs, the intermingling of the isolates on the neighbour-joining tree, and by the FST analysis. This is perhaps not surprising as these locations straddle the Lao/Thai border. It is possible that disease transmission occurs via the trading activities of villagers in the border area, commuting, or tourism. This may be due to both human movement and movement of mites via animals.
There is no evidence of overlap between the Thai isolates and the Lao isolates from the north (Luang Namtha) or the south (Salavan). There is, however, evidence of transmission between Vientiane and Luang Namtha, particularly with respect to the most common cluster observed in our study, CC86. In contrast, most of the strains from Salavan appear to form a loose cluster which is not closely related to isolates from any of the other regions.
Salavan and Luang Namtha have remained relatively undisturbed rural hinterlands with environments which may be particularly suited to the maintenance of large populations of chigger mites. In contrast, Vientiane (the capital of Laos) is rapidly expanding into surrounding paddy fields and the rural hinterland [31]. This anthropogenic disturbance has likely had a major impact on the life cycle, ecology and behaviour of O. tsutsugamushi bacteria, their mite vectors, and their rodent and human hosts to limit the spread of the disease. A study on spatial distribution of scrub typhus in Vientiane demonstrated that the prevalence of scrub typhus IgG antibodies among patients from rural villages is significantly higher than that in patients from urban settings. Moreover, many urban patients are positive for O. tsutsugamushi IgG suggesting prior exposure to scrub typhus, possible in rural settings [31].
|
10.1371/journal.pcbi.1004032 | Laminar and Dorsoventral Molecular Organization of the Medial Entorhinal Cortex Revealed by Large-scale Anatomical Analysis of Gene Expression | Neural circuits in the medial entorhinal cortex (MEC) encode an animal’s position and orientation in space. Within the MEC spatial representations, including grid and directional firing fields, have a laminar and dorsoventral organization that corresponds to a similar topography of neuronal connectivity and cellular properties. Yet, in part due to the challenges of integrating anatomical data at the resolution of cortical layers and borders, we know little about the molecular components underlying this organization. To address this we develop a new computational pipeline for high-throughput analysis and comparison of in situ hybridization (ISH) images at laminar resolution. We apply this pipeline to ISH data for over 16,000 genes in the Allen Brain Atlas and validate our analysis with RNA sequencing of MEC tissue from adult mice. We find that differential gene expression delineates the borders of the MEC with neighboring brain structures and reveals its laminar and dorsoventral organization. We propose a new molecular basis for distinguishing the deep layers of the MEC and show that their similarity to corresponding layers of neocortex is greater than that of superficial layers. Our analysis identifies ion channel-, cell adhesion- and synapse-related genes as candidates for functional differentiation of MEC layers and for encoding of spatial information at different scales along the dorsoventral axis of the MEC. We also reveal laminar organization of genes related to disease pathology and suggest that a high metabolic demand predisposes layer II to neurodegenerative pathology. In principle, our computational pipeline can be applied to high-throughput analysis of many forms of neuroanatomical data. Our results support the hypothesis that differences in gene expression contribute to functional specialization of superficial layers of the MEC and dorsoventral organization of the scale of spatial representations.
| Higher brain functions such as spatial cognition are carried out in specialized brain areas. Within a specialized brain area nerve cells with different functions are organized in layers and gradients. It is possible that this topographical organization reflects underlying differences in molecular organization of the brain. However, systematic comparison of the expression patterns of tens of thousands of genes at the resolution of layers and borders is challenging. Here we develop a new computational pipeline that addresses this problem. We apply this pipeline to analysis of the medial entorhinal cortex (MEC), a brain structure that is important for spatial cognition. Our analysis shows that the MEC is highly organized at a molecular level, identifies related groups of genes that might underlie functional specialization, and implicates energy-related genes in vulnerability of certain neuronal populations to neurological disorders including Alzheimer’s disease. Our computational pipeline may have general utility for high-throughput and high-resolution analysis of brain anatomy. Our results support the notion that molecular differences contribute to functional specialization of higher cognitive circuits.
| Spatial cognition emerges from interactions between specialized neuronal populations in the hippocampal-entorhinal system [1]. The medial entorhinal cortex (MEC) is of particular importance for cognitive functions that rely on estimation of spatial position and orientation [2]. Neurons in each layer of the MEC represent distinct information, have differing connectivity, and can be distinguished by their morphological and biophysical properties [3–7]. For example, layer II has a relatively high density of neurons with grid firing fields, whereas deeper layers contain a higher proportion of neurons with firing also modulated by head direction [3, 8]. Further topographical organization is present orthogonal to cell layers along the dorsoventral axis in that the scale of spatial representations, local and long-range connectivity, synaptic integration and intrinsic electrophysiological properties all vary with dorsoventral position [9–15]. While this specialization of encoding and cellular properties is well established, the extent to which molecular specialization defines neuronal populations within the MEC or contributes to their distinct functions is not clear.
Insights into molecular substrates for topographical organization in other brain regions have been gained through large-scale analysis of differences in gene expression [16–21]. Our understanding of the architecture and functions of MEC may benefit from similar approaches. While detailed anatomical and histochemical studies have shown that certain genes, including reelin, calbindin [22] and some cadherins [23], identify cell populations associated with particular layers of the MEC, we know very little about the identity, laminar or dorsoventral organization of the vast majority of genes expressed in the MEC. This is a difficult problem to address for the MEC as its borders with adjacent structures are ambiguous, it has a dorsoventral as well as a laminar organization and its similarity to other cortical structures is unclear. As a result, key questions about its molecular organization are currently unanswered. For example, are the layers and borders of the MEC unambiguously delineated by coordinated expression patterns of multiple genes? Are genes differentially expressed along the laminar and dorsoventral axes? Do genome-wide laminar or dorsoventral differences in gene expression lead to mechanistic predictions regarding the organization of functional properties in the MEC? The MEC is ontogenetically distinct from both the 3-layered hippocampus, and from neocortex [24], with which it shares a similar laminar organization, but does this similarity to neocortex reflect a common molecular organization?
The topographical organization of the MEC extends to pathological signatures of common disorders in which it is implicated. Layer II exhibits neuronal loss [25] in patients with mild to severe Alzheimer’s disease (AD) and altered excitability in animal models [26]. Layer II is also affected in individuals with schizophrenia where there is evidence of abnormalities in cell size, organization and RNA expression [27, 28]. In contrast, epilepsy is primarily associated with loss of layer III neurons in humans [29] and in animal models [30]. Yet, the mechanisms that predispose different cell populations in the MEC to particular disorders are not known. Given the evidence of genetic associations with these diseases [31–33], it is possible that the molecular profiles of particular MEC neurons confer vulnerability. However, testing this hypothesis requires knowledge of the laminar and dorsoventral organization within the MEC of genes that are causally involved in disease.
To better understand the molecular basis for its function and pathology, we aimed to establish a genome-wide approach to define the laminar and dorsoventral organization of the MEC transcriptome. Large-scale investigations into gene expression patterns have previously used either in situ hybridization (ISH) or RNA sequencing (RNA-Seq) to identify genes with differential expression in the neocortex [16, 34], hippocampus [17] and sub-cortical structures including the striatum [20]. While transcriptomic approaches such as RNA-Seq provide a robust platform for quantification of transcripts, accurate isolation of cell populations at the resolution of cell layers is challenging [16], limiting the current applicability of this approach. In contrast, ISH is more useful for identifying patterns of gene expression because the precise location of transcripts can be examined. The Allen Brain Atlas (ABA), a high-throughput ISH database, contains brain-wide data for over 20,000 genes and has been used to identify laminar borders, and to distinguish regions and cell types in the somatosensory cortex [34], hippocampus [17], and cerebellum [35]. However, in its currently accessible form ABA data is searchable at best at a resolution of 100 µm [34] and is further limited in its utility for automated comparison of laminar gene expression because the magnitude of error in alignment accuracy of brain sections is comparable to the width of narrow individual cortical layers. Small alignment errors can therefore easily lead to incorrect assignment of genes to layers, thereby confounding systematic analysis of differences between layers.
To address these issues, we established a computational pipeline for registration and automated analysis of ISH data at a resolution of approximately 10 µm, enabling us to compare precise spatial expression patterns of over 80% of genes in the ABA dataset [34, 36]. We combined analysis of this high spatial resolution data with RNA-Seq analysis of gene expression in dorsal and ventral regions of the MEC. We demonstrate that while very few genes are uniquely expressed in the MEC, differential gene expression defines its borders with neighboring brain structures, and its laminar and dorsoventral organization. We propose a new molecular basis for distinguishing the deep layers of the MEC and provide evidence that at a molecular level deep layers of the MEC are relatively similar to those of neocortex. Superficial layers are substantially more divergent between neocortex and MEC. Analysis of genes with differential expression suggests roles in layer-specific and dorsoventral specialization of calcium-ion binding molecules, ion channels, adhesion molecules and axon guidance-related molecules. We find that differential laminar expression patterns do not extend to genes directly implicated in disease, but selective expression of related genes may provide a context that confers vulnerability to pathology in neurodegenerative diseases such as AD. Our data establish a genome-wide framework for addressing the organization of circuit computations and pathology in the MEC.
To be able to systematically compare expression of genes in the adult mouse MEC at laminar resolution we extended the precision with which the localization of expressed genes in the ABA dataset can be compared. To achieve this we implemented methods to warp ISH images and their corresponding processed expression images, in which pixel intensity is used to represent the relative total transcript count [34], into a standard reference frame (see Materials and Methods, Fig. 1A-B and S1A Fig.). Our re-registered ABA data set contains at least 1 image meeting our quality criteria and that contains the MEC for 16,639 genes (81.2% of genes in the ABA sagittal dataset) (S1B Fig.). Non-linear registration provides a striking improvement in the spatial resolution at which the localization of gene expression can be compared (Fig. 1C). Prior to re-registration, exploration of the organization of gene expression is confounded by variability in the shape and size of brain sections used in different ISH experiments. For example, averaging ISH expression images for 1000 genes prior to registration results in diffuse images without any laminar organization (Fig. 1C). In contrast, following re-registration of the same images, neocortical layers are clearly distinguishable from one another (Fig. 1C). Other landmarks, for example the white matter border with the striatal region, the hippocampal pyramidal layer and layer I of the piriform cortex, also become clearly identifiable (Fig. 1C). Within the MEC, the resolution is such that multiple layers and a dorsoventral organization can be recognized. Thus, our computational pipeline for image registration enables high-resolution comparison of gene expression between layers and along the dorsoventral axes of the MEC, as well as with other brain regions.
To validate gene expression data extracted from the ABA we compared mean pixel intensity values across dorsal and ventral MEC with RNA-Seq data acquired from the same regions (see Materials and Methods, S1A Fig.). Our RNA-Seq analysis detected 20,106 of the 38,553 genes (52.7%) in the Ensembl mouse database (release 73), including 15,496 protein-coding genes. In comparison, of the 16,639 genes from the ABA dataset for which we successfully registered images, our analysis revealed that 9,873 (59.3%) are expressed in the MEC, including 8,941 genes that we could identify in the Ensembl database (Fig. 1D). Of the registered Ensembl genes, 8,064 (90.2%) were also detected by RNA-Seq, indicating a high degree of consistency between the two approaches (Fig. 1D). This is supported by a significant positive correlation between RNA-Seq transcript FPKM (fragments per kilobase of exon per million fragments mapped [37]) and mean pixel intensity of ABA images (r = 0.40, p < 2.2 × 10-16) (Fig. 1E). It is possible that the 877 genes that appear to be expressed in ABA data, but are not detected by RNA-Seq, are false positives, while the 3,297 genes detected using RNA-Seq that are not detectable in the ABA data, may reflect false negatives in the ISH data, for example due to errors in probe design, staining or image processing. A further 6,463 genes detected using RNA-Seq that are not in the re-registered ABA dataset (Fig. 1D) include 1,939 pseudogenes and 791 long intergenic non-coding RNAs, as well as 1,855 protein-coding genes.
Together, these data demonstrate the potential of combining advanced image processing tools for high resolution alignment and analysis of ISH data sets with RNA-Seq. RNA-Seq enables quantification of thousands of transcripts over a large dynamic range, while automated analysis of ISH data reveals gene expression at laminar resolution that can be quantified and compared within and across brain regions.
Previous investigation using microarrays to compare tissue harvested from multiple brain regions has shown that gene expression in the entorhinal area is most similar to that of neocortex and hippocampus and least similar to non-telencephalic regions [38]. However, it is not clear if there are individual genes that distinguish these brain regions, whether differences show laminar or dorsoventral specificity or if this pattern applies to the MEC specifically rather than the entorhinal area as a whole.
To first determine whether gene expression in the MEC alone shows a relationship to other brain areas similar to that established by microarray analysis, we used the re-registered ABA dataset to isolate ISH gene expression data from several brain regions (Fig. 2A). Consistent with microarray data [38], we found that gene expression in MEC correlates most strongly with neocortex (r = 0.958, p < 2.2 × 10-16), amygdala (r = 0.936, p < 2.2 × 10-16) and the hippocampal pyramidal layer (r = 0.942, p < 2.2 × 10-16) and more weakly with the caudate putamen (r = 0.887, p < 2.2 × 10-16) (Figs. 2B, S2A-B). The relatively high correlations between the MEC and the other regions suggests that differential expression of relatively few genes is likely to underlie functional differences between these areas.
To investigate whether the expression of single genes could distinguish MEC from other regions, we identified genes with at least 4-fold higher mean pixel intensity in MEC compared with each of the other regions (S2C Fig., Materials and Methods). This analysis revealed 118 genes that are expressed at higher levels in MEC than neocortex, 93 for piriform cortex and 54 for the hippocampus, compared with 318 for the amygdala and 1,162 for the caudate putamen. These numbers decrease as the threshold difference in mean intensity between MEC and the other regions is increased (S2D Fig.). Section-wide average images reveal that the expression within the MEC of these genes is often not uniform, but can be concentrated in specific layers (e.g. MEC vs neocortex) or dorsoventrally organized (MEC vs piriform cortex and amygdala) (Fig. 2C). They also reveal that genes selectively enriched in MEC compared with one region are, on average, also strongly expressed in other regions (Fig. 2C). Nevertheless, 3 genes could be identified as uniquely enriched in the MEC compared with the other 5 regions (Fig. 2C), although expression of each followed a laminar organization and did not mark the MEC as a whole. We also asked if combinations of expressed genes might better distinguish the MEC from other regions. However, we found that only in a minority of pairs of genes with converging expression in MEC (14/456) does expression fully colocalize to the same laminar and dorsoventral regions (S2E-F Fig.).
Together, these data further validate our quantification of MEC gene expression and indicate that few, if any, individual genes or pairs of genes are likely to distinguish the MEC as a whole from other brain regions. Thus, specific attributes of the MEC are unlikely to be a product of highly specific expression of a few genes. Instead, our data are consistent with combinatorial expression of larger sets of genes defining differences between cell populations in the MEC and other brain areas (c.f. [39]). Our data also highlight a limitation of regional comparison of gene expression in that genes which are co-expressed in a given brain region may not be colocalized to the same cell layer or dorsoventral area. Therefore, to better understand the laminar and dorsoventral organization of gene expression in the MEC, and the relationship between the organization of the MEC and neocortex, with which it has the most similar overall gene expression [38], we took advantage of our pipeline for large-scale comparison of gene expression to analyze expression at laminar and sub-laminar resolution.
Borders of the MEC, which we consider here as the region also previously referred to as the caudal entorhinal field [4, 40, 41], have typically been defined on the basis of classical cytoarchitectonic criteria, chemoarchitecture and connectivity [11, 40, 42]. However, because these criteria don’t always converge, ambiguity exists regarding the definition and location of the borders with adjacent regions including the parasubiculum [43, 44] as well as with more ventral structures (c.f. [42, 45, 46]). We therefore sought to determine whether ISH data, and in particular our re-registered ABA sagittal data set, would enable a clearer resolution of dorsal and ventral MEC borders, which can be viewed unambiguously in the sagittal plane.
We focused initially on identifying genes that delineate the dorsal border of the MEC. In some atlases this region is considered as the retrosplenial or perirhinal region [47], and in others as the ectorhinal region [34, 36]. However, cytoarchitectonic, histological and electrophysiological studies in rats suggest that part of this region corresponds to the superficial layers of the parasubiculum [43, 48]. By comparing relative pixel intensity between the dorsal MEC and adjacent regions (see Materials and Methods, Fig. 3A), we identified a number of genes with expression that appears to stop at the dorsal border of the MEC (e.g. Wdr16 and Fabp5) (Fig. 3B). For some of these genes, expression is only absent within a wedge-shaped region before resuming in more dorsal cortical areas (e.g. Nov), a pattern which is highly consistent across different medio-lateral sections (Figs. 3C, S3A-B). We therefore asked if there are genes that are expressed in the wedge-shaped region, but not the adjacent regions. We identified 9 such genes (see Materials and Methods), including Igfbp6 and Kctd16 (Fig. 3D). All of these genes are also expressed in the parasubicular region medial to the MEC (Fig. 3E). Of these, 7/9 also have sparse expression in superficial parts of MEC (Fig. 3D). These observations support the view that the parasubiculum extends to wrap around the dorsal border of MEC [43, 48].
At the ventral aspect of the MEC, cytoarchitectonic analysis delineates a border with the medial entorhinal field [40, 47]. We asked whether differential gene expression supports the presence of this ventral border and whether it can clarify its position. By analyzing gene expression in regions either side of the approximate location of this border (Fig. 3F), we identified genes with expression that drops off sharply (Fig. 3G, S3C-D Fig.). These genes include apparently layer-specific genes such as Nef3, Cutl2 and Col5a1 in layers II, III and V/VI respectively. We also identified genes with the converse pattern of high ventral expression and low MEC expression, including Kctd6, Sema3c and LOC241794 (Fig. 3H). These expression patterns are consistent across different mediolateral sections (S3C-D Fig.). Thus, the ventral border of the MEC can be identified by genes with sharply increased or reduced expression.
Cytoarchitectonic and developmental studies indicate that the MEC is a type of periarchicortex (paleocortex), a transitional structure between 6-layered neocortex and 3-layered archicortex, with 5 cytoarchitecturally distinct cell body layers [24, 40, 45]. However, while layer II and III are easily distinguished by their cytoarchitecture and connectivity (cf. [40]), differentiation of cell populations within layers V and VI is less well established, although there is evidence that the cell bodies and dendrites of distinct cell types are differentially distributed within these layers [4, 6, 40, 49]. Layer IV corresponds to the cell free lamina dissecans [40]. To better define the laminar organization of the MEC, and to be able to compare its structure to other cortical regions, we therefore asked if differential gene expression distinguishes the superficial from deep layers or clarifies laminar borders within the deep layers (Figs. 4 and S4).
We first identified 159 genes specifically expressed in layers II, III or V/VI (see Materials and Methods and S4A-C Fig.). We refer to these genes, which show no consistent expression in other layers, as layer-specific genes (see Materials and Methods). We initially analyzed deep layers (V—VI) together because their divisions and borders are not easily distinguished by cytoarchitectonic criteria. Since genes with layer-specific expression patterns are of particular interest as neuroscience tools for isolating laminar functions, we examined their likely validity. Almost all layer-specific genes could be detected in our RNA-Seq analysis (149/159 with mean FPKM ≥ 0.1) and 62/159 had substantial levels of expression (mean FPKM ≥ 10). We also identified a further set of 622 genes, which we define as strongly differentially expressed (DE)(see Materials and Methods, S4B Fig.). These genes are expressed at higher levels in at least one layer than another, but are not necessarily exclusive to one layer. Both layer-specific and DE genes show consistent expression patterns across mediolateral sections (S6 Fig.). Only 37 of the layer-specific genes and 144 of the DE genes are amongst the 1000 most viewed genes in the ABA [50]. Thus, layers of the MEC can be distinguished by layer-specific and DE genes, many of which have received little previous attention suggesting they may represent new targets for future exploration.
Further examination of genes specific to the deep layers revealed three separate divisions of layers V and VI. First, a narrow zone at the deep border of layer IV is distinguished by expression of 5 genes, including Etv1, Grp and Nts (Figs. 4A, S4D). A second narrow zone of cells that is adjacent to the white matter is delineated by the expression of 8 genes, including Jup and Nxph4 (Figs. 4A, S4D). Finally, the wide intervening region is distinguished by 20 genes, including Thsd7b, Cobll1 and Col5a1 (Figs. 4A, S4D). Because layer V has been suggested to have a narrow superficial and wider deep zone [40], we refer here to the two more superficial subdivisions as layer Va (narrow) and Vb (wide), and we refer to the layer bordering the white matter as layer VI [40]. This delineation of layers Va, Vb and VI is supported by patterns of expression within the deep layers for the larger set of DE genes (layer Va (n = 24), Vb (n = 55) and VI (n = 13); S4E Fig.). A further 27 DE genes are expressed in both layer Va and VI, but not Vb (S4E Fig.). Thus, patterns of gene expression enable differentiation of divisions within the deep layers.
Previous cytoarchitectonic and electrophysiological studies have indicated that within layer II a subset of cells are clustered in islands [51–54]. In mice, neurons within islands express calbindin, whereas neurons outside islands express reelin [22, 53, 54]. Cells in these islands are of particular interest as they differ from reelin-positive cells in both their electrophysiology and projection targets [22, 53, 54]. Taking all DE genes within MEC, we found 30 genes that within layer II are predominantly expressed in apparent islands (Figs. 4A, S4F). Of these, just 8 are also specific to layer II within the MEC (S4G Fig.), including the calcium-binding protein, calbindin (Calb1). A further 19 are strongly expressed in the MEC deep layers but not layer III. The island genes include 13 genes that are expressed in the wedge-shaped patches of presumed parasubiculum adjacent to dorsal MEC, 3 of which we identified earlier (e.g. Mrg1, S4F Fig.). We also identified 37 genes with the converse, ‘Inter-island’, pattern (S4F Fig.). Of these genes 23 are specific to layer II, including Reln (reelin) and Il1rapl2 (S4G Fig.). A further 11 are also strongly expressed in the MEC deep layers, in particular in layer Va. The remaining layer II-specific genes do not appear to be uniquely expressed in either island or inter-island regions (S4G Fig.). Thus, differential gene expression distinguishes cell populations within layer II, shows that cells within and outside islands may be distinguished from cells in other layers by expression of common genes and provides evidence of similarities between the layer II island cells and parasubiculum.
What is the relationship between laminar organization of the MEC and other regions of cerebral cortex? While MEC has greatest similarity in gene expression to neocortex and also shares a similar laminar structure, classic ontogenetic evidence indicates that the two regions are developmentally distinct [24]. However, it is unclear if these ontogenetic differences are associated with later molecular differences between specific layers of the mature cortices.
To address this we first systematically examined overall expression in visual cortex and somatosensory (SS) cortex of genes with layer-specific expression in MEC. When we averaged expression of all genes selectively expressed in layers V/VI of MEC we found them to have a similar laminar organization in neocortex (Fig. 5A). In contrast, mean expression in neocortex of genes localized specifically to layer II or III of the MEC has a less distinct laminar organization (Fig. 5A). To quantify these differences we measured the distribution of gene expression intensity as a function of distance from the corpus callosum to the pial surface (Fig. 5B, Materials and Methods). We found that the three groups of MEC layer-specific genes have differing expression patterns in neocortical regions (Figs. 5C, S5A; Mixed Model Analysis, F = 12.3, p < 0.001). To assess the degree to which the laminar organization of each group of MEC layer-specific genes is maintained in neocortex, we first calculated the ratio of their expression in deep layers (V and VI) to superficial layers (II-IV). We then calculated the difference between these ratios and their expected values of 1 for deep genes, and zero for superficial genes. This difference was significantly smaller for deep layer-specific genes compared with superficial layer-specific MEC genes (Fig. 5D; MANOVA, p = 0.002 and p = 0.004 for effect of MEC layer-specific group in visual and SS cortex respectively). A similar relationship is apparent when we consider the expression patterns of individual genes (S5B, C, D Fig.). Around 92% of deep layer-specific MEC genes are expressed in the neocortex and 61–64% of all these genes are enriched in deep visual and SS cortex, respectively (S5C Fig.). In contrast, 77% of superficial layer genes are expressed in SS or visual cortex, but just 28–43% are enriched in superficial layers (S5C Fig.). Thus, our analysis of layer-specific genes supports the idea that deep layers of MEC have greater similarity to neocortical regions than superficial layers.
To investigate whether the relationship between layers of the MEC and neocortex extends beyond layer-specific genes, we took all genes in the re-registered ABA data set and examined the correlations in pixel intensity between layers in different cortical regions (Fig. 5E). MEC deep layers together correlate most strongly with neocortical layer VI (r = 0.96,0.94, p < 2.2 × 10-16), while layer II and III of MEC are more strongly correlated with neocortical layer V (r = 0.93–0.95, p < 2.2 × 10-16) than II or III (r = 0.90–0.91, p < 2.2 × 10-16) (Fig. 5E). To establish whether these correlations differ from those between neocortical regions, we investigated correlations between visual and SS cortices across all genes. We found that all corresponding layers correlated strongly (r > 0.96, p < 2.2 × 10-16) (Fig. 5E). Similarly, when we examined expression patterns of SS cortex layer-enriched genes [34, 36], we found a similar laminar organization of expression between SS and visual cortex (S5E Fig.). Thus, while laminar organization of gene expression is maintained between neocortical regions, gene expression within superficial layers of MEC, in particular, diverges from corresponding layers of neocortex.
Given the overall similarity between gene expression in deep layers of MEC and neocortex, we examined possible relationships between particular deep layers in each region. Of genes specifically expressed in particular deep layers we found that MEC layer VI-specific genes are almost always also expressed in layer VIb of neocortical regions (n = 7/8; Figs. 5A, S5D) (c.f. [55]). Meanwhile, layer Vb-specific genes are more commonly expressed in layer VIa of neocortex than layer V (S5D Fig., n = 15 vs 7 / 19). Moreover, MEC deep layers together correlate most strongly with neocortical layer VI (r = 0.96,0.94, p < 2.2 × 10-16), and more weakly with layer V (r = 0.94, 0.91, p < 2.2 × 10-16)(Fig. 5E). This is consistent with our observation that MEC layer Vb genes are more commonly expressed in layer VIa of neocortex than layer V, as the layer Vb region occupies the majority of the area of the MEC deep layers. Thus, at the level of gene expression MEC layers Vb and VI can be considered most closely related to neocortical layers VIa and VIb, respectively.
In summary, our analysis provides molecular evidence for an organization in which deep layers of MEC and neocortex implement similar gene expression programs, whereas superficial layers of MEC and neocortex express more diverse sets of genes.
Our analysis suggests specialized gene expression in different layers of the MEC. If this reflects an underlying functional organization then it could be reflected in the functions associated with layer-specific or DE genes. Given known differences in electrical intrinsic properties, morphology, connectivity and organization of cells between MEC layers [4, 6, 40, 45], we hypothesized that genes involved in cell excitability and communication might be differentially expressed across layers. To test this, we focused initially on DE genes as their greater number gives more statistical power in identifying over-represented gene attributes. We identified Gene Ontology (GO) annotations and pathways that are overrepresented amongst DE genes (n = 722 Ensembl-identified genes) relative to all genes expressed in the MEC (n = 9,057 Ensembl-identified genes). To reduce redundancy and identify diverse functions of interest, we clustered enriched terms into groups. Consistent with our prediction, genes associated with neuronal projections (n = 75, padj = 1.85 × 10-9), particularly synapses (n = 45, padj = 4.57 × 10-5), and those involved in calcium ion binding (n = 61, padj = 2.72 × 10-7), cell adhesion (n = 54, padj = 9.48 × 10-8), and axon guidance (n = 23, padj = 3.14 × 10-4) are overrepresented amongst DE genes (Fig. 6A). We also found strong enrichment of genes involved in ion channel activity (n = 46, padj = 7.05 × 10-7) and synaptic transmission (n = 25, padj = 2.36 × 10-4) (Fig. 6A). Amongst ion transport-related genes, cation channel activity (n = 34, padj = 3.82 × 10-5) is particularly enriched whereas anion channel activity is not. We asked if attributes enriched among DE genes were also identifiable amongst layer-specific genes. In addition to being significantly overrepresented amongst all DE genes, cell adhesion, axon guidance and calcium ion binding-related genes were also significantly overrepresented amongst the group of layer-specific genes (Fig. 6B). Given critical roles of these genes in neuronal signaling, these data support the idea that laminar differences in gene expression within the MEC support laminar organization of computations within MEC microcircuits.
Are genes within the functional groups that are overrepresented amongst DE genes enriched in particular layers or are they distributed across layers? Comparison of expression patterns for individual genes revealed genes with enriched expression in each layer (Fig. 6C, Materials and Methods). This analysis highlights a number of genes of potential functional importance. For example, ion channel-related genes include the potassium channel subunits Kcna4 and Kcnmb4, which control excitability and are enriched in layers II and Vb respectively, while axon guidance/adhesion-related genes enriched in layer II include Lef1, Lhx2 and Dcc as well as the ephrin receptor gene Epha4 (Fig. 6C). A possible role for the latter genes could be to control guidance of axons to newborn granule cells in the dentate gyrus [56]. Cell adhesion-related genes are also selectively expressed and significantly overrepresented in all layers and include several cadherins and protocadherins (Fig. 6B-C). These data suggest that subsets of each functional group of DE genes are expressed in each layer.
Together these data reveal candidate categories of genes that are most likely to distinguish the functions of different layers within the MEC. Our analysis also identifies molecules with highly specific laminar expression that could contribute to particular electrical and synaptic properties.
Topographical organization of intrinsic features along the dorsoventral extent of the MEC has received considerable interest because the characteristics of grid cells vary systematically along this axis [9, 10, 12, 13, 57]. The extent to which gene expression parallels this organization is not currently known. We took two approaches to addressing this issue, one using our re-registered ABA dataset, with its advantage of high spatial resolution, and the other using RNA-Seq analysis, which enables quantification across a wide dynamic range and the ability to test the reproducibility of gradients. This combined approach therefore enabled us to question not only dorsoventral differences in gene expression, but also their laminar organization.
We first calculated the ratio of pixel intensity between dorsal and ventral regions in images from the re-registered ABA dataset (Fig. 7A). We defined genes with at least 20% more expression in the dorsal than ventral area as being expressed higher dorsally (D>V) and those with at least 20% more in the ventral area as being expressed higher ventrally (V>D) (see Materials and Methods, S7A Fig.). As a result, we identified 3,188 D>V genes compared with 1,352 V>D genes (Fig. 7B). We next used RNA-Seq analysis to compare gene expression from microdissected regions of dorsal and ventral MEC. This also identified genes with dorsoventral differences in their expression (Fig. 7C), of which 1,467 D>V genes and 1,198 V>D genes satisfied our criteria of 20% more expression in one of the areas than the other. Of these genes 452 and 347, respectively, had statistically significant differences in expression (Cuffdiff 2 [58]: FDR < 0.05) across 4 replicate samples (Fig. 7C).
To establish whether similar populations of dorsoventrally expressed genes are identified by RNA-Seq and in the re-registered ABA dataset, we correlated the ratio of dorsal to ventral expression determined by each method. First, to avoid confounds from genes with different expression between layers, we focused on genes expressed in only one layer. We found that measures of differential expression are strongly correlated between ABA and RNA-Seq datasets for layer II, III and V/VI (Fig. 7D, LII: slope = 0.85, r = 0.74, p = 0.0007, LIII: slope = 1.5, r = 0.97, p = 0.0002, LV/VI: slope = 2.4, r = 0.77, p = 0.0038). Second, we compared gene expression for all genes found to be significantly differentially expressed across biological replicates in RNA-Seq data. We again found a significant correlation between the datasets (S7B Fig., r = 0.52, p < 2.2 × 10-16).
Do dorsoventral differences in gene expression manifest differently across layers? Average images indicate that layer II has the strongest D>V pattern, while the deep layers have the strongest V>D pattern (Fig. 7B). To test this, we compared the average ratio of ventral to dorsal expression for all layer-specific genes. We found significant differences in the ventral to dorsal ratio for deep layer-specific genes compared to layer-II or III-specific genes (1-way ANOVA F = 7.47, p = 0.0008. Post-hoc Tukey’s HSD LII vs. Deep p = 0.0016, LIII vs Deep: p = 0.010), with deep layers enriched for a V>D expression pattern (Fig. 7E). Indeed, while 20.6% of layer II and 14.3% of layer III-specific genes show significant D>V expression, only 1.8% of deep layer genes do (Figs. 7D, S7C).
Together our data provide convergent evidence for systematic organization of gene expression along the dorsoventral axis of the MEC, identify dorsally and ventrally enriched sets of genes, and suggest differences in the laminar organization of dorsoventral gradients.
Are the roles of genes with differential dorsoventral expression related to the cellular and system-level organization of function in the MEC? Taking all significant D>V and V>D genes identified by RNA-Seq, we investigated their possible functions using a GO and pathway analysis. By using clustering to distinguish enriched terms into key groups of interest (see Materials and Methods), we found that D>V genes are enriched for a number of attributes, particularly axon ensheathment (n = 15, padj = 2.08 × 10-7) and channel activity (n = 32, padj = 2.81 × 10-7) (Fig. 8A). We next used the re-registered ABA data set to examine the expression patterns of the identified gene groups. The D>V pattern found using RNA-Seq is replicated for the majority of axon ensheathment- (n = 9/12) and channel activity-related genes (n = 19/24) that show expression in the re-registered ABA dataset (Fig. 8B). We then investigated the layers in which gradients are strongest. Axon ensheathment genes show consistent D>V gradients in the superficial (n = 11/12) and deep (n = 9/12) layers (Fig. 8B). In contrast, genes involved in channel activity are more likely to show D>V gradients in the superficial (n = 22 / 24) than in the deep (9/24) layers (Fig. 8B).
In contrast to D>V genes, genes with V>D expression in the RNA-Seq data set are most strongly enriched for the neuroactive ligand-receptor pathway (n = 24, padj = 4.04 × 10-13), which is related to G-protein coupled receptor activity (n = 35, padj = 2.2 × 10-9), and for the extracellular region (n = 44, padj = 2.16 × 10-9) (Fig. 8C). Of the 11 identified neuroactive ligand-receptor pathway genes that are expressed in our re-registered ABA data set, 7 have consistent overall V>D patterns in the ABA (Fig. 8D) while 3 show no detectable expression. A total of 15/24 extracellular region-related genes are consistent with ABA data (Fig. 8D), with a further 12 showing no detectable expression. A possible reason for the discrepancies between ABA and RNA-Seq measures is that in the ABA analysis of the overall V>D gradient, V>D gradients that are only present in the deep layers may go undetected. This is because in some ABA images, ventral deep layers become narrower towards the medial border of the MEC and therefore ventral gene expression may be overshadowed by expression in the superficial layers or may not be present in the image. In support of this, we found that most V>D gradients found using RNA-Seq could be detected in ABA data when gene expression was specifically measured in the deep layers (Neuroactive: 8/11, Extracellular: 20/24, Fig. 8D).
Pathological changes in the MEC have been observed in a number of neuro-developmental and neurodegenerative disorders. Whereas layer III appears to be most consistently affected in epilepsy patients and in animals models of epilepsy [29, 30], cell number and disrupted organization within layer II are consistently reported in Alzheimer’s disease (AD) [51, 59], as well as in Huntington’s (HD) and Parkinson’s disease (PD) [60], schizophrenia [61] and autism [62]. This laminar specificity suggests that particular features of these layers, whether genetic or network-based, hard-wired or experience-driven, confer vulnerability. One possibility is that genes with mutations causally linked to particular disorders have layer-enriched expression. Alternatively, broadly expressed causal genes might cause specific pathology in layers with enriched expression of genes that confer vulnerability.
To address whether normal adult gene expression in the mouse provides insight into vulnerability, we first explored the laminar expression patterns of genes involved in signaling pathways that are disrupted in disease. Images showing the average expression pattern of genes involved in KEGG neurodegenerative disease pathways (AD, HD and PD) [63] indicate high expression in layer II, particularly in dorsal regions (Fig. 9A). To test whether this reflects significant enrichment of neurodegenerative disease pathway genes in layer II, we took all DE genes that exhibit high expression in layer II (see Materials and Methods) and compared representation of disease-related genes to their representation in the MEC as a whole. We found that AD, PD, and HD pathway genes are all overrepresented amongst layer II-enriched genes (Fig. 9C, Exact Fisher Test with Benjamini-Hochberg correction: Log2 Fold Enrichment > 1.19, padj < 0.024). Thus, basal gene expression may confer vulnerability of layer II in neurodegenerative diseases.
In the absence of KEGG pathway information, we used several database resources to identify genes related to schizophrenia [64, 65], autism [66] and epilepsy [67] (see Materials and Methods). Average images reveal weak, if any laminar organization for schizophrenia and epilepsy-related genes, with some evidence of layer II enrichment for autism-related genes. However, an enrichment analysis shows that schizophrenia-related genes are overrepresented amongst layer III-enriched genes (Fig. 9C), while both autism- and schizophrenia-related genes are enriched amongst RNA-Seq defined D>V genes (Fig. 9C), suggesting that pathology related to these diseases could show dorsoventral differences.
Since layer II enrichment of AD pathway genes corresponds with layer II vulnerability to AD, we explored whether genes with variants that have been established to confer increased risk of AD show layer-specific expression. AD possesses several key genetic risk factors, namely APP, PSEN1, and PSEN2 [68], but meta-analyses of genome-wide association studies have also shown that ApoE, ABCA7, Clu, Bin1, Cd33, Cd2ap, Epha1, Ms4a6A-E, Picalm, Sorl1, Ptk2b, NME8, FERMT2, CASS4, Inpp5d, Dsg2, Mef2c and Cr1 are strongly associated with late-onset AD [69–71]. We found that 16 out of the 20 of these genes that are in our re-registered ABA data set are expressed in the MEC. However, none are specifically expressed in layer II and only a minority show strong differential expression across layers (Fig. 9D). Pathology in layer II is therefore unlikely to be the result of layer-specific expression of AD risk genes. It could instead reflect enriched expression of signaling pathways linked to neurodegeneration in AD (Fig. 9A). Indeed, further analysis of the laminar expression patterns of AD pathway genes (Fig. 9E, Materials and Methods) reveals that almost all those with moderate laminar enrichment show highest expression in layer II, and that many (n = 16 / 26) are mitochondrion-associated genes, suggesting that cells in layer II may have higher energy demands than cells in other layers. This is consistent with the strong cytochrome oxidase staining observed in layer II [48]. Given that mitochondrial dysfunction is a feature of neurodegenerative disease [72] and that genes related to metabolism are altered in the MEC of patients with mild cognitive impairment [73] and AD [74], layer II vulnerability could be due to or compounded by enriched expression in layer II of signaling pathways that confer vulnerability to AD pathology.
To investigate the molecular organization of the MEC we combined a new pipeline for large-scale comparison of gene expression at high spatial resolution with RNA-Seq analysis. We show that differences in gene expression define the dorsal and ventral borders of MEC, its layers and its dorsoventral organization. We find that the MEC is closely related to neocortex through gene expression. This similarity is strongest for the deep layers, whereas superficial layers appear more specialized. Enriched topographical organization of genes related to synaptic communication and excitability indicates that laminar and dorsoventral organization of spatial coding within the MEC may have specific molecular substrates. Identification of laminar organization of AD-related pathways, but not risk genes, suggests that specific layers of the MEC may be particularly vulnerable to triggers of pathology in AD and other neurodegenerative diseases. The data sets generated by our study are a new resource for investigating molecular substrates for spatial coding and computation by the MEC and the structures with which it interacts, while the computational pipeline we have developed may have general applications for neuroanatomical investigation requiring comparison of many probes at high spatial resolution.
By developing a pipeline for automated comparison of brain sections at 10 µm resolution we were able to identify genes whose expression pattern delineates the borders and layers of the MEC (Fig. 3 and Fig. 4). Validation of this pipeline against RNA-Seq data indicates that relative expression levels estimated with the two approaches are consistent (Fig. 1 and Fig. 7). Recent work using double ISH labeling validates the layer-specific expression patterns we find for cadherins in the MEC [23], while other well-characterized genes such as reelin and calbindin [22, 75] also have expected expression patterns. Our analysis identifies a further 767 genes with layer-specific or enriched expression and 799 genes with dorsoventral expression. Nevertheless, our current analysis is limited by the availability of genes in the ABA data set (20,495 / 38,553 Ensembl genes, most of which are protein-coding), by the likelihood of false negative data in the ABA ISH data where true gene expression has been missed (estimated 3,297 / 14,054 by comparison with RNA-Seq data) and by limitations in image processing and registration accuracy that prevent us making use of the entire ABA data set. While our analysis is restricted to sections in parasagittal planes containing the MEC, it could be extended to include other brain regions through additional planes and to other species including humans [76]. In principle our approach could also be extended to analysis of images from three-dimensional datasets obtained using different methods [77, 78].
Our results resolve dorsal and ventral borders of the MEC, provide molecular evidence for laminar divisions of its deep layers, and identify numerous new molecular markers for the well-established separation of the superficial layers. While dorsal and ventral borders can be distinguished unambiguously in sagittal images, medial and lateral borders are better resolved in horizontal sections, so demarcation of these borders may require use of additional horizontal data sets. Delineation of deep layers is of particular interest as they are believed to relay hippocampal output to neocortex (c.f. [4]), but their organization and functional properties have received relatively little attention. We distinguish a narrow region deep to the laminar dissecans as layer Va, consistent with that described by [45]. We also identify a distinct division of the deeper layers into Vb and VI, a narrow region of cells that appears continuous with neocortical layer VIb (Fig. 4). We suggest that the divisions previously reported within layer V [6, 40, 79] correspond to the superficial layer Va and a deeper layer Vb that we identify here. Definitive laminar delineations within the deep layers will require analysis of shared gene expression, dendritic morphology and axonal connectivity. Our results also identify new markers for island cells and, to our surprise, suggest their similarity to neurons in the parasubiculum. It will be interesting to establish whether this similarity extends to functional properties [43].
Our analysis and data sets provide a resource for future functional investigation of laminar organization of functions in the MEC. This includes identification of markers for distinguishing cell populations (Fig. 4), particularly for layer III and the deep layers, for which there are currently few specific markers. Our delineation of layers Va, Vb and VI identifies several genes in each layer whose promoters may be usable for generation of driver lines to target that cell population. We also identify common expression patterns between MEC and neocortex that may underlie shared functional roles (Fig. 5). For example, the cortical layer VI-specific immunoglobulin heavy chain gene, TIGR accession TC146068 [34], also shows expression in MEC layer Vb. It is unlikely that this similarity in expression between deep layers of MEC and neocortex reflects biased selection of exemplar genes [80] as it is present when considering all expressed genes as well as those with laminar selectivity (Fig. 5). Instead, our analysis demonstrates that deep layers of MEC show greater similarity to corresponding cortical layers than do more superficial layers. Because our analysis includes the majority of protein-coding genes (Fig. 1), it also leads to novel predictions about gene expression underlying specialized function. As well as identifying candidates for electrophysiological differences between neurons from different layers [6], many cell adhesion and axon guidance molecules are enriched amongst patterned genes (Fig. 6). Of particular interest are cell adhesion-related genes such as Cdh13, Lef-1 and Dcc that show a similar expression pattern to Reln, which marks the subset of excitatory layer II cells that project to the dentate gyrus [22]. One possibility is that these genes play roles in forming connections with new born granule cells.
Genome-wide views of cortical organization can inform investigation of disease mechanisms by identifying convergent expression of molecular components of disease pathways [34, 81]. We found no evidence of layer-specific expression of genes causally implicated in disease pathology (Fig. 9). Instead, our analysis suggests that differential gene expression may underlie layer-specific pathology by predisposing specific cell populations to disease-causing mechanisms. For example, enriched expression of energy-related genes may reflect susceptibility of this layer to degeneration in AD. These functional and pathological predictions should be testable in future experimental studies.
The dorsoventral organization of the resolution of grid firing fields and the corresponding organization of excitable and synaptic properties of layer II stellate cells [9, 10, 12, 13] suggests that cellular mechanisms for grid firing may be identifiable by comparison of key features of dorsal and ventral MEC circuits. However, until now there has been little evidence for molecular differences that could underlie this organization (cf. [12]). We provide converging evidence from re-registered ABA data and from RNA-Seq data for systematic coordination of gene expression along the dorsoventral axis of the MEC. Consistent with key roles of superficial layers in the generation of grid fields, D>V gradients were most often found in layer II and III (Fig. 7). We also found evidence for genes with the opposite V>D pattern of expression, but these were most prominent in deeper layers, suggesting that control of dorsoventral differences by molecular pathways differs across layers. A potential caveat of our analysis is that dorsoventral differences in gene expression could reflect differences in the proportions of certain cell types. While approaches such as transcriptomic analysis of isolated cells will be required to resolve this, our finding that many layer-specific genes are not significantly differentially expressed along the dorsoventral axis, while dorsoventral genes have continuous rather than all or nothing changes in intensity (Fig. 8), argues for gradients reflecting coordination of gene expression levels within populations of a single neuron type.
By taking a genome-wide approach to differences in gene expression we obtained unbiased estimates of gene functions that are enriched among dorsoventral genes. Strikingly, we found enrichment among genes with higher dorsal expression of axon ensheathment and ion channel activity (Fig. 8). This is in accordance with previous evidence for dorsoventral differences in synaptic transmission and ionic conductances [12, 13, 15], and in immunolabelling for myelin [48]. Enrichment of 10–20 genes associated with each function indicates that the corresponding cellular differences may involve coordinated control of gene expression modules. For example, our analysis extends candidates for dorsoventral differences in excitability from HCN and leak K+ channels [12], to include non-selective cation channels such as Trpc5 [82] and voltage-dependent potassium channels such as Kcnq3 [83] and Kcnk1 (Twik1) [84]. Similarly, we identify myelin-related genes such as Mbp and Plp1, as well as related adhesion molecules such as Cntn2 (Tag-2), as candidates for dorsoventral differences in coordination of axon ensheathment [85]. Future gene manipulation studies will be required to establish causal roles of these genes in dorsoventral tuning of cell properties and of spatial firing. They may also provide insight into the role of topographic gene expression in the development and maintenance of topographical connectivity between the MEC and hippocampus. Additional investigation will also be required to establish whether dorsoventral coordination of transcription is complemented by similar coordination of translational and post-translational mechanisms.
Neurons in the MEC encode representations of space [9] that are critical for spatial learning and memory [86]. An unresolved question is whether this computation requires a specialized cortical circuit, or whether it is an example of a generic computation to which canonical cortical circuits can easily be adapted. Evidence for the former comes from findings that in layer II, which contains the highest density of cells with grid firing fields, excitatory stellate cells are only able to communicate indirectly via inhibitory interneurons [87–89], whereas in other cortical regions excitatory layer II principal neurons synapse with one another [90]. Consistent with this view our molecular analysis suggests considerable divergence between superficial layers of MEC and neocortex. In contrast, deeper layers of MEC appear much more similar to neocortex. Together with the dorsoventral organization of ion channel and axon ensheathment genes, our findings suggest that specialization important for spatial circuits is particularly striking within the superficial layers of the MEC. The functions within the MEC of the individual genes and functional gene groups that we identify as having laminar and dorsoventral organization have for the most part not been investigated and likely will be important targets for future exploration.
All animal experiments were carried out according to guidelines laid down by the University of Edinburgh’s Animal Welfare Committee and in accordance with the UK Animals (Scientific Procedures) Act 1986.
Brains were rapidly extracted from 13 male 8-week-old C57Bl/6JolaHsd mice and maintained in modified oxygenated artificial cerebrospinal fluid (ACSF) of the following composition (mM): NaCl 86, NaH2PO4 1.2, KCl 2.5, NaHCO3 25, CaCl2 0.5, MgCl2 7, glucose 25, sucrose 75), at approximately 4ºC. One 400 µm thick sagittal slice containing the right MEC was cut from each brain using a Leica Vibratome VT1200 system [91]. Dorsal and ventral regions were microdissected under a dissection microscope (S1A Fig.), with care taken to avoid inclusion of ventral entorhinal cortical regions, parasubicular or postrhinal regions and subicular regions. Tissue sections were collected into separate RNase-free eppendorf tubes before being quickly frozen on dry ice. Frozen tissue was stored at -80ºC for several weeks before RNA extraction.
We compared RNA from dorsal and ventral MEC of 4 groups of mice. To minimize the effects of inter-animal variability and variability in the dissection, whilst maintaining sufficient power to detect dorsoventral differences, samples were pooled with 3 or 4 mice in each group. RNA was extracted using RNeasy Lipid Tissue Mini Kit (Qiagen Cat:74804). RNA integrity was assessed using a Agilent 2100 Bioanalyzer. All sample RINs were between 7.1 and 8.5. cDNA was synthesized and amplified using the Ovation RNA-Seq System V2 (NuGEN Cat:7102) using 120 ng of starting material for each sample. The samples were fragmented and sequenced by the Ark-Genomics facility using Illumina HiSeq with multiplexed paired-end analysis on two lanes. Raw data were processed using Casava 1.8. Sequenced fragments were aligned using TopHat v2.0.8. After sequencing and alignment, absolute RNA expression and differential expression were computed using Cuffdiff 2 software on the output BAM files [58]. We chose Cuffdiff 2 to ensure accurate counting of transcripts in the presence of alternatively splicing. Reported gene expression therefore reflects the summed expression of all transcripts/isoforms of a gene. Cuffdiff 2 was run on the Edinburgh Compute and Data Facility (ECDF)[92] cluster on 4 cores each with 2GB of RAM. The reference genome used was Ensembl 73, downloaded 12th Nov 2013.
Transcripts were classified as expressed if their mean fragments per kilobase of exon per million fragments mapped (FPKM) [37] across samples ≥ 0.1 (c.f. [16]) in at least one of the dorsal or ventral regions (Fig. 1D). We also only considered transcripts for inclusion if Cuffdiff 2 analysis revealed them to have a minimum number of 10 alignments in a locus (default value). Transcripts were only tested for differential expression if mean FPKM across samples ≥ 1.
The steps for processing and extraction of data from ABA images are summarized in S1 Fig. and described in detail below. Code used in this section is available at https://github.com/MattNolanLab/Ramsden_MEC.
Image download from ABA. Images were downloaded from the ABA database using the application programming interface (API: http://www.brain-map.org/api/index.html). Since the ABA sagittal reference atlas begins at 3.925mm laterally, and as the MEC is located between approximately 3.125 and 3.5 mm laterally, images between 0 and 1400 μm (refers to distance from most lateral point) were selected for download for each image series. Two files were downloaded for each image: an ISH image file and a corresponding expression image file. Images were downloaded using the API files: http://www.brain-map.org/aba/api/imageseries/[enterimageseries].xml and http://www.brain-map.org/aba/api/image?zoom=3& top=0& left=0& width=6000& height=5000&mime=2&path=[path specified in xml file]. Images were approximately 500KB each. Approximately 120,000 images were downloaded in total and they were stored on a cluster provided by ECDF [92].
Preprocessing and cerebellar segmentation. ABA images, of variable dimensions, were first downsized by a factor of 1.25 and pasted onto the center of a new image of 1200 (width) x 900 pixels (height) using the Python Image Library. ISH images were then processed to improve image segmentation and registration (S1A Fig.: steps 2–5). No further changes were made to the expression image files until application of a segmentation mask (step 6).
Image preprocessing proceeded as follows. (a) Background subtraction was carried out on the ISH images using ImageJ [93](S1A Fig.), with radius set to 1 pixel as this is approximately the size of a cell at the chosen resolution. (b) Images were thresholded using the ImageJ ‘Min_error’ automatic thresholding method such that all visible objects in the image, including anatomical features and cells with very low staining, were retained. The aim of this step was to minimize gene expression-specific information in the images whilst retaining anatomical detail to facilitate image registration based on landmark features. (c) To aid feature extraction, a smoothing filter (ImageJ) was applied to the images to smooth them prior to processing.
Because the cerebellum could impair performance of the registration algorithm we developed an automated segmentation workflow to remove the cerebellar region from images prior to registration (S1A Fig.). (d) An edge detection algorithm was applied to background-subtracted images (FeatureJ Edge detection [94]). This image was thresholded to provide two outlined regions: the forebrain and cerebellum. These regions occasionally featured internal gaps caused by very low pixel intensity brain regions. We used an ImageJ algorithm to identify the two regions as objects (defined by the complete perimeter) and to fill in any such gaps (ImageJ/Process/Binary/Fill Holes). These regions could then be detected as separate objects, using the ImageJ particle analysis tool (ImageJ/Analyze/AnalyzeParticles), and only the largest object, corresponding to the forebrain, was subsequently included in the segmentation mask. This mask could then be applied to the ISH and corresponding expression images.
Segmentation failed for images with low ISH labeling (because of edge detection failures), where the cerebellum and forebrain overlapped (due to mounting errors), and where erroneous staining prevented typical boundary detection. It was not feasible to examine all images and check those in which segmentation had failed, so we developed a method for automatically detecting successful segmentation. For each image within an image series, we used a binary support vector machine (SVM) classifier with a linear kernel to classify image masks based on success. To classify images, it is first necessary to extract features of the image that represent the patterns found in them. We used the VLFeat toolbox in Matlab [95] and custom-written code [96] to extract scale invariant feature transform (SIFT) features [97] from the images. The toolbox extracts SIFT features at 4 different scales to provide a spatial histogram that contains information about the positioning of features in space (PHOW features). For each image a feature histogram containing 4000 values was generated for input to the classifier. The SIFT feature library was provided by [96]. We trained the SVM classifier on PHOW feature vectors from 800 correctly segmented images and 250 poorly segmented images and used a further 800 positive and 250 negative images for validation and tuning of the regularization parameter. The classifier was able to separate positive and negative validation images with over 98% accuracy with a tuned regularization parameter. We therefore used the SVM model with the same parameters to obtain a score for all remaining (~120,000) images that estimated their chance of success. The majority of images were assigned positive scores but we flagged any image with a score below 1 (11% of images) as being potentially erroneous.
Generation of reference images. To enable the extraction of information from 2D ISH images with precision, we generated reference images for five planes covering the medio-lateral extent of the MEC and its borders (S1A Fig.). The central image (C) was our primary data extraction image, images in the adjacent lateral plane (L1) supplemented this information, while images in more lateral (L2) and in medial (M1 and M2) planes were used for reference but not for data extraction.
Reference images were generated using hand-selected ISH images that were chosen based on (1) relatively uniform expression in the MEC, (2) good tissue quality, and (3) medium ISH staining intensity. Approximately 15–20 images were chosen for each of the 5 reference images (See Figs. 1A, S1A). Pre-processed images were rigidly aligned using an ImageJ plugin “Align Image by Line ROI” (http://fiji.sc/Align_Image_by_line_ROI) [93]. Given images in which the user has marked 2 corresponding points on each image, this plugin finds an optimal transformation (translation, rotation, scale) in closed form that aligns the images into the same location.
Images then underwent group registration using a Matlab library, the Medical Image Registration Toolbox [98] (Fig. 1A). Images were registered by two-dimensional non-linear deformation to one another, with the aim of finding the group match with the greatest similarity. We chose to group register 15–20 images to capture a sufficient degree of variance without requiring excessive memory (∼ 5GB RAM) or time (∼ 20 hours). A Gaussian filter with a window size of 13 and a standard deviation of 3 was applied iteratively three times to each, followed by contrast enhancement, to enhance image structures at the relevant spatial scale. We chose to use cubic β-splines to represent the possible class of transforms and mutual information as the similarity measure, as it is relatively resistant to differences in contrast. The output of group registration is a series of transforms that correspond to each image. We generated reference images by applying these transforms and then calculating the median of the transformed images (S1A Fig.).
Classifying images based on their medio-lateral location. The images downloaded from the Allen Brain Atlas could be assigned either to one of the five reference image groups or to a sixth group for images not containing MEC. To ensure the ~120,000 images were appropriately classified based on medio-lateral extent, we used classification to identify, for each image series, the image most similar to our central reference image, ImrefC. We used a Support Vector Machine (SVM) library for Matlab [95] with a linear kernel and binary classification. The SVM provided a score for each image across all image series reflecting the chance that the image was approximately in the same medio-lateral plane as ImrefC. We could then compare scores for all images that had been downloaded for a given image series and choose the image with the highest score. Images from each image series that corresponded to more medial and lateral reference images could then be identified based on their relative medio-lateral location (calculated using the ABA API database xml file corresponding to the relevant image series using the position and referenceatlasindex xml tags), since images within image series were always separated by 100, 200 or 400 μm.
The procedure for the SVM followed several stages. (a) A Gaussian filter was applied to preprocessed ISH images, with a window size of 15 and a standard deviation of 3. (b) We trained the SVM classifier with PHOW features extracted from images manually classified for 501 genes into mediolateral groups. (c) We optimized the regularization parameter of the SVM and tested performance with a further 483 genes. After training, all the remaining images were run through the SVM and assigned a score. The highest scoring image from each image series was then assigned to an ImrefC folder for manual inspection. Any images that did not belong in ImrefC were removed. Images that were distantly located from ImrefC images were placed in a No-ML folder. We manually checked this folder and any images that appeared to match a reference image were moved to the appropriate folder. Records were kept of all movements and this process of checking continued until we were satisfied that images had been assigned to a reference with ∼95% accuracy.
Registration gene images to a reference image. Each pre-processed ISH image was Gaussian-filtered and contrast-enhanced to facilitate extraction of large anatomical features and then 1-to-1 registered to its respective reference image, also Gaussian filtered, using the MIRT toolkit (S1A Fig.). 1-to-1 registration is unlikely to be as accurate as group registration but group registration would be unfeasible, in terms of memory and time required, for the number of images involved (∼ 20,000). Images were registered using cubic β-splines with mainly default settings, with the exception of the transformation regularization weight, which sets a limit on the scale of the deformation. We increased this from the default value of 0.01 to 0.1 to prevent large deformations. MI was used as a similarity measure because of its invariance to differences in contrast. The output of the algorithm was a transform describing the deformation of all points.
Apply registration transformation to expression images. The transform, calculated using the thresholded images, was also applied to the original ISH images (for visual assessment of registration success), as well as to the expression images (for extracting pixel intensity). All images underwent the same procedure.
Image quality check. To assess the accuracy of registration we used several measures automatically collected from all images: an MI-related score from the registration algorithm, a cross-correlation score on the final image, and a classification score from a classifier trained on poorly registered images. The MI score of the final deformation for each image reflects its similarity to the reference image. There is a clear distinction between the distribution of scores before and after registration (S1A Fig.). To determine how well these scores represent correct alignment, we chose a random sample of 100 images and manually rated their registration accuracy, then plotted their scores against the final mutual information result. This allowed us to set a threshold so that we could flag images that were potentially poorly registered. A total of 17% of images in the central plane were flagged compared with 14% in the L1 plane.
The MI score reflects registration accuracy across the whole image and therefore could overestimate the accuracy of registration in the MEC. Therefore as a second test we used the Matlab function (normxcorr2) to cross-correlate the MEC-containing region of a registered image with a larger region of the respective reference image. This cross-correlation function provides both a normalized maximum fit score and the location of the maximum fit, thereby enabling us to estimate the offset between the posterior MEC border within the registered image and within the reference image. We validated scores by using the image set used for validation of MI analysis. The majority of images were given a high manual rating of 5 and had a cross-correlation offset of near zero. A total of 7% of central images were flagged as having an offset that was potentially too large.
To detect image flaws including holes in the tissue created by bubbles, and aberrant detection of the pial surface as an RNA-expressing cell, which artificially increases the mean pixel intensity of the image, we used an image classifier. We decided to use image features to capture erroneous elements in the expression images that could subsequently be detected using an SVM. Features were extracted from the region including the MEC and immediate surround. We then trained a binary SVM classifier with a radial basis function kernel on all the images that we had visually assessed as having significant errors (n=∼ 50) against high quality images (n = 300). We used the LIBSVM package in Matlab for this [99]. We used cross validation to optimize the regularization parameter, C, and the hyperparameter of the radial basis function, gamma. We then tested all images with the classifier, giving a probability estimate that each image was erroneous. We again compared the probability estimate with visual assessment of a subset of the images used previously to estimate accuracy for the MI score, and flagged images with scores greater than 0.13. The classifier distinguishes images with large registration errors and pial surface errors from high-quality images (S1A Fig.). However, this method is naive to anatomy and not particularly sensitive to minor misalignment errors as it extracts features from the entire MEC region that are scale and alignment-invariant. 29% of images were flagged based on these results. Images were assigned error statuses and defined as not meeting the quality criteria if they had poor MI scores or were flagged by at least 2 of the other error measures. Error statuses were updated for all visually assessed images. In summary, 15,447 / 20,032 (77%) genes had images meeting these quality criteria in the central plane and 12,814 (64%) had images meeting the criteria in the L1 plane (S1B Fig.).
Extraction of pixel intensities from ABA images. Custom python scripts were written for all analysis of 8-bit ABA expression TIFF images. Expression images were used instead of raw ISH images because stages of processing that control noise, background illumination and contrast invariance across the images have already been performed as part of development of the ABA [34, 36, 100] (see ABA Informatics Data processing white paper). In addition, pixel intensity information represents overall expression level of individual cells that have been detected as expressing the gene of interest and should not contain structural information present in brightfield images that is not gene-specific, such as densely fibrous regions,
Genes were classified as being expressed if the mean expression in either the custom-defined dorsal or ventral region was ≥ 1 (scale up to 255). All data presented are based on pixel intensity values from 8-bit grayscale expression images. Expression images have been shown with a 16-color lookup table for visualization purposes. Average images are shown either based on absolute intensity or intensity normalized by the mean of the MEC region, as indicated.
Regional gene expression was estimated by manually outlining regions using Bezier lines (ImageJ), using the selections to create a binary mask (black on white pixels) that could be imported using custom Python or Matlab scripts, then using the mask to select elements of the original expression images.
Some genes are represented more than once in the ABA dataset, either because multiple probes have been used to detect different transcripts (n = 351 / 20,334 genes), or where ISH experiments with a single probe have been replicated (n = 1,011 genes). Where multiple probes are used to target a single gene we analyze images for each probe separately, but in population analyses we report the gene once whether only 1 probe or all probes were detected. For replications we found that relative intensities across different regions were similar in each image set, but overall image intensities could vary. We therefore analyzed average images generated by obtaining the mean relative pixel intensity across regions of interest for each image series and multiplying this by the mean pixel intensity of each whole image averaged across all relevant image series.
We extracted pixel intensity information from central and adjacent lateral images, which showed highly similar patterns of gene expression. Comparisons in mean pixel intensity between corresponding MEC layers of central and adjacent lateral images showed correlations of at least 0.945, which was higher than correlations with non-corresponding layers (< 0.938). When relative mean pixel intensities were compared between corresponding layers we found correlations of at least 0.66, compared with < 0.21 for non-corresponding layers.
Comparison of ABA and RNA-Seq expression. To compare the results from ABA data and RNA-Seq data, we first used the Ensembl Biomart database to match Ensembl data from RNA-Seq to Entrez IDs and Gene symbols. To correlate ABA and RNA-Seq expression, we took the mean pixel intensity of the dorsal and ventral regions, averaged across the ImrefC and ImrefL1 planes (where images were available) and compared this to the mean FPKM of RNA-Seq dorsal and ventral samples (S1A Fig.).
We describe below methods used for analyses associated with each main figure. Pearson correlation coefficients and linear regression analyses were performed using the statistics linear regression package, lm, in R. We define absolute intensity as the pixel intensity measurement in 8-bit images (range 0–255), and relative intensity as a measure reflecting the ratio of pixel intensities between two or more regions, for example layers or brain regions.
ABA regional comparisons and combinatorial analysis (Fig. 2). The neocortical, hippocampal, caudate putamen, amygdala, piriform and MEC regions were manually outlined using the central composite reference image and Allen Mouse Reference Atlas [34, 36] as guides. Genes with a mean pixel intensity in MEC ≥ 5 were defined as being MEC-enriched if their relative intensity compared to the comparison regions was ≥ 0.8. Genes with mean pixel intensity in MEC < 5 and ≥ 2 were defined as being MEC-enriched if their relative intensity was ≥ 0.99. In both cases the mean intensity of the compared brain region also had to be < 5. The proportion of MEC-enriched genes also expressed in the other brain regions was calculated by finding genes that did not satisfy these criteria. MEC-unique genes were identified using the same thresholds applied in comparison to all regions.
We identified pairs of genes with overlapping expression by first finding all genes that are enriched in the MEC relative to at least one other brain region and pairing them with each other. We then identified those pairs for which at least one of the pair was present in all five MEC-enriched lists. Images of genes were overlaid using ImageJ and manually inspected for degree of overlap. To aid visual assessment, genes with MEC expression < 5 were only included if expression was restricted to a particular subregion, as uniform expression at this intensity appears to most likely reflect non-specific ISH staining or uneven illumination across the tissue.
Detection of borders (Fig. 3). To identify genes defining the dorsal and ventral borders of MEC, we outlined regions dorsal and ventral to the approximate location of the borders using the central reference image (Fig. 3A, 3F). All gene images meeting the quality criteria in the re-registered data set with mean pixel intensity < 15 in one region and with a differential pixel intensity of > 15 between the regions were selected for manual validation. We supplemented this search with an ABA differential expression search [34, 36] (Target structure: Parasubiculum, Contrast structure: Medial Entorhinal Cortex), with an expression threshold of 3.5, which identified an additional 8 genes that demarcated the dorsal border of the MEC.
Identification of layer-specific and differentially expressed (DE) genes (Fig. 4). Genes were defined as layer-specific if they show no consistent expression in other MEC layers or differentially expressed (DE) if they show substantially higher expression in at least one layer than in another. Exact criteria for identification of layer-specific and DE genes are as follows.
(1) Comparison of laminar regions within the re-registered ABA dataset. Layers were defined using the composite central and adjacent lateral reference images (S4A Fig.). We primarily used data from the central reference image, only using information from the lateral plane when no central image was available. Since layers possess multiple types of cells that may themselves differentially express genes, we compared both average expression across layers, and the distribution of high-intensity expression. For each image a high-intensity pixel was defined as having intensity ≥ mean + 2 x S.D of all pixels in the MEC. The absolute pixel intensity (Lxabs = Lxrel x 3 x MECmean, where x refers to layer) and proportion of high-intensity pixels were then calculated for each of the layers II, III and V/VI. Relative laminar mean intensity (Lxrel) and relative proportion of high-intensity pixels (Lxprop) were calculated by dividing layer measures by their sum. We also identified the first (Lmax = maximum of (LIIrel + LIIprop, LIIIrel + LIIIprop, LVrel + LVprop)) and second highest expressing layers (Lmid (not Lmin or Lmax)). We then calculated an absolute mean pixel intensity difference between these layers (Lmaxdiff = ((Lmaxrel-Lmidrel)2 x Lmaxabs / 255) x 30) and the joint expression of the two highest expressing layers (Lmaxabsjoint = ((Lmaxrel + Lmidrel−Lminrel)2 x Lmaxabs / 255) x 30). These values acted to penalize low absolute pixel intensities. We did not initially distinguish layers V and VI as the border between these layers is not clearly defined in the sagittal plane. Only genes with a mean overall pixel intensity ≥ 1 and mean pixel intensity ≥ 2 in at least one layer were evaluated.
To quantify laminar differences in expression we calculated patterning scores (PS) as follows:
We assigned weights to the measures relative intensity (wrel = 0.4), relative proportion of high-intensity pixels (wprop = 0.5) and absolute pixel intensity (wabs = 1- (wrel + wprop)). If the image had no high-intensity pixels, wrel = 0.9.
We calculated a PS for single layer enrichment: PSsingle = wrel x Lmaxrel + wprop x Lmaxprop + wabs x min([Lmaxabs,1])
We calculated a PS for joint layer enrichment: PSjoint = wmean x (Lmaxrel + Lmidrel) + wprop x (Lmaxprop + Lmidprop) + wabs x min([Lmaxabsjoint,1])
Genes with PSsingle ≥ 0.65 or PSjoint ≥ 0.88 (n = 1,314) were marked as candidates for DE genes, including the subset of layer-specific genes.
(2) Cross correlation of genes within the registered data set. We also used the re-registered data set to find genes with similar patterns of intensity to genes identified through differential laminar expression, but that may have been missed in the previous analysis due to their occupation of very small areas. Using a SciPy cross-correlation function (Alistair Muldal:https://github.com/oleg-alexandrov/projects/blob/master/fft_match/norm_xcorr.py), we used Nxph4 as the seed gene to find other layer VI genes and Mrg1 as the seed to find other potential island genes. To identify genes only expressed in the narrow VI layer, we compared a small dorsal region and searched only images with mean pixel intensity ≥ 1 and < 5 that had a sum of squares difference (SSD) of zero with the target gene. We checked images for the 30 genes with the highest cross-correlation. For island genes, we searched using a small dorsal region including layer II and checked the top 50 genes with mean intensity ≥ 1 and SSD = 0.
(3) Identification of genes through ABA search tools. Since our aim was to make this resource as comprehensive as possible, we extended our search beyond our re-registered data set to make use of ABA differential expression tools and knowledge of cortex-enriched genes [34, 101]. We initially identified strongly differentially expressed ‘seed’ genes through manual exploration using the ABA differential expression tool, Fine structure Annotation and Anatomical Gene Expression Atlas (AGEA; http://mouse.brain-map.org/agea) (total checked = 922). Taking at least 8 genes with the strongest expression for each layer, we used the ABA NeuroBlast tool to identify all other genes with an expression correlation of at least 0.5 with any one of these ‘seed’ genes in the retrohippocampal (RHP) region. This provided us with over 4,000 potential DE genes, but no clear indication of their laminar expression profile. We visually assessed all those that did not have an image meeting the quality criteria in our re-registered ABA data set (n = 959). For the identification of layer-specific genes we also scanned all those that our analysis suggested had borderline differential expression (n = 163) or that had a NeuroBlast expression correlation greater than 0.7 (n = 1,288).
Given the potential for images to be poorly registered in both ABA and our re-registered data set, we also visually inspected differentially expressed neocortical genes (n = 302) using lists provided by [101](http://www.nature.com/nrn/journal/v8/n6/suppinfo/nrn2151.html) and those annotated as having a ‘high’ specificity score in the Somatosensory cortex Annotation on the ABA website (http://help.brain-map.org/download/attachments/2818169/SomatosensoryAnnotation.xls?version=1&modificationDate=1319171046372) [34, 36] to determine whether they also showed laminar specificity in MEC.
(4) Visual validation. To minimize the number of false positives in the data, all candidate layer-specific and DE genes were validated through visual inspection. Layer-specific genes were confirmed as showing consistent expression in a single layer using the original ISH image and, where possible, images in more lateral and medial planes. Genes with DE expression had to show consistently higher expression in at least one layer than in another. For ABA re-registered data, 703 / 1314 candidates could be added to the set of DE genes. An additional 4 genes with layer-specific expression were identified using cross-correlation and an additional 36 DE including 10 layer-specific genes using the cortex-enriched lists referred to above. The NeuroBlast data identified an additional 121 DE genes, 13 of which were layer-specific. During manual validation of DE genes, we also recorded particular patterns of gene expression, including island or inter-island expression and specific laminar expression within the deep layers.
To generate final lists of layer-specific genes, we included all genes visually validated as layer-specific, independent of PS score. For DE genes, we included all layer-specific genes, all genes in the re-registered ABA data set that had a single layer PS ≥ 0.65 or joint PS ≥ 0.88 and that were visually assessed as strongly differentially expressed, and all genes acquired using ABA tools and cortex layer-enriched lists that we validated as showing differential expression. See S4B Fig. Layer-specific and DE genes showed consistent expression patterns across mediolateral sections (S6 Fig.).
Analysis of laminar similarities and differences between MEC and neocortex in ABA data (Fig. 5). Taking the neocortical region used in our analysis of regional expression, we associated all pixel intensities for each gene image with a normalized location relative to the corpus callosum (or subicular border for MEC) and the nearest point along the pial surface. For all MEC layer-specific genes (S4C Fig.), we calculated mean pixel intensities at different normalized locations throughout the three cortical regions (Fig. 5B). We plotted histograms by binning the distances into 20 regions. Statistically significant differences in the laminar gene list expression patterns were detected using Mixed Model Analysis in SPSS (v21) with an unstructured covariance matrix. Fixed effects were the list of genes, location and their interaction. Random effects were the list of genes and location with image series as subject.
To test whether deep and superficial layer-specific expression patterns correspond between MEC and neocortex, we used genes previously identified as having ‘high’ specificity in SS cortical layers [34] (S5E Fig.) to divide the cortical regions into deep (layers V/VI) and superficial (II-IV) regions (cf. [101]). We also used these genes to estimate an approximate border between visual and SS cortex. For each gene with mean pixel intensity ≥ 2, we calculated the ratio of pixel intensity in the deep region to the superficial region. If gene expression in MEC and visual or SS cortex corresponds, we would expect this ratio to be 1 for MEC deep layer-specific genes and 0 for MEC superficial layer genes. To test this prediction, we subtracted the ratio for each gene from the expected value and performed a MANOVA using SPSS (v21) with type of MEC specificity as the between-subjects variable and the visual and SS results as dependent variables. Post-hoc tests were performed with Tukey’s HSD. The percentage of genes that are enriched in superficial or deep regions was calculated by including genes with a deep: superficial ratio of less than 0.4 or greater than 0.6 (50% difference). Genes with mean intensity < 2 in the neocortical region were not included in the analysis.
To calculate the correlations between cortical layers across all ABA genes, we used the estimated laminar boundaries described previously to generate binary array masks corresponding to each layer. Pixel intensities were averaged across all pixels within laminar boundaries then Pearson correlation coefficients calculated.
Functional analysis for genes with laminar organization (Fig. 6). We used the GOElite gene ontology tool [102] with recent versions of the GO OBO database and gene annotation file for mus musculus (24/1/2014) and Kyoto Encyclopaedia of Genes and Genomes (KEGG) database [63] to extract all enriched terms. For the reference set for DE genes, all DE genes and genes with weak differential or uniform expression and threshold mean intensity ≥ 2 in MEC (n = 9,057 Ensembl identifiers) were included. This ensured we were not simply selecting for brain-enriched genes but for genes differentially expressed within the MEC. Terms were defined as enriched if associated with a p value < 0.05 after a one-sided Fisher overrepresentation test followed by Benjamini-Hochberg false discovery rate adjustment for multiple tests.
To reduce redundancy and identify clusters of meaningful GO and KEGG terms we calculated a kappa similarity measure (described in [103]) to identify terms sharing a higher proportion of genes than chance. We then used hierarchical clustering (adapted from code written by Nathan Salomonis: http://code.activestate.com/recipes/578175-hierarchical-clustering-heatmap-python/) on the kappa similarity matrix to cluster terms with at least 5 genes and fewer than 100 genes into groups. Within each cluster we then extracted all kappa similarity scores. Only terms with less than modest overlap (kappa < 0.7) with a more significant term and fewer than 100 genes were presented in the summary figure (Fig. 6A). Clusters were labelled using their most significant term. To establish the significance of the over or under representation of layer-specific genes in the selected lists we used R implementation of a 2-way Exact Fisher test (fisher.test) followed by multiple corrections analysis (p.adjust.M). In heatplots, data are shown for images in the plane corresponding to the central reference image, where available, or for the adjacent lateral plane.
Investigating dorsoventral differences in gene expression (Fig. 7). For analysis of RNA-Seq data Cuffdiff 2 [58] was used to identify all differentially expressed genes with an FPKM of at least 1 and difference in expression of at least 20% (log2(1.2) = 0.2630). Differences in the direction of differential expression between genes with different layer-specific expression (Fig. 4) were detected using ANOVA followed by post-hoc Tukey’s HSD tests, performed in R.
For analysis of ABA images dorsal and ventral regions were manually outlined using the two primary custom reference atlases. Average pixel intensities within each region and across central and the adjacent lateral sections containing MEC were calculated for each ABA image series. Only ABA images with a threshold mean intensity ≥ 2 in the dorsal or ventral region were included in the differential expression analysis. We used a threshold log fold enrichment of 0.2630 (increase of 20%) to define differential dorsoventral expression, the same value used for RNA-Seq.
For comparison between Cuffdiff 2 analysis and ABA analysis, Cuffdiff 2 results were matched to ABA Entrez values or gene official symbols using the Ensembl Biomart tool. Two comparisons were made: one specific to ABA layer-specific genes, and the other only including genes determined to be significantly differentially expressed according to RNA-Seq data.
Functional analysis for dorsoventral genes (Fig. 8). A similar method was used to that described above for laminar genes. For the reference set we included all genes with an FPKM ≥ 1 in MEC that had a sufficient number of alignments in a locus (measured using Cuffdiff 2) to be tested for differential expression (Ensembl identifiers = 13,954).
To estimate the presence of gradients in MEC in re-registered ABA data, we divided the MEC region in both the central and adjacent lateral planes into 5 subregions along the dorsoventral axis. For each layer and subregion, we calculated the mean pixel intensity and used these values to calculate linear regression gradients (SciPy stats.linregress) for each layer and across the whole MEC. For heatmaps, genes were sorted according to this gradient and color-coded according to whether their dorsoventral difference was sufficiently large for them to be defined as D>V or V>D-expressed in the analysis performed in Fig. 6.
Disease analysis of layer-patterned genes (Fig. 9). Genes involved in pathways for Alzheimer’s disease, Huntington’s disease and Parkinson’s disease were acquired from the KEGG pathway database [63]. This tool contains manually drawn pathways of molecules known to be perturbed, either through environment or genes, in certain diseases, as well as molecules that are therapeutic markers or diagnostic markers. Autism-related genes include mouse genes for which the human homolog has a score of 1–4 (high-low evidence) or syndromic in the SFARI AutDB database (n = 238 (Ensembl)). Schizophrenia genes include the results of a computational analysis of meta-analysis results from [64, 104] (n = 175 (Ensembl)). Epilepsy genes include those associated with any form of epilepsy in the Disease database (n = 90 (Ensembl)) [67]. Genes that have been causally associated with early or late-onset AD were obtained from key reviews and meta-analyses in the AD literature [69, 72, 105].
To generate lists of layer-enriched genes we took all DE genes and identified those visually validated as being enriched in particular layers. We computed the significance of over or underrepresentation of disease genes amongst layer-enriched genes using Fisher’s Exact Test in R, as described for gene ontologies in Fig. 6. To determine the associated ontology terms of all AD pathway genes with at least moderate layer enrichment we used a PSsingle threshold of 0.60.
|
10.1371/journal.pgen.1001081 | Mutation in the Gene Encoding Ubiquitin Ligase LRSAM1 in Patients with Charcot-Marie-Tooth Disease | Charcot-Marie-Tooth disease (CMT) represents a family of related sensorimotor neuropathies. We studied a large family from a rural eastern Canadian community, with multiple individuals suffering from a condition clinically most similar to autosomal recessive axonal CMT, or AR-CMT2. Homozygosity mapping with high-density SNP genotyping of six affected individuals from the family excluded 23 known genes for various subtypes of CMT and instead identified a single homozygous region on chromosome 9, at 122,423,730–129,841,977 Mbp, shared identical by state in all six affected individuals. A homozygous pathogenic variant was identified in the gene encoding leucine rich repeat and sterile alpha motif 1 (LRSAM1) by direct DNA sequencing of genes within the region in affected DNA samples. The single nucleotide change mutates an intronic consensus acceptor splicing site from AG to AA. Direct analysis of RNA from patient blood demonstrated aberrant splicing of the affected exon, causing an obligatory frameshift and premature truncation of the protein. Western blotting of immortalized cells from a homozygous patient showed complete absence of detectable protein, consistent with the splice site defect. LRSAM1 plays a role in membrane vesicle fusion during viral maturation and for proper adhesion of neuronal cells in culture. Other ubiquitin ligases play documented roles in neurodegenerative diseases. LRSAM1 is a strong candidate for the causal gene for the genetic disorder in our kindred.
| Sensory motor neuropathies are diseases of the peripheral nervous system, involving primarily the nerves which control our muscles. These can result from either genetic or non-genetic causes, with genetic causes usually referred to as Charcot-Marie-Tooth (CMT) disease after the three clinicians who first described the key diagnostic markers. CMT patients lose muscle function, mainly in their arms and legs, with increasing severity during their lives. There are almost two dozen known genes that can mutate to cause CMT, and these fall into a wide variety of biochemical cellular pathways. We identified a group of patients with CMT from a small rural community, with good reason to suspect a genetic basis for their disease. Using high-throughput mapping and DNA sequencing technologies developed as part of the Human Genome Project, we were able to find the likely mutated gene, which was not any of the previously known CMT genes. Based on its sequence, the gene, called LRSAM1, probably plays a role in the correct metabolism of other proteins in the cell. Among the known CMT genes, some are also involved in protein metabolism, suggesting that this is a generally important pathway in the neurons that control muscle activity.
| Charcot-Marie-Tooth disease (CMT) comprises a set of genetically heterogeneous disorders of the peripheral nervous system, affecting motor and sensory function. CMT is the most common inherited neuromuscular disorder, with a wide range of clinical presentations, but as described by OMIM (118200), the salient features of CMT include a slowly progressive weakness and atrophy of the musculature, predominantly of the distal lower limb. This weakness often affects the patients ability to walk or run, and eventually can progress to reach the upper extremity. Within the broad group of patients defined clinically, there are various categories of CMT defined by neurophysiological subphenotypes, pathological findings on biopsy, modes of familial transmission, and specific mutated genes identified in individual patients. These criteria have been extensively reviewed in recent literature [1]–[17]. A query of OMIM for genes causing Charcot-Marie-Tooth yields 26 separate entries with allelic variants; the database of inherited peripheral neuropathies notes 31 gene entries for CMT plus an additional 7 described as causing distal hereditary motor neuropathy. Nonetheless, mutations in new genes associated with CMT continue to be reported[18].
The functions of genes whose mutation yields a CMT or closely related motor neuropathy phenotype span a wide range of disparate biochemical activities including structural components of myelin (PMP22, P0), a mitochondrial transport and fusion protein (MFN2), transcription factors (SOX, EGR2), components of protein degradation pathways (DNM2, RAB7, LITAF), tRNA synthetases (GARS, YARS), a nuclear structural component (LMNA) and others [19]. Thus, novel CMT genes are difficult to predict through selection of biological candidates for sequencing in unexplained patients. The best approach for identifying the genetic cause of unexplained CMT remains linkage mapping in multiplex families, with adequate statistical power dependent on the mode of transmission, the specifics of pedigree and local population structure.
We report the mapping of a novel form of autosomal recessive axonal CMT through homozygosity mapping in an extended consanguineous pedigree of a local founder population. The identified gene appears to play a role in vesicle metabolism, consistent with some other CMT genes.
In the course of clinical work, we ascertained a patient with Charcot-Marie-Tooth disease, most closely similar to subtype AR-CMT2 (recessive, axonal), although this clinical presentation has sometimes been included as a type of CMT4[16]. The index patient noted the gradual onset of weakness around age 20, particularly affecting his distal lower extremities, but also present in the hands. He noted episodic muscle cramping of extremity and trunk muscles. He lost the ability to run in his early 20s. He denied sensory symptoms. He had erectile dysfunction and urgency of urination, but no other autonomic symptoms or evidence of spasticity. At the time of examination he demonstrated bilateral pes cavus, with marked wasting of distal lower extremity muscles and mild wasting of hand intrinsic muscles. Fasciculations were present in upper and lower extremity muscles. In the lower extremities he had grade 4 out of 5 ankle dorsiflexion strength (MRC scale), grade 4 hand intrinsic muscle strength and other muscles were grade 5. He could not walk on either the toes or heels. There was no gait ataxia. Upper and lower extremity tendon reflexes were absent. He had mild loss of sensation on the fingertips and severe loss of sensation in the feet and legs, most markedly to vibration, but also involving proprioception and pain perception. Laboratory investigation demonstrated an elevated serum creatine kinase (CK) from 1082 to 1921 U/L (18-199 U/L). Nerve conduction studies and needle electromyography demonstrated a diffuse sensorimotor peripheral neuropathy. There was no evidence of a primary muscle disorder. The predominant electrophysiological pattern was consistent with axonal degeneration (see Table S1). Sensory nerve action potentials were small or absent. All of the upper extremity motor nerve conduction velocities were faster than 38 m/s. The ulnar compound muscle action potential amplitude was small and a repeat study 2 years later demonstrated both median and ulnar compound muscle action potential amplitudes were small with normal motor conduction velocities. These are accepted criteria for an axonal CMT [1]. Upper and lower extremity muscles demonstrated ample denervation and partial reinnervation, with fibrillation and reduced recruitment of large motor unit potentials. Denervation of paraspinal muscles indicated axonal degeneration was present at very proximal nerve levels. Temporal dispersion was seen in some motor nerve conductions, but no conduction block, which may be an indication of an element of secondary demyelination, but the predominant electrophysiologic pattern was axonal.
The proband is a member of an extended multiply consanguineous family derived from a rural eastern Canadian population isolate; the extended pedigree includes five additional affected individuals with similar suites of symptoms (Figure 1A). Other affected family members exhibited sensory and motor dysfunction with pes cavus. Autonomic symptoms have not been consistently reported. Weakness and wasting has usually been moderate and predominantly in distal lower extremity muscles. The onset of symptoms has usually been in early adult years. One patient was not aware of any difficulties, but had examination abnormalities in his 40's. Some of the affected individuals are able to ambulate into later years, though others have become wheelchair dependent. Sensory symptoms are sometimes not reported, but sensory examination is consistently markedly abnormal, with loss of vibration sense often up to proximal legs and hips. Proprioception loss has been severe in some affecteds with accompanying sensory ataxia. Laboratory abnormalities that are available in only a few patients include mild increased CSF protein and increased serum CK. One patient had significant essential tremor, but that has not usually been reported. When EMG data is available, the pattern is typically predominantly axonal degeneration with only mildly slowed or normal motor nerve conduction velocities and no upper extremity motor nerve conduction velocities slower than 38 m/s. One other patient had evidence of paraspinal muscle denervation, with a normal MRI of the spine, suggesting axonal degeneration at very proximal nerve levels from the neuropathy.
Based on transmission of the trait in the pedigree, the genetics are consistent with an autosomal recessive disorder. Given the isolation of the regional population, it seemed likely that all affected individuals in our cohort might be homozygous for the same causal mutation, sharing a chromosomal haplotype around the causal gene. We sampled DNA from six affected patients and related family members. We performed high density genome-wide SNP genotyping of five affected individuals. Formal linkage analysis using a recessive model was not deemed useful, given the highly consanguineous pedigree structure and also the impossibility of obtaining reliable marker allele frequencies for this small subpopulation. Instead, we used the homozygous haplotype (HH) method to test for linkage to any of 23 known relevant CMT loci. The HH method is a rapid non-parametric algorithm that utilizes the subset of completely homozygous markers in samples from affected individuals, and looks for consistent loci by excluding regions where affected individuals are homozygous for different alleles of a given SNP [20], [21]. The method is robust due to the high density of commercial genotyping panels. In this case, HH confidently excluded all of the known relevant CMT loci, under the assumption that all five affected individuals in our pedigree are homozygous for the same causal allele. HH flagged three chromosomal regions as potentially linked, on chromosomes 9, 15 and 17 (Figure 1B).
Subsequently we genotyped additional pedigree members including one more affected, and looked for regions of extended homozygosity shared identical-by-state (IBS) in the six affected individuals but not in unaffecteds. As shown in Table 1, among the longest series of consecutive homozygous SNPs, a region on chromosome 9 appeared as a clear outlier. This region corresponds to that predicted from HH analysis, and extends from rs1324475 at 122,423,730 Mbp to rs10987845 at 129,841,977 Mbp. It is interrupted by several single heterozygous SNPs, mostly in one particular sample; these presumably represent false heterozygote genotype calls. In contrast, the potential regions found by HH on chromosomes 15 and 17 were not homozygous in all six affected individuals when all marker data was considered. The likely linked interval is 7.42 Mbp in size on chromosome 9, and includes 84 RefSeq annotated genes, including a cluster of 14 olfactory receptor genes which were not considered likely candidates.
We prioritized genes likely to have neuronal or neuromuscular function based on manual review. In all we sequenced 314 coding exons of 18 genes (HSPA5, DENND1A, RABGAP1, RAB14, STXBP1, DNM1, SPTAN1, DAB2IP, LHX2, TOR1A, GSN, LHX6, LMX1B, CDK9, CDK5RAP2, FPGS, SH2D3C, LRSAM1), until we observed a particular homozygous variant in the gene LRSAM1 (Figure 2A). This variant changes a coding exon consensus splice acceptor AG dinucleotide to an AA. There are three RefSeq annotated isoforms of LRSAM1, differing in the 5′ noncoding region, generating transcripts of either 25 or 26 exons. All three splice forms predict the same open reading frame; the variant identified in our patients lies in the penultimate coding exon, either 24 (isoforms 1, 2) or 25 (isoform 3). The variant was found homozygous in all six affected individuals, and either wild type or heterozygous as expected among sequenced parents and siblings (Figure 1A). This variant is not present in dbSNP build 130 which includes 2 million novel SNPS recently submitted by the 1000 Genomes project, nor was it detected in any of 150 local control (a mix of anglo- and franco-phonic individuals) or 96 CEPH Caucasian control samples, totalling almost 500 control chromosomes. No other homozygous coding variants were detected by sequencing this set of candidate genes.
The variant in question changes the consensus splice acceptor site. We tested three splice site prediction programs (Berkeley Drosophila Project, NetGene2 and SplicePort) to see whether they were sensitive to the alternative site used in the homozygous patients. All three programs correctly predicted the bona fide splice acceptor site in the wild type sequence. The Berkeley tool failed to predict the alternative AG two nucleotides internally in the mutant sequence, while NetGene2 and SplicePort predicted this acceptor site though with low confidence. We were able to test directly whether splicing of the exon was altered, using total RNA extracted from a fresh blood sample from one affected patient (1702). By qualitative RT-PCR, we saw a product of the appropriate size in both a control sample and the affected patient sample, at roughly equivalent intensities (d.n.s.) Although the resolution of the electrophoresis was much less than single nucleotide, sequencing of the sample product from the affected patient showed that splicing was to the next AG directly following the true acceptor site, two bases into penultimate exon 24 (or 25 as per isoform 3) (Figure 2B). This causes an obligatory frameshift, leading to an altered open reading frame and premature truncation of the protein after 643 (out of 723) residues in all three spliced isoforms. The effect of this change on protein expression was tested directly by Western blot using EBV-transformed B-lymphocyte cell lines (B-LCL) derived from a healthy control and from one of the affected CMT patients. While a single strong band was detected by the anti-LRSAM1 antibody in the control B-LCL (molecular weight approximately 78 kDa), no protein was detected in the B-LCL derived from the CMT patient (Figure 2C). Either the truncated protein is rapidly degraded, or else is rendered non-reactive with our antibody. In either case, the result is most likely to be a significant loss-of-function of the gene product, although unusual gain-of-function effects of a truncated protein can be imagined (though these might be expected to behave in a dominant not recessive fashion).
LRSAM1, leucine rich repeat and sterile alpha motif containing 1, is predicted to be an E3 type ubiquitin ligase [22]. It is also known as TAL (TSG101-associated ligase) and RIFLE. TSG101 itself is a tumor suppressor gene, with a reported role in maturation of human immunodeficiency virus, and LRSAM1 is implicated in its metabolism directly by polyubiquitination. TSG101 is involved in retroviral vacuolar budding. Interestingly, another TSG101-ubiquitinating ligase is known, (Mahogunin, or MGRN1), for which knockout mice exist and exhibit a neurodegenerative phenotype. Moreover, the known CMT gene LITAF, also called SIMPLE, interacts with mouse ubiquitin ligase gene product NEDD4 [23], also potentially with TSG101 [24], and may itself be an E3 ubiquitin ligase [25], These related findings support the interpretation that mutation of LRSAM1 is probably causal in our patients. It remains to be determined whether the pathogenic effects of mutations in these protein degradation pathway genes act directly via specific neuron-specific proteins (such as PMP22) or more generally through decreasing cell viability.
The currently recommended diagnostic paradigm for Charcot-Marie-Tooth entails a complex flow chart combining clinical, familial and molecular genetic analyses [3]. While this approach makes sense when DNA sequencing technologies are cost-limiting, this mixed paradigm could soon be replaced by a more comprehensive and pre-emptive molecular analysis. With the advent of whole genome reagents such as all-exon hybridization capture oligonucleotide libraries, together with the tremendous cost-reductions in DNA sequencing using next-generation nanotechnologies, it should soon be feasible to sequence either entire patient genomes, or entire exomes, for less than the cost of traditional Sanger-based fluorescent capillary sequencing of sets of candidate genes [26]–[28]. We envisage an analysis paradigm whereby all patients with a potential genetic diagnosis, across any medical subdiscipline, may first be sequenced to identify likely pathogenic variants, which can then be cross-indexed with clinical parameters to flag likely causal genes. This approach has recently been shown to be feasible in a research context, including detection of a pathogenic variant in a family segregating a known form of CMT [29]–[32].
Approval for the research study was obtained from the Capital Health research ethics board. Patients were identified in the course of routine clinical ascertainment and treatment of movement disorders in the neurology clinic at the Halifax Infirmary. All sampled family members provided informed consent to participate in the study. DNA was obtained from blood samples using routine extraction methods.
Whole-genome SNP scanning was performed at the McGill University and Genome Quebec Centre for Innovation, using the Illumina Human610-Quadv1_B panel. Data were scanned using the Bead Array Reader, plate Crane Ex, and Illumina BeadLab software, on Infinium II fast scan setting. Allele calls were generated using Beadstudio version 3.1 with genotyping module. Data are generated in three different output formats, AB, Forward strand, and Top strand (as defined by Illumina). We used AB format for all linkage analyses.
Homozygosity haplotype (HH) analysis was performed according to the method of Miyazawa [21]. The source code of HH program was modified to customize the format of output. The parameter LARGEGAP defined in the header file, which is used to define large gap of two consecutive SNPs like centromere, was changed from the default value 300,000 bp to 400,000 bp to accommodate some non-centromere spaces for HumanHap610 genotypes. The revised C source code of HH program was compiled with GNU compiler on a Linux-based operating system Fedora. HH analysis requires a SNP annotation file, which includes SNP name, physical coordinates, genetic distances, and minor allele frequencies. The SNP annotation file provided by HH software is for the Affymetrix 500K GeneChips Human Mapping Array Set. The HH format annotation of Illumina HumanHap610 for CEPH population was created from the SNP annotations downloaded from Illumina website. The genetic distances of SNPs with empty value, inconsistent value, or zero were interpolated according to the physical coordinates of their flanking SNPs. HH analysis was performed with a cutoff value 3.0 cM. Homozygosity analysis was performed using customized scripts and manual inspection comparing samples from affected and unaffected pedigree members.
Annotated coding exons were amplified by PCR using standard methods, and sequenced at Dalhousie University, using Sanger fluorescent sequencing and capillary electrophoresis. Sequence traces were analyzed using MutationSurveyor (Soft Genetics, Inc.) Specific primers for amplification of LRSAM1 exons and PCR conditions are provided in Table S2.
EBV-transformed B-LCL cells derived from a healthy subject or CMT patient 1675 were cultured in RPMI with 10% FBS and 1% pen/strep in 5% CO2. Cells were pelleted and lysed in lysis buffer (50 mM Tris-HCL, pH 7.4, 150 mM NaCl, 2 mM EDTA, 0.2% Triton X-100 with 1 mM PMSF and protease inhibitor tablet (Sigma) added to ice cold buffer immediately prior to use). Cells were broken by vortexing for 1 minute. Cell debris was removed by centrifugation at 16000×g for 10 minutes. Protein concentration was determined by the Bradford method (Sigma). Samples were diluted to 6 microg/microL in lysis buffer, then to 2 microg/microL in sample dye (125 mM Tris-HCL ph 6.8 with 20% glycerol, 4% SDS, 0.04% bromophenol blue, 10% 2-mercaptoethanol). Samples were heated to 95°C for 5 minutes prior to separation of 50 ug sample on a 7.5% SDS-PAGE gel. Benchmark pre-stained protein ladder (Invitrogen) was included on the gel. Protein was transferred by wet transfer to methanol-wetted PVDF membrane in transfer buffer (25 mM Tris-base, 192 mM glycine). Membranes were blocked overnight in blocking buffer (5% skim milk powder, 0.05% Tween 20, in PBS pH 7.4). Anti-LRSAM1 antibody (abcam) diluted 1∶500 in blocking buffer was incubated overnight at 4 degrees. Blots were washed in PBS- (0.05% Tween 20 in PBS pH 7.4) 15 minutes plus 3×5 minutes. HRP labelled secondary anti-mouse antibody, diluted 1∶2500 in blocking buffer, was incubated for 1 hour at room temperature. Blots were washed as above. HRP was visualized using SuperSignal West Pico Substrate (Fisher Scientific) and exposing to X-ray film for 3-5 minutes. Protein transfer to the gel was confirmed by staining the PVDF membrane with Fast Green.
The URLs for the data and analytic approaches presented herein are as follows:
Online Mendelian Inheritance in Man (OMIM), http://www.ncbi.nlm.nih.gov/Omim/
UCSC Genome Browser, http://genome.ucsc.edu/
NCBI, http://www.ncbi.nlm.nih.gov/
Database of inherited peripheral neuropathies, http://www.molgen.ua.ac.be/CMTMutations/Home/Default.cfm
|
10.1371/journal.pmed.1002073 | Associations between Mental Health and Ebola-Related Health Behaviors: A Regionally Representative Cross-sectional Survey in Post-conflict Sierra Leone | Little attention has been paid to potential relationships between mental health, trauma, and personal exposures to Ebola virus disease (EVD) and health behaviors in post-conflict West Africa. We tested a conceptual model linking mental health and trauma to EVD risk behaviors and EVD prevention behaviors.
Using survey data from a representative sample in the Western Urban and Western Rural districts of Sierra Leone, this study examines associations between war exposures, post-traumatic stress disorder (PTSD) symptoms, depression, anxiety, and personal EVD exposure (e.g., having family members or friends diagnosed with EVD) and EVD-related health behaviors among 1,008 adults (98% response rate) from 63 census enumeration areas of the Western Rural and Western Urban districts randomly sampled at the height of the EVD epidemic (January–April 2015). Primary outcomes were EVD risk behaviors (14 items, Cronbach’s α = 0.84) and EVD prevention behaviors (16 items, Cronbach’s α = 0.88). Main predictors comprised war exposures (8 items, Cronbach’s α = 0.85), anxiety (10 items, Cronbach’s α = 0.93), depression (15 items, Cronbach’s α = 0.91), and PTSD symptoms (16 items, Cronbach’s α = 0.93). Data were analyzed using two-level, population-weighted hierarchical linear models with 20 multiply imputed datasets. EVD risk behaviors were associated with intensity of depression symptoms (b = 0.05; 95% CI 0.00, 0.10; p = 0.037), PTSD symptoms (b = 0.10; 95% CI 0.03, 0.17; p = 0.008), having a friend diagnosed with EVD (b = −0.04; 95% CI −0.08, −0.00; p = 0.036), and war exposures (b = −0.09; 95% CI −0.17, −0.02; p = 0.013). EVD prevention behaviors were associated with higher anxiety (b = 0.23; 95% CI 0.06, 0.40; p = 0.008), having a friend diagnosed with EVD (b = 0.15; 95% CI 0.04, 0.27; p = 0.011), and higher levels of war exposure (b = 0.45; 95% CI 0.16, 0.74; p = 0.003), independent of mental health. PTSD symptoms were associated with lower levels of EVD prevention behavior (b = −0.24; 95% CI −0.43, −0.06; p = 0.009).
In post-conflict settings, past war trauma and mental health problems are associated with health behaviors related to combatting EVD. The associations between war trauma and both EVD risk behaviors and EVD prevention behaviors may be mediated through two key mental health variables: depression and PTSD symptoms. Considering the role of mental health in the prevention of disease transmission may help fight continuing and future Ebola outbreaks in post-conflict Sierra Leone. This sample is specific to Freetown and the Western Area and may not be representative of all of Sierra Leone. In addition, our main outcomes as well as personal EVD exposure, war exposures, and mental health predictors rely on self-report, and therefore raise the possibility of common methods bias. However, the findings of this study may be relevant for understanding dynamics related to EVD and mental health in other major capital cities in the EVD-affected countries of West Africa.
| Little attention has been paid to how mental health, trauma, personal exposures to Ebola virus disease (EVD) relate to health behaviors in post-conflict West Africa.
This study was done to test the links between mental health and trauma and health behaviors related to EVD—behaviors that increase the risk of spreading EVD and behaviors that help to prevent the spread of EVD.
At the height of the EVD epidemic (January–April 2015), 1,008 adults selected at random from 63 census enumeration areas in the Western Rural and Western Urban districts of Sierra Leone were surveyed to examine the relationships between war exposures, post-traumatic stress disorder (PTSD) symptoms, depression, anxiety, personal exposure to EVD, and EVD-related health behaviors.
Individuals reporting greater intensity of depression symptoms and higher rates of PTSD symptoms also reported higher rates of behaviors that increase the risk of spreading EVD, while individuals reporting previous exposure to war or having a friend diagnosed with EVD reported lower rates of such behaviors.
Individuals reporting higher levels of anxiety, having a friend diagnosed with EVD, and greater previous exposure to war (taking into account their mental health status) reported increased rates of behaviors that help to prevent the spread of EVD, while individuals reporting higher rates of PTSD reported lower rates of such behaviors.
These findings suggest that in areas with a recent history of war or conflict, past war trauma and certain mental health problems are linked with health behaviors that are important in stopping the spread of EVD.
Considering the role of mental health, in particular depression and PTSD, in preventing the spread of disease may help fight ongoing and future Ebola outbreaks in Sierra Leone.
While this study is specific to the Western Area of Sierra Leone, its implications are important for countries in West Africa affected by conflict and EVD.
| In many post-conflict settings, years of protracted violence have left the populace exposed to war trauma and loss, which contribute to unmet mental health needs. A history of conflict in West Africa has destroyed health systems, undermined community relations, and reduced trust in state institutions, increasing risks for health crises and social unrest [1]. In Sierra Leone, the health system was decimated over the course of the 11-y civil war (1991–2002). An estimated 50,000 persons were killed, and over 20,000 children and adolescents were involved with armed groups [2]. In Sierra Leone today, health services remain underdeveloped, of poor quality, and fraught with barriers to accessing care [3]. Mental health and social services are nearly nonexistent [4].
Direct and indirect effects of war likely contributed to the ability of Ebola virus disease (EVD) to ravage the country in 2014–2015. This EVD outbreak in Sierra Leone is reported to have resulted in 8,704 confirmed cases (13,823 confirmed, probable, and suspected cases) and 3,589 confirmed deaths (3,955 confirmed, probable, and suspected deaths) [5] and laid bare the inadequacies of the health system, which was unable to manage the crisis at its peak. To prevent future EVD outbreaks in Sierra Leone, it is important to understand how unmet mental health and social service needs influence both EVD risk and health behaviors critical for EVD prevention.
Research has documented associations between war exposures, mental health problems, and impairments in functioning among survivors of trauma including war [6–11]. However, in the context of EVD in West Africa, associations between past trauma, ongoing mental health difficulties, and health behaviors remain unexplored. Drawing from theories examining adaptive coping to life stressors as well as research investigating disease risk behaviors in the context of HIV/AIDS [12–15], Fig 1 sets out a conceptual model whereby exposure to past trauma may affect both EVD risk behaviors and the uptake of prevention messaging and EVD prevention/health-promoting behaviors, and where this effect may be mediated by related mental health difficulties. Our hypotheses are grounded in previous research on mental health and the dynamics of infectious disease prevention. For instance, poor adherence to treatment for HIV/AIDS has been linked to depression as well as post-traumatic stress disorder (PTSD) and other anxiety disorders [16–20]. Similarly, we hypothesized that in the context of EVD, past trauma and depression would relate to poor problem-solving in the context of EVD. We theorized that PTSD and depression might contribute to hopelessness and a foreshortened sense of the future, which might increase the likelihood of risk behavior. Understanding the relationship between past war exposures, mental health, and EVD risk and EVD prevention behaviors might provide critical knowledge for EVD prevention in Sierra Leone and other post-conflict settings.
Institutional review board approval was obtained from the Harvard T. H. Chan School of Public Health (Protocol #15145, Approval #17) and the Sierra Leone Ministry of Health and Sanitation Ethics and Scientific Review Committee. With institutional review board approval, consent was obtained orally with a witness due to low literacy among participants. A local community advisory board, comprising community members and healthcare professionals, reviewed and advised the research.
This survey was conducted January–April 2015 in the Western Urban (including the capital, Freetown) and the Western Rural districts of Sierra Leone (Fig 2). These districts represent diversity in ethnic composition and degrees of war exposure and were the epicenter of EVD cases during the 2014–2015 outbreak, together accounting for over 40% of confirmed cases.
The sampling population included adults 18 y and older living in the Western Urban and Western Rural districts, including Sierra Leone’s capital. Participants were selected using multistage cluster sampling. Census enumeration areas (EAs) were the primary sampling unit (see Fig 3). A list of EAs for the two districts and maps defining the EA boundaries were obtained from Statistics Sierra Leone [21]. Sixty-three EAs were randomly sampled from the list. Sixteen households within each EA were selected using random geographic sampling techniques. For each EA, a map was provided by Statistics Sierra Leone with a specified number of streets (2–5) indicated encompassing the EAs for the Sierra Leone 2004 Population and Housing Census. On those streets, the interviewers then selected a proportional sample of equally distanced households (with “household” defined as persons residing together). Each household was approached, and among adults who happened to be available on first contact, one was chosen at random from the household by alphabetizing first names in ascending order and choosing the first one. When a selected individual was unable or refused to participate, another individual within the selected household was randomly selected. After three attempts over the course of one day, if no household member was available, another household was selected using the same geographic randomization techniques.
The EVD risk behaviors outcome (Cronbach’s α = 0.84) was measured by 14 three-point items assessing likelihood (0 = very unlikely, 1 = somewhat likely, 2 = very likely) of engaging in certain behaviors if a family member or oneself experienced EVD-like symptoms (fever, fatigue, malaise and weakness, reddened eyes, joint and muscle pain, headache, nausea and vomiting). Such risk behaviors included waiting to see if the symptoms would go away, trying to treat the symptoms at home, and trying treatments such as hot salt water baths (see S1 Table for full list of risk behavior items). The EVD prevention behaviors outcome comprised a set of 16 items (α = 0.88) scored on a four-point frequency scale (0 = never, 1 = sometimes, 2 = often, 3 = always); EVD prevention behaviors included frequent hand washing, seeking EVD prevention information from a healthcare worker, and avoidance of large public events, public transport, and touching dead bodies, consistent with prevention messaging being used in Sierra Leone at the time of the study (see S2 Table for full list of prevention behavior items).
Independent variables (predictors) included an adapted version of the War Trauma Questionnaire, comprising 16 items, used in Liberia and the Democratic Republic of the Congo [22,23] to assess conflict-related experiences (eight items) and witnessing experiences (four items). Based on frequency of occurrence in the sample, item–total correlations, and internal consistency estimates, the 16 war exposure items in the original questionnaire were reduced to eight items addressing five victimization experiences (being forced to flee, having one’s home destroyed, being physically beaten, being forced to work or commit violence, being threatened with death) and three witnessing experiences (looting and destruction, beating or torture, killing of a household member). These were coded as 1 = occurrence or 0 = no occurrence and combined into one scale (α = 0.85). Anxiety and depression were assessed using a Krio (the lingua franca of Sierra Leone) adaptation of the Hopkins Symptom Checklist-25 (HSCL-25), scored on past week symptom intensity (1 = not at all, 2 = a little, 3 = quite a bit, 4 = extremely), previously adapted for Sierra Leone [24–26]. Both the depression (α = 0.91) and anxiety (α = 0.93) subscales had excellent internal consistency. To assess symptoms of PTSD, participants responded to 16 items (0 = no, 1 = yes) about the most distressing event they had ever experienced or witnessed, using the PTSD Symptom Scale–Interview adapted for use in Liberia (see S3 Table; α = 0.93) [27]. To assess personal EVD exposure, participants were asked five questions to determine whether a household member, a family member whom they did not live with, a friend, a neighbor, or someone in their community had been diagnosed with EVD over the past 12 mo (0 = no, 1 = yes).
Mental health scales used had been previously validated or used in Sierra Leone, Liberia, or other conflict-affected countries in sub-Saharan Africa [24–29]. All measures new to this setting were reviewed by local collaborators for face validity, examined item by item for local comprehension, and forward- and backward-translated following a standard protocol [30–32].
All continuous scales are averaged; thus, they are interpreted in terms of their original response scale.
Local research assistants (eight women and seven men) were trained to administer the survey using Android tablets running Kobo Toolbox digital data collection software [33]. All research assistants received a 3-d training on survey administration and research ethics. Interviewers were assigned to same-sex participants. Interviews were conducted one-on-one in an outdoor setting (due to ongoing EVD outbreak) that provided privacy and confidentiality. Participants were given contact information to use if they had any questions and were offered a small gift of foodstuffs (value US$2) for agreeing to be interviewed. Risk of harm cases were referred to local mental health or social services for follow-up care (two cases were identified and referred).
All descriptive statistical analyses were conducted in STATA 13.0 SE [34]. Using HLM 7.0 [35], we estimated two-level hierarchical linear models (also called “multilevel” or “mixed” models) to accommodate the clustering of data into EAs and to prevent bias in estimation of standard errors. Because this was an equal allocation sampling design based on 16 households per EA, weights were applied based on the population of each EA to allow generalization to the Western Urban and Western Rural districts from which the EAs were sampled. Households in two EAs were undersampled because the EAs had been improperly identified (clerical error), so the remaining households were up-weighted so that those EAs would be properly represented in the sample. Furthermore, the households that had been misidentified were placed in their proper EAs, resulting in oversampling of those EAs; households in these EAs were, therefore, down-weighted to prevent overrepresentation in the weighted sample. Robust standard errors were used for all model interpretation.
In accordance with our conceptual model, multilevel models predicting EVD prevention behaviors and EVD risk behaviors were fit, containing the three mental health variables—anxiety, depression, and PTSD symptoms—as well as war exposures and EVD diagnosis of people in their family or community (details described above). To examine PTSD symptoms, anxiety, and depression as mediators of EVD risk and prevention behaviors, and to estimate the direct and indirect effects of war exposures, we fit multilevel models in which each of the outcomes (EVD prevention behaviors and EVD risk behaviors) was regressed on war exposures (with measures of personal exposure to EVD and anxiety as controls) and each of the mental health outcomes, and also models in which two of the mental health variables (PTSD symptoms and depression symptoms) were regressed on war exposures [36]. Because the temporal sequencing of EVD-related health behaviors and anxiety was less clear than that of the other two mental health variables, anxiety was not considered as a mediator. Mediation analyses assumed that the effect of the exposure and mediator on the outcome were unconfounded.
In order to make use of all available data (n = 1,008) and avoid possible bias associated with listwise missing value deletion, 20 multiply imputed datasets were created for the analyses, using all variables included in the analyses plus additional demographic variables (participant age, sex, marital status, education, and household wealth, based on household land and asset ownership) and participant report of daily hardships for the imputations. The number of missing values for a given variable can be determined by comparing the n for a given variable in Tables 1 and 2 (e.g., 979 for depression symptoms score) to 1,008, the total number of participants interviewed.
Population-weighted estimates for the adult household-based population living in the Western Urban and Western Rural districts of Sierra Leone are presented in Table 1. The sample comprised 505 women (50.8%) and 503 men (49.2%). The mean age of participants was 34.2 y (95% CI 33.2, 35.2 y) and median age was 30 y, ranging from 18 to 84 y. Over 42% (42.7%; 95% CI 38.8%, 46.7%) of the participants reported being in a marital relationship or a partnership; an equal proportion (43.2%; 95% CI 39.4%, 47.1%) reported never being married or living with a partner. About a fifth of the participants reported no formal education (18.5%; 95% CI 14.6%, 23.2%) or incomplete primary schooling (4.7%; 95% CI 3.5%, 6.3%). The main ethnic groups were Temne (40.1%; 95% CI 34.3%, 46.1%), Mende (15.6%; 95% CI 12.7%, 19.2%), and Limba (13.6%; 95% CI 10.6%, 17.2%), and a majority of the participants (88.4%; 95% CI 85.3%, 91.0%) spoke Krio at home. Less than a quarter of participant households (22.1%; 95% CI 18.2%, 26.7%) owned their land, and out of a list of nine common assets, a mean of 3.5 (95% CI, 3.2, 3.9) assets was owned.
Thirty percent of participants (29.7%; 95% CI 24.2%, 35.8%) were exposed to at least one war-related event (Table 2), and the average number of war exposures was 0.88 (95% CI 0.69, 1.08). Of note, 20.8% (95% CI 16.1%, 26.4%) of the sample reported being forced to flee during the war, and 14.7% (95% CI 11.5%, 18.7%) had seen their homes/property destroyed. Just over 2% (2.1%; 95% CI 1.1%, 4.0%) of participants reported being beaten during the war, and 7.3% (95% CI 5.0%, 10.6%) of the sample had been threatened with death. Twenty percent of the sample (20.3%; 95% CI 16.1%, 25.3%) had witnessed looting/destruction of their home or belongings, 12.9% (95% CI 9.7%, 16.9%) had witnessed beating or torture of others, and 8.5% (95% CI 6.2%, 11.6%) had witnessed the killing of a household member by an armed group.
Levels of mental health problems were noteworthy. On the HSCL-25, the mean score was 1.38 (95% CI 1.30, 1.46) for the depression section and 1.29 (95% CI 1.21, 1.37) for the anxiety section. On the PTSD Symptom Scale, among those individuals who endorsed a traumatic war-related event or who chose to discuss a troubling event without describing it (n = 563), an estimated weighted prevalence in this sample for meeting the criteria of likely PTSD was 11.3% (95% CI 7.7%, 16.3%).
Close to half the participants reported having EVD cases in their community (44.7%; 95% CI 35.6%, 54.3%), and one in five reported cases among neighbors (20.5%; 95% CI 14.3%, 28.4%). Fewer reported cases among friends (11.0%; 95% CI 6.8%, 17.3%), family members outside the household (11.1%; 95% CI 7.2%, 16.6%), and household members (6.9%; 95% CI 3.8%, 12.3%).
Initial analyses showed that when personal EVD exposures were considered together, only EVD diagnosis of a friend showed a significant relationship with individual EVD prevention or risk behaviors; the other EVD exposures were thus dropped from all models for parsimony.
Relationships between war exposures, personal EVD exposure, mental health, and EVD-related health behaviors are displayed in Table 3. EVD risk behaviors were positively associated with depression symptom severity (b = 0.05; 95% CI 0.00, 0.10; p = 0.037) and PTSD symptom severity (b = 0.10; 95% CI 0.03, 0.17; p = 0.008) and inversely associated with war exposures (b = −0.09; 95% CI −0.17, −0.02; p = 0.036) and having a friend diagnosed with EVD (b = −0.04; 95% CI −0.08, −0.00; p = 0.036). Both PTSD (b = 0.47; 95% CI 0.31, 0.63; p < 0.001) and depression (b = 0.51; 95% CI 0.26, 0.76; p < 0.003) were associated with war exposures. The unstandardized coefficients are in terms of their original units as displayed in Table 2; for example, having a friend who had EVD is associated with an overall decrease in the risk behavior score of 0.04, which is about 17.6% of a standard deviation of that score. A decrease in the average response to the severity of depression scale of roughly 0.55 (a little more than halfway between responses on the Likert scale)—which corresponds to one standard deviation—results in an approximate 0.03 change in the outcome, which amounts to a 11.5% of a standard deviation change. Considering all variables in the model predicting EVD risk behaviors displayed in Table 3 and the additional steps to assess mediation, war exposures have an indirect positive effect on EVD risk behaviors through PTSD symptoms and depression; however, accounting for this effect, there is a negative direct effect of war exposures (i.e., tending to reduce EVD risk behaviors). Because the direct and indirect effects are in opposite directions, they act against each other and thus the total effects are closer to zero and not significant. Table 4 displays the direct, indirect, and total effects for depression and PTSD as mediators of war exposure’s effect on EVD risk and prevention behaviors.
EVD prevention behaviors (Table 3) such as frequent hand washing and avoiding mass gatherings were associated with higher levels of anxiety (b = 0.23; 95% CI 0.06, 0.40; p = 0.008). Depression was not significantly associated with EVD prevention behaviors. Having a friend diagnosed with EVD (b = 0.15; 95% CI 0.04, 0.27; p = 0.011) and higher levels of war exposure (b = 0.45; 95% CI 0.16, 0.74; p = 0.003) were associated with greater EVD prevention behaviors, while PTSD symptoms were associated with fewer EVD prevention behaviors (b = −0.24; 95% CI −0.43, −0.06; p = 0.009). As above, the coefficients are in the original units of the scale. For example, having a friend diagnosed with EVD is related to an increase in the level of preventative behaviors of 0.15, which is about 24.6% of a standard deviation. A change in the response for anxiety symptoms, about one standard deviation, was related to an increase of about 0.13 on the preventative behavior scale, which is about 21.3% of a standard deviation. Similar to the effect seen above, war exposures had negative indirect effects on EVD prevention behaviors, but in the presence of mental health mediators, there was a positive direct effect of war exposures on preventative behaviors (Table 4).
Consistent with our conceptual model, higher scores on measures of PTSD symptoms and depression were associated with higher EVD risk behaviors, and symptoms of PTSD were associated with lower levels of EVD prevention behaviors. In other words, chronic mental health difficulties associated with war trauma were associated with higher levels of risk behaviors in the context of the EVD epidemic. Furthermore, we showed that both PTSD and depression were in part associated with exposure to traumatic events during Sierra Leone’s civil war.
For both EVD prevention and EVD risk behaviors, the behaviors of participants were associated with EVD diagnosis of a friend, more so than with diagnosis of EVD among neighbors or other community members, suggesting that intimate knowledge of affected patients may be more important than geographic proximity in shaping one’s personal response to the epidemic.
After accounting for the association between war exposures and adverse mental health consequences, higher levels of exposure to war-related events were associated with EVD prevention behaviors; such a relationship might be indicative of individuals with higher levels of war exposure being persons with stronger survival skills or who have become more risk adverse.
Our results are consistent with relationships established between depression and PTSD in research on HIV risk behaviors [15]. As has been found in HIV research, in the presence of common mental health problems such as depression, good judgment was possibly distorted, leading to our finding that higher levels of depression symptoms were associated with higher levels of EVD risk behaviors. Traumatic stress reactions may also play a similar role whereby, in the presence of PTSD symptoms, the ability of the individual to attend to EVD prevention activities was less. However, higher levels of anxiety symptoms were associated with more EVD prevention behaviors such as hand washing and seeking out EVD information via healthcare workers, suggesting that anxious persons may maintain a higher level of vigilance or concern about the EVD epidemic or that vigilance may have raised anxiety in some individuals who were more actively taking precautions. It is also difficult to disentangle the relationship of anxiety to the EVD epidemic. It is possible that an unmeasured trait, such as vigilance, underlies both anxiety and preventive behaviors.
Although this is, to our knowledge, the first study of its kind to examine the relationship between both past war exposure and mental health and present-day EVD-related health behaviors, limitations must be noted. Although we used randomly selected census EAs to derive a representative sample of households in the Western Area at the height of the epidemic, this sample is specific to Freetown and the Western Area and not representative of Sierra Leone’s other 13 districts. However, this sample is of relevance for understanding the dynamics related to EVD and mental health in other major capital cities in the EVD-affected countries of West Africa. The analyses may also be subject to some degree of confounding. In addition, the measures of our main outcomes as well as personal EVD exposure, war exposures, and mental health predictors rely on self-report, and therefore raise the possibility of common methods bias.
In post-conflict settings, high levels of unaddressed mental health problems are common and are associated with war exposure. Such mental health difficulties can in turn shape the uptake of sensitization campaigns and public health messages aimed at reducing EVD risk. The findings in this study indicate the need for greater attention to the role of PTSD, depression, and other common mental health problems in counteracting risks for EVD in post-conflict Sierra Leone and may have implications for other war-affected regions. More generally, the findings suggest that successful uptake of EVD prevention messages among individuals with poor mental health in post-conflict low- and middle-income countries requires specific and targeted approaches that take into account the nature of war traumas, the resulting behavioral implications, and the potentially re-traumatizing effect of communication around fear and death, which are exacerbated in the context of such crises but remain underaddressed. The EVD epidemic laid bare the weaknesses in Sierra Leone’s health services. The country’s highly underdeveloped mental health and social services system must be strengthened to respond to the reality of compounded adversity due to both war and the recent epidemic and to prevent future outbreaks [37].
|
10.1371/journal.ppat.1007293 | Potent neutralizing antibodies in humans infected with zoonotic simian foamy viruses target conserved epitopes located in the dimorphic domain of the surface envelope protein | Human diseases of zoonotic origin are a major public health problem. Simian foamy viruses (SFVs) are complex retroviruses which are currently spilling over to humans. Replication-competent SFVs persist over the lifetime of their human hosts, without spreading to secondary hosts, suggesting the presence of efficient immune control. Accordingly, we aimed to perform an in-depth characterization of neutralizing antibodies raised by humans infected with a zoonotic SFV. We quantified the neutralizing capacity of plasma samples from 58 SFV-infected hunters against primary zoonotic gorilla and chimpanzee SFV strains, and laboratory-adapted chimpanzee SFV. The genotype of the strain infecting each hunter was identified by direct sequencing of the env gene amplified from the buffy coat with genotype-specific primers. Foamy virus vector particles (FVV) enveloped by wild-type and chimeric gorilla SFV were used to map the envelope region targeted by antibodies. Here, we showed high titers of neutralizing antibodies in the plasma of most SFV-infected individuals. Neutralizing antibodies target the dimorphic portion of the envelope protein surface domain. Epitopes recognized by neutralizing antibodies have been conserved during the cospeciation of SFV with their nonhuman primate host. Greater neutralization breadth in plasma samples of SFV-infected humans was statistically associated with smaller SFV-related hematological changes. The neutralization patterns provide evidence for persistent expression of viral proteins and a high prevalence of coinfection. In conclusion, neutralizing antibodies raised against zoonotic SFV target immunodominant and conserved epitopes located in the receptor binding domain. These properties support their potential role in restricting the spread of SFV in the human population.
| Foamy viruses are the oldest known retroviruses and have been mostly described to be nonpathogenic in their natural animal hosts. Simian foamy viruses (SFVs) can be transmitted to humans, in whom they establish persistent infection, as have the simian lenti- and deltaviruses that led to the emergence of two major human pathogens, human immunodeficiency virus type 1 (HIV-1) and human T lymphotropic virus type 1 (HTLV-1). Such cross-species transmission of SFV is ongoing in many parts of the world where humans have contact with nonhuman primates. We present the first comprehensive study of neutralizing antibodies in SFV-infected humans. We showed high titers of neutralizing antibodies in the plasma of most SFV-infected individuals. Neutralizing antibodies target the dimorphic portion of the envelope protein surface domain that overlap with the receptor binding domain. SFV-specific antibodies target epitopes conserved over 8 million years of co-speciation with their nonhuman primate host. Greater neutralization potency in infected individuals was statistically associated with smaller SFV-related hematological changes. In conclusion, our results suggest the protective action of neutralizing antibodies against SFV infection and spread in the human population.
| Simian foamy viruses (SFVs) are complex retroviruses that are widely prevalent in nonhuman primates (NHPs) [1]. In animals, SFV replicate in the superficial cell layers of the buccal cavity [2] and are mostly transmitted through bites and licking [3]. Humans are not natural hosts of SFV, but can be persistently infected over several decades after a cross-species transmission event [4–7]. Most SFV-infected people were bitten by a NHP and are thus the first hosts of a zoonotic virus [6, 8]. Human infection with zoonotic SFV is thus a natural model to study the key steps of the emergence of retroviruses. Several NHP species live in the tropical forests of Central Africa, and people from rural areas are frequently exposed to their body fluids through hunting, butchering, and meat consumption. Several new zoonotic agents have emerged from simian reservoirs populating this region, including human immunodeficiency virus-1 (HIV-1), human T-cell leukemia virus-1 (HTLV-1), and Ebola and Monkeypox viruses [9]. We have established that the prevalence rates of SFV in South Cameroon are approximately 0.3% of the general population and greater than 20% for people who have been bitten by a NHP [8, 10, 11]. Our work and other studies on people infected with SFV from African NHP species have consistently reported the persistence of replication-competent virus and the presence of SFV DNA in blood cells [12–18]. Blood gorilla SFV DNA loads vary between 1 and 1000 copies/105 cells [8, 18]. This is the range observed for blood HIV-1 DNA levels in HIV-1 infected humans [19].
From the immunological point of view, human SFV infection corresponds to the efficient control of a zoonotic retrovirus: SFVs persist throughout the lifetime of the host, but no major clinical impact for the infected host nor transmission to other human hosts has yet been described [4–6, 8]. SFVs show broad organ tropism in NHPs [2, 20]. In vitro, they are highly cytopathic for most cell lines [21]. In macaques, neutralizing antibodies prevent SFV transmission through blood transfusion [22]. SFV induces type I interferon (IFN) production in in vitro infected cells [23], and are susceptible to type I and type II IFNs and several restriction factors [7, 24–28]. Concerning immune responses, the only studies on human samples have shown SFV genome editing by apolipoprotein B mRNA-editing catalytic polypeptide (APOBEC) cytidine deaminases [29, 30] and the presence of neutralizing antibodies in one worker infected with a chlorocebus SFV [15].
SFV-specific antibodies induced in response to a zoonotic infection probably contribute to the restriction of viral replication in the infected hosts and the inhibition of human-to-human SFV transmission, two key steps in pathogen emergence. The antiviral function and properties of SFV-specific antibodies have yet to be described in humans. Accordingly, we aimed to perform an in-depth characterization of neutralizing antibodies raised by humans infected with a zoonotic SFV and their relationship with viral genotypes. We focused on gorilla SFV because these strains are found in approximately 70% of infected individuals in Cameroon and Gabon [8, 10, 31]. Furthermore, gorillas are phylogenetically close to humans. We quantified the neutralizing capacity of plasma samples from 58 SFV-infected hunters against primary zoonotic SFV strains [12]. We then studied the relationship between the neutralization specificity of plasma samples and the genotype of the strain infecting each hunter, defined the region of the SFV envelope protein (Env) targeted by the neutralizing antibodies, characterized the cross-recognition of various SFV genera, and investigated whether neutralization was associated with the characteristics of the SFV infection [32].
The primary strains isolated from our study population, representative of the two genotypes, were the tool of choice for the initial evaluation of neutralizing antibodies in the absence of data on the neutralization of gorilla SFV [12, 33]. The primary zoonotic GI-D468 and GII-K74 strains were neutralized by autologous plasma collected at two time points. Neutralization titers were very high (> 1:2,000) for plasma samples of BAD468 and moderate (< 1:200) for those of individual BAK74 (Fig 1A and 1B).
We quantified the neutralizing activity of plasma samples from 44 gorilla SFV-infected individuals (S1 Table). Thirty-four (77%) neutralized the GI-D468 strain, with a median titer of 1:496 (range 1:10–1:14,724; Fig 1C). Twenty-two (50%) neutralized the GII-K74 strain, with a median titer of 1:20 (range 1:10–1:2,279, Fig 1C). Plasma samples from eight uninfected hunters living in the same villages as infected individuals (S1 Table) had titers < 1:20 against both strains (Fig 1C). The frequency of samples with neutralizing activity were significantly different between infected and uninfected hunters (Fisher’s exact test P values were < 0.0001 and 0.01 for the GI-D468 and GII-K74 strains, respectively).
The 44 individuals infected with gorilla SFV can be defined as single, dual, or non-neutralizers: 24 plasma samples neutralized a single strain, 16 neutralized both strains, and four had no neutralizing activity. The lowest dilution tested for all plasma samples was 1:20, due to limited sample volume for some participants. Low neutralization activity observed at a 1:20 dilution only was confirmed using lower plasma dilutions (1:10 and 1:5). Among the 40 reactive plasma samples, five had low to moderate titers (< 1:200), 24 had high titers (≥ 1:200 and < 1:2,000), and 11 had very high titers (≥ 1:2,000). Thus, most humans infected with a gorilla SFV produced high titers of antibodies that neutralized primary zoonotic strains circulating in the same geographical region.
We hypothesized that the strain infecting each individual was the major determinant of genotype-specific neutralization. Among the 44 gorilla SFV-infected people, for whom neutralization assays were performed, 34 had had their env gene PCR-amplified and sequenced in a previous study [33]. Twenty-six SFV strains belonged to the GI genotype and eight to the GII genotype. Fig 1D and S2 Table show the neutralization titers against both strains and the genotype of the SFV strain for each participant: the infecting and neutralized strains were of the same genotype (15 GI, 5 GII) for all single neutralizers. Dual neutralizers were infected with strains from GI (n = 10), GII (n = 2), or undetermined (n = 4) genotype. Their neutralization titers against the GI-D468 and GII-K74 strains were not related: for example, plasma from individual BAK74 neutralized the GI-D468 strain more efficiently than its autologous strain (neutralization titers: 1:340 vs 1:62). The results at this point support a match between the specificity of neutralization and viral genotype for single, but not dual neutralizers.
We hypothesized that dual neutralizers were coinfected with strains from the two genotypes. We thus established two genotype-specific PCR assays to demonstrate coinfection. Eight people had blood cell DNA samples positive by both GI and GII-specific PCR, showing coinfection by strains of the two genotypes. Twenty-one people were infected with a GI strain only (19 previously classified as GI and two undetermined, S2 Table) and 10 with a GII strain only (seven previously classified as GII and three undetermined, S2 Table). The coinfection status was unknown for five individuals because we ran out DNA or both genotype-specific PCRs gave negative results. Re-examination of the neutralization pattern and results from genotype-specific PCR showed that one of the eight coinfected individuals neutralized a single strain, whereas seven neutralized both strains (Fig 1E, S2 Table). The nine other dual neutralizers tested positive by either GI- or GII-specific PCR only or were undetermined. In conclusion, coinfection by SFV strains of the two genotypes occurred in eight of 39 (20%) infected individuals, and most coinfected individuals raised antibodies that neutralized strains of both genotypes.
We next investigated the localization of the epitopes targeted by the antibodies. The match between genotype(s) from neutralized and infecting strains supports the direct interaction of genotype-specific amino acids (aa) with the neutralizing antibodies or their indirect involvement with the proper conformation of the epitopes. Alignment of Env protein sequences from the GI-D468 and GII-K74 strains showed that the central region of gp80SU harbors a variable region (SUvar) which differs significantly in aa composition, with 42% (101 of 243) nonidentical aa (Fig 2A). In contrast, the surrounding bipartite conserved portion of the SU (SUcon) contains only 2.5% nonidentical aa: two N-terminal and a cluster of three C-terminal aa of the SUvar region. The gp18LP and gp48TM subunits vary by one and three aa, respectively, corresponding to less than 1% nonidentical aa.
We generated foamy virus vector particles (FVV) enveloped by wild-type GI Env (EnvGI) and chimeric GI/II Env with exchange of SU (EnvGI-SUGII), SUvar (EnvGI-SUvarGII), or SUcon (EnvGI-SUconGII) sequences using chimeric Env packaging constructs (Fig 2). Neutralization assays were performed with the FVV, using the same cell line (GFAB) and same moi (100 IU/well) as for the replicating viruses. Susceptibility to neutralization by human plasma samples of FVV with EnvGI and EnvGI-SUGII Env was similar to the one of replicating GI-D468 and GII-K74 viruses with less than a two-fold difference in neutralization titers (S1 Fig).
We first tested samples from single neutralizers: plasma sample from individual BAD463, which neutralized the GI-D468 virus only, efficiently neutralized the vectors carrying EnvGI and chimeric EnvGI-SUconGII, but not chimeric EnvGI-SUGII or EnvGI-SUvarGII (Fig 3A). Plasma BAD551, which neutralized only the GII-K74 virus, gave the opposite pattern of neutralization (Fig 3B). Thus, neutralization epitopes recognized by these two plasma samples are encoded by the SUvar region. These results were confirmed, with 14 plasma samples neutralizing the GI-D468 virus only and six neutralizing the GII-K74 only: when diluted at 1:80, the plasma samples reduced the relative infectivity of vectors expressing the corresponding SUvar region by five to 100-fold (Fig 3C).
We next tested plasma from individuals coinfected with strains of two genotypes against the four vectors. Neutralization titers depended on the expression of the SUvar region. For example, plasma from individual BAK177 had a high neutralization titer against the GI-D468 virus (1:1,025) and a low titer against the GII-K74 virus (1:36). Its neutralization titers against vectors carrying EnvGI and EnvGI-SUconGII were high (1:790 and 1:655) and those against EnvGI-SUGII and EnvGI-SUvarGII vectors were low (1:66 and 1:89) (Fig 3D). Other plasma samples showed similar neutralization patterns (Fig 3E and S3 Table). In conclusion, the SUvar domain carries the neutralizing epitopes recognized by antibodies from single neutralizers and coinfected individuals.
The plasma of nine individuals neutralized both viral strains, whereas they were only positive in a single genotype-specific PCR. There are two possible explanations for this result: they were coinfected, but only a single strain was detectable by PCR, or their neutralizing antibodies target epitopes that are located in regions conserved between the two genotypes. For most, the titers were clearly different against the two replicative viruses (Fig 1E), arguing against the neutralization of conserved epitopes. Indeed, titers depended upon expression of the SUvar region by vector particles, as shown in Fig 3F and 3G and S3 Table. Thus, dual neutralizers were most likely coinfected by strains of two distinct genotypes.
Four of 18 individuals infected with a GII strain did not neutralize the GII-K74 strain (three were infected with a GII strain only and one was infected with GI and GII strains but neutralized the GI-D468 strain only). One individual had no neutralizing activity against either strain and was confirmed to have been infected on the basis of a Gag-specific WB and an LTR-specific PCR only, whereas the Pol, Env, and Env genotype-specific PCR assays (this paper) were negative [8, 33]. Of note, all 31 individuals infected with a GI strain were able to neutralize the GI-D468 strain. In conclusion, neutralizing antibodies were detected in the plasma of 100% GI SFV-infected individuals and 78% of those infected by GII SFV.
We therefore investigated whether the higher proportion of non-neutralizers among GII-infected individuals was related to differences in Env sequence variability across strains. Env sequences derived from genotype-specific PCRs amplicons were aligned for the 31 GI and 18 GII strains infecting our study population and the viral strains used in the neutralization assays, with sequences ordered as a function of their neutralization titers (S2 Fig). Sequence variability was similar for the GI and GII strains: 16% and 18% of the positions had at least one aa change among the analyzed GI and GII sequences, respectively. Analysis of 12 full-length GI and seven full-length GII Env sequences showed similar variability [33]. Thus, there was no clear relationship between Env protein diversity within each genotype and the detection of neutralizing antibodies.
Among the four GII-infected non-neutralizers, individual CH101 had full length env gene sequence data available. There was a single nonconservative F420Y change in the CH101 SUvar sequence, using the GII-K74 sequence as a reference. In an ELISA assay, plasma from CH101 did not react with peptides corresponding to GII-K74 and CH101 amino acids 411 to 430, whereas plasma from BAK74 recognized both peptides. Thus, the antigenic difference between the infecting and neutralized strains does not explain the lack of neutralization, at least for individual CH101. We were unable to obtain a second sample from all non-neutralizers to confirm the absence of plasma neutralization and amplify the env gene.
The formerly used classification of viral strains was based on their susceptibility to neutralization (serotyping) and showed segregation of SFV according to the host species. Among SFV infecting African NHPs, those from chlorocebus, papio and chimpanzee belong to four distinct serotypes, with chimpanzee SFV strains divided into two [21]. Gorilla SFV had not been previously typed. We therefore examined the susceptibility of SFV strains to neutralization according to their origin. We started by testing two laboratory-adapted chimpanzee SFV strains with defined serotypes: PFV, belonging to the CI genotype and serotype 6, and SFV7, belonging to the CII genotype and serotype 7 [34]. Plasma samples from eight individuals infected with a chimpanzee SFV were tested and neutralized CI-PFV (n = 3), CII-SFV7 (n = 6), GI-D468 (n = 6) or GII-K74 (n = 4) (Table 1). The plasma from seven individuals neutralized a single chimpanzee SFV and one was a dual neutralizer.
We further characterized the cross-neutralization of gorilla and chimpanzee SFV by testing the plasma of the 44 gorilla SFV-infected individuals: 17 neutralized CI-PFV only, eight CII-SFV7 only, 10 both strains, and nine none of them (S2 Table). In contrast, plasma from cercopithecus SFV-infected individuals had no or low neutralization titers against gorilla and chimpanzee SFVs (Fig 4A). Thus, gorilla and chimpanzee SFV share at least some neutralizing epitopes, but are not inhibited by plasma from hunters infected with monkey SFV.
Combining the data from gorilla and chimpanzee SFV-infected individuals (n = 52 samples), neutralization of homologous strains (i.e. phylogenetically related SUvar domains) were mostly concordant. Neutralization of GI-D468 and CI-PFV were concordant for 42 samples (12 negative and 30 positive) and discordant for 10 samples, which neutralized only GI-D468 (Fig 4B). In reactive plasma samples, titers against GI-D468 and CI-PFV were correlated (Spearman’s ρ = 0.696, P < 0.0001). The second genotype showed a similar profile (Fig 4C): neutralization of GII-K74 and CII-SFV7 were concordant for 48 samples (24 negative and 24 positive), and discordant for four, which neutralized either GII-K74 (n = 2) or CII-SFV7 (n = 2) viruses. In reactive plasma samples, titers against GII-K74 and CII-SFV7 were correlated (ρ = 0.823, P < 0.0001). Neutralization titers against viruses with unrelated SUvar domains were not correlated (P = 0.32 for GI-CII and P = 0.99 for GII-CI, Fig 4D and 4E). These data confirm that the SUvar domain is a major determinant of recognition by neutralizing antibodies raised against both gorilla and chimpanzee SFVs.
We compared the neutralization of the laboratory-adapted CII-SFV7 strain to that of zoonotic primary chimpanzee SFVs that we isolated from Cameroonian hunters [12]. Both CII-D327 and CII-AG15 belonged to the CII genotype and neutralization titers against both viruses were strongly correlated (Spearman’s rho = 0.932, P = 0.0002). CII-D327 was neutralized by plasma from 23 of 52 individuals infected with either gorilla or chimpanzee SFV. Neutralization of the CII-D327 and both CII-SFV7 and GII-K74 strains was mostly concordant, and in reactive plasma samples, the titers were strongly correlated (Fig 4F and 4G). Conversely, neutralization of CII-D327 was unrelated to that of CI-PFV and GI-D468 (Fisher’s exact test P = 0.26 and P = 0.33, respectively). In conclusion, laboratory-adapted and primary SFV strains show similar susceptibility to neutralization. Nevertheless, the occurrence of discordant responses against strains belonging to the same genotype support some inter-individual variability in the interactions between neutralizing antibodies and susceptibility sites on SFV Env.
We investigated whether neutralizing antibody levels were related to the parameters of SFV infection. The neutralizing activities of plasmas was tested against four SFV strains, corresponding to the two genotypes from each of the two host-specific SFV clades (GI-D468, GII-BAK74, CI-PFV and CII-SFV7). We considered two quantitative measures of neutralization: magnitude, defined as neutralization titers against the species- and genotype-matched strains and breadth, defined as the number of neutralized strains. Neither magnitude nor breadth of neutralizing antibodies were associated with age, duration of infection, or blood SFV DNA levels (Fig 5). Neutralization titers varied over a wide range, leading us to consider a composite measure to better quantify the neutralization capacity of plasma samples against the panel tested. A score was assigned for each plasma sample and each viral strain, based on neutralization titers. Neutralization titers < 20, ranging from 1:20 to 1:200, 1:200–1:2000, and > 1:2000 correspond to 0, 1, 2, and 3 points, respectively [35]. The sum of the points for the four tested strains defined the neutralization score of a plasma sample. The neutralization score was not associated with age, duration of infection, or blood SFV DNA levels (Fig 5C, 5F and 5I). The duration of infection was different for the two ethnic groups, Bantus and Pygmies (medians: 13 and 21 yrs for Bantus and Pygmies, respectively, Mann-Whitney test P = 0.02), and SFV DNA levels tended to differ (22 and 32 copies/105 cells, P = 0.09). Bantus and Pygmies had similar neutralization magnitude (Medians: 1:395 vs. 1:965, P = 0.11) and breadth (Medians: 2 vs; 2, P = 0.68). In Pygmies, higher breadth was associated with higher SFV DNA levels (Fig 5J), whereas the SFV DNA loads were very homogeneous and not related to the neutralization breadth in Bantus (Fig 5J). Neutralization titers against the four viruses were stable for samples collected from seven individuals at two time points, one to seven years apart (Table 2).
Finally, we searched for associations between neutralization levels and hematological parameters, of which the levels were significantly different between SFV-infected individuals and matched uninfected controls [32]. Hunters infected with SFV had lower hemoglobin levels than uninfected hunters. Here, among 22 SFV-infected individuals, those with a higher neutralization breadth or score had higher hemoglobin levels, as well as a higher hematocrit and erythrocyte levels (Fig 6A–6I). Higher neutralization magnitude also tended to be associated with higher hemoglobin levels and hematocrit. Only 33% of individuals that neutralized three or four strains had mild or moderate anemia (hemoglobin level < 13 g/dl), whereas as the frequency was 69% for those who neutralized two strains or less. In addition, SFV-infected individuals had higher urea levels than uninfected controls, which were lower for those with a higher neutralization breadth or score (Fig 6K and 6L). We also repeated the analyses with a higher neutralization cut-off (1:80) and observed similar associations or trends for associations. Analyses stratified by ethnic group showed similar associations in both groups, except for a strong correlation between neutralization breadth and protein levels in Bantus only (Spearman’s rank test rho = 0.874, P = 0.01). In conclusion, individuals who produced neutralization antibodies with a larger breadth displayed higher hemoglobin and lower urea levels than those with a narrower neutralization range.
We report the presence of high titers of neutralizing antibodies in the plasma of most SFV-infected individuals. Ape SFV species comprise two genotypes that cocirculate in humans and NHPs from Central-Africa [33]. We detected predominantly genotype-specific neutralization. We constructed vectors with chimeric Env, based on naturally occurring sequence variations, and used them to map dominant viral susceptibility sites to the dimorphic portion of the surface domain that overlaps the receptor-binding domain. The frequent cross-neutralization of gorilla and chimpanzee SFVs supports the recognition of highly conserved epitopes. Our description of the neutralization patterns of SFV-infected humans provides two key pieces of information concerning their virological status: gorilla SFVs appear to continuously or sporadically express their viral proteins and the coinfection rate is over 30%. Several hematological markers differ between SFV-infected individuals and uninfected controls [32]. We found that a larger neutralization breadth is statistically associated with smaller hematological changes, supporting the beneficial impact of the humoral response on the clinical outcome of infected hosts. SFV has a high capacity to cross the host-species barrier and persistently infect humans. Overall, our data demonstrate potent neutralization, targeting mostly conserved and immunodominant epitopes located in the receptor binding domain, produced by approximatively 90% of gorilla SFV-infected humans. These properties may have helped to block the spread of SFV in the human population.
We show here that most of the 52 individuals infected with gorilla or chimpanzee SFV produced neutralizing antibodies. Titers varied by over 1,000-fold between participants and were high or very high in 75%. A key element of the cross-species transmission of viruses is their replication in the new host. A side-by-side comparison of SFV viral load, replication sites, and immune response in human and animal samples is impossible to perform for apes living in the wild. However, neutralization titers (< 1:20 to 1:15,000) in human plasma were similar to those reported in chimpanzee and gorilla samples (< 1:32 to 1:2,500) tested against chimpanzee SFV [36]. SFV-specific IgG was found to be slightly lower in occupationally infected humans than in captive chimpanzee samples using a different assay of SFV-specific antibodies (semi-quantitative WB assay) [37]. A humoral response of comparable magnitude in humans and apes reflects the lack of an intrinsic restriction of SFV replication in vivo in humans, in accordance with the lack of in vivo virus adaptation after a zoonotic infection [12].
The humoral response in gorilla SFV-infected people is consistent with persistent exposure to viral antigens during chronic viral infection, as neutralization titers were (1) high in most individuals; (2) of similar magnitude among people who have been chronically infected for 1 to > 40 years; and (3) stable over a one to seven year-long longitudinal follow-up. Some Brazilian workers with serological evidence of infection by New World Monkey (NWM) SFV were shown to lose their seroreactivity in western-blot assays [38]. Blood SFV DNA levels for NWM strains are usually undetectable in humans, in contrast to NHP [38–40]. Overall, both the frequency of SFV DNA detection in human blood samples and antibody levels support higher expression for SFV that originate from the closest human relatives [8, 38, 39, 41]. In conclusion, high neutralization titers in humans infected with ape SFV strongly suggest that these zoonotic viruses express their proteins in their new human hosts.
We describe the paradoxical situation of high neutralization titers, an indicator of viral protein expression, and undetectable viral RNA in blood and saliva of the same study population [14]. Blood antibodies reflect viral protein expression in the organism from months to years, whereas qRT-PCR assays are limited to the tissue analyzed (blood or buccal samples), at a given time point. Thus, the lack of SFV RNA detection in blood or saliva does not prove the absence of viral replication in the human hosts. Furthermore, circulating lymphocytes—the major blood cell type carrying SFV DNA [13]–are usually in a resting state not permissive to SFV replication, whereas proliferating lymphocytes are mostly located in lymphoid organs and tissues. In vitro, SFV latency is characterized by persistent expression of the Bet protein only, in the absence of Env expression [42–44]. Thus, the presence of high levels of Env-specific antibodies provide indirect evidence against in vivo latency of gorilla SFV in humans.
We recently observed that several hematological parameters differ between SFV-infected hunters and age-matched uninfected hunters living in the same area [32]. The hematological alterations consisted of reduced hemoglobin and red-cell levels and elevated protein, urea, creatinine, creatine phosphokinase, and lactate dehydrogenase levels. Individuals with a large neutralization breadth had smaller hematological alterations than those with a narrow neutralization breadth. Neutralization breadth depends upon the targeted epitopes, diversification of viral sequences in the infected host, efficient cross-talk between B and T lymphocytes, and the maturation and selective expansion of high affinity antibodies. If the neutralization breadth of SFV-specific antibodies reflects their in vivo efficacy, as shown for other viral infections [45], our results may reflect the protective action of neutralizing antibodies against hematological changes observed in SFV-infected individuals.
Approximately 10% of gorilla SFV-infected individuals had detectable neutralizing antibodies. We were unable to obtain a second blood sample from these individuals, precluding confirmation of the diagnosis and the neutralization pattern. For one non-neutralizer, CH101, the sequence variation between the autologous virus and the strain used in the neutralization assay is unlikely to explain the negative results. Another non-neutralizer, BAK235, was included in our recent clinical study [32]. He was not immunosuppressed, and his hematological parameters were close to the median value of the group (Fig 6). All participants were apparently healthy at the time of sampling and all had negative HIV serological assays. Thus, undetectable neutralization was probably not a consequence of poor health. The lack of neutralizing antibodies may be related to low viral load, as suggested by undetectable SFV DNA in the qPCR assay for individual BAK235 [8], and failure to amplify the env gene for individuals BAK235, CH86, and H10GAB79. We indeed observed a positive correlation between antibody titers and blood SFV DNA levels for Pygmies. Overall, undetectable plasma neutralization in some participants may be explained by low viral replication.
Here, we showed that neutralizing epitopes targeted by human plasma are mostly located in the SU domain of the Env protein and demonstrated the relation between the genetic and antigenic properties of gorilla SFV strains. For chimpanzee SFV, the initial description of two serotypes [46] is consistent with the latter demonstration of infection of these NHP species by these two genotypes [33]. Here, we directly demonstrated that chimpanzee SFV strains belonging to serotypes 6 and 7 [36, 46] were neutralized by the plasma of individuals infected with CI and CII genotypes, respectively. Feline foamy virus (FFV) genotypes were shown to determine the type-specificity of neutralizing antibodies [47] and the fragment of the genome comprising the env and bel1 genes to carry the determinants for FFV genotype-specific neutralization [48, 49]. The two genotypes of FFV are present in domestic and wild felids of various geographic regions [50], and several Asian and African NHP species are infected by SFV strains segregating into two genotypes [33]. Overall, targeting of the bimorphic region located in the Env protein surface domain is a cardinal feature of antibodies that neutralize thus far studied foamy viruses.
Most plasma samples neutralized both gorilla and chimpanzee SFV, indicating high conservation of some of the sites targeted by the neutralizing antibodies. The chimpanzee/gorilla split occurred eight million years ago [51]. The phylogeny of the conserved and variant portions of the env gene differ: for the central SU region, SFVs form two clades, corresponding to the two genotypes, and within each clade, isolates are distributed according to their host [33]. Consequently, aa identity of the SUvar region is ≈ 70% for GI-CI and GII-CII pairs of viral strains, and only ≈58% for GI-GII and CI-CII pairs. In the conserved backbone of Env, aa identity is ≈ 78% for GI-CI and GII-CII pairs and above 95% for GI-GII and CI-CII pairs. These data are fully concordant with the cross-neutralization of gorilla and chimpanzee SFVs for strains from the same genotype group. However, some plasma samples only neutralized strains from a single host species (19% and 8% between GI-CI and GII-CII strains, respectively). Such a narrow pattern may reflect inter-individual variation in Env sequences and/or antibody binding sites. Minor antigenic differences between serologically related strains have indeed been detected for FFV and chimpanzee SFV [34, 52]. The genetic stability of foamy viruses in each host and among hosts of the same species limit their potential to escape neutralizing antibodies. In humans, the genetic stability of SFV is high (98.6 homology in the Pol protein in blood samples taken four years apart [15]), and no adaptation to the human host has been described, even decades after infection [12, 53]. Viruses from human-NHP transmission pairs had almost identical sequences [15, 54]. Overall, neutralization epitopes are expected to be conserved because of the previously reported genetic stability of SFV and we further demonstrate their conservation during the coevolution of SFV with their NHP hosts.
Physical blockade of the interaction between viruses and their cellular receptor is an important and common mechanism of virus neutralization [55]. The SFV receptor is unknown, but ubiquitous. Interference experiments showed that simian, feline, bovine, and equine FV share the same receptor [56]. The particle-associated envelope glycoprotein of FV is composed of trimers arranged in interlocked hexagonal assemblies, with the SU bearing the receptor binding site located at the top of the spikes [57]. Here, we demonstrated that immunodominant neutralizing epitopes are predominantly located in the dimorphic central part of the SU that overlaps the receptor binding domain [58]. Polymorphic aa located outside of the SUvar region are not involved in recognition by neutralizing antibodies, nor required for the proper conformation of Env, as chimeric Env built for this project mediated the entry of vector particles. The suppression of viral transmission by effective neutralizing antibodies may induce evolutionary pressure that selects antibody escape mutations but immune evasion may be limited if it results in variant viruses with impaired replicative capacity. The Env sequence from SFV is less variable than that of Gag [59]. This pattern is opposite to that observed for most other retroviruses and probably reflects strong structural constraints on the Env protein, which in turn may favor the efficient neutralization of SFVs.
One third of gorilla SFV-infected individuals were dual neutralizers and one fifth had evidence of coinfection by strains from the two genotypes. All but one coinfected individual were dual neutralizers. Discordance between serological and molecular assays for genotype-specific SFV detection may have several causes. Nucleic acid tests give negative results when the viral load is low or strains carry variant nucleotides in the primer sequences. The sensitivity of serological detection of an infection rarely reaches 100%, due to the lack of an antibody response in some individuals and/or minor antigenic differences between viruses. Here, we only detected coinfection by strains belonging to distinct genotypes, and true coinfection rates may be higher than those estimated by serological or molecular detection of viral genotypes.
The rate of coinfection in Cameroonian and Gabonese hunters is in the same range as that reported in Bangladeshi infected with macaque SFV [60]. Dual neutralization has been reported in sera drawn from chimpanzees housed in a primate research center, as well as bush chimpanzees [46, 61] and FFV-infected cats [47, 62]. In Bangladesh, the analysis of viral sequences from macaques revealed that at least one of four adult animals was infected with at least two strains [60]. Single genome amplification of wild chimpanzee samples has provided evidence of coinfection with genetically diverse viruses [63, 64]. In a wild community of chimpanzees in Ivory-Coast superinfection rates increased with age and were above 80% in adult animals [65]. Being injured while hunting apes in Africa is infrequent, well-recalled, and occurred more than once for only one individual of our study group. The high rate of coinfection in gorilla SFV-infected hunters is most probably the result of simultaneous transmission of several strains during a single exposure event. Indeed, we tested a series of ape samples and found that 26% of chimpanzees and 37% of gorillas from Cameroon and Gabon were coinfected by strains of both genotypes (S4 Table).
SFV coinfection is thus frequent in animals, and transmission of more than one virus strain to humans is frequent for macaque and gorilla SFV. Recombination is efficient in vitro [66] and has been detected in viruses infecting chimpanzees and macaques [60, 63]. Our results concerning coinfection are particularly relevant for the emergence of retroviruses in the human population, as simian immunodeficiency viruses have evolved through several recombination events in NHP hosts coinfected by two strains, progressively acquiring viral proteins able to antagonize host restriction factors, eventually allowing successful emergence of HIV-1 [9, 67, 68].
In conclusion, most SFV-infected humans produce neutralizing antibodies at levels similar to those detected in simian hosts. SFV-specific neutralizing antibodies display key features of protective responses: they target conserved epitopes and their breadth is associated with reduced biological changes in the infected hosts. The receptor-binding domain is the immunodominant region targeted by neutralizing antibodies. Whether subdominant cross-reactive antibodies recognizing the SUvar domain and/or antibodies targeting the conserved SU regions are present in plasma samples is still an open question. The overall robust SFV-specific neutralization calls for further studies dedicated to fine epitope mapping, Fc-mediated antiviral functions, and molecular characterization of human monoclonal antibodies [69–72].
Participants gave written informed consent. Ethics approval was obtained from the relevant national authorities in Cameroon (the Ministry of Health and the National Ethics Committee) and France (the Commission Nationale de l'Informatique et des Libertés (CNIL) and the Comité de Protection des Personnes Ile de France IV).
Field studies were performed on adult populations living in villages and settlements across rural areas of the rainforest in Cameroon and Gabon [8, 10, 31]. SFV infection was diagnosed by a clearly positive Gag doublet on Western blots, using sera from the participants, and the amplification of the integrase gene and/or LTR DNA fragments by PCR, using cellular DNA isolated from blood buffy-coats [8]. The origin of the SFV was identified by phylogenetic analysis of the sequence of the integrase gene, as described [8]. Plasma samples from 66 participants were used for this study, 53 Cameroonians and 13 Gabonese (S1 Table). Eight participants were not infected with SFV, eight were infected with a chimpanzee SFV, 44 with a gorilla SFV, and six with a cercopithecus SFV. All but four SFV-infected participants had been injured by a NHP. All participants were male, 37 were Bantus and, 29 were Pygmies. Their median (interquartile range, IQR) age was 53 (43–61) years. The duration of infection was estimated as the time elapsed between receiving the wound and the sampling and was 19 (11–28) years. Blood SFV DNA levels were determined by qPCR carried out on buffy-coats from 30 of the 44 gorilla SFV-infected participants, as described in a previous study [8]. Blood tests were carried out at the Centre Pasteur du Cameroun in Yaoundé for 22 individuals [32].
Gorilla FAB (GFAB) cells are baby hamster kidney (BHK)-21-derived clones containing the β-galactosidase gene under the control of the LTR from SFVggo-hu.BAK74 LTR [73]. They were cultured in DMEM-5% fetal bovine serum (FBS) supplemented with 300 μg/ml G418 (Sigma-Aldrich, Lyon, France). Human embryonic kidney 293T cells (Cat. N° 12022001, Sigma) were cultured in DMEM-10% FBS.
Virus stocks were produced by infection of subconfluent BHK-21 cells that were passaged twice a week. Cultures displaying an extensive cytopathic effect were lysed by three freeze-thaw cycles. Cell lysates were clarified, filtered (0.45 μm pore size), and stored as single-use aliquots. Titrations were performed as described [73]. Briefly, serially diluted viral solutions were incubated with indicator cell lines in 96-well plates. Cells were fixed after 72 h with 0.5% glutaraldehyde in phosphate-buffered saline (PBS) for 10 min at room temperature (RT). Cells were washed with PBS and incubated for 1 h at 37°C with an X-Gal staining solution [2 mM MgCl2, 10 mM potassium ferricyanide, 10 mM potassium ferrocyanide, and 0.5 mg/ml 5-bromo-4-chloro-3-indolyl-B-D-galactopyranoside in PBS]. An S6 Ultimate UV Image analyzer (CTL Europe, Bonn, Germany) was used to count X-Gal stained cells. One infectious unit (IU) was defined as a blue cell or syncytia.
Original and first passage cell lysates were used to produce the viral stocks from zoonotic primary strains isolated from infected hunters [12]. The genotypes were defined in [33] and are abbreviated as follows: GI and GII for gorilla SFV genotypes I and II, respectively; CI and CII for chimpanzee SFV genotypes I and II, respectively. SFVggo_huBAD468 (JQ867465) and SFVggo_huBAK74 (JQ867464) belonged to the GI and GII genotype, respectively; SFVptr_huAG15 (JQ867462) and SFVptr_huBAD327 (JQ867463) belonged to the CII genotype. Two laboratory-adapted chimpanzee SFV isolates, SFVpsc_huPFV (KX087159, [74]) and SFVpve_Pan2 (SFV-7, env gene KT211269) belonged to the CI and CII genotypes, respectively. Here, strain nomenclature [75] was replaced by short names summarizing the genotype and strain names. For example, in the short name “GI-D468”, GI stands for GI genotype and -D468 for SFVggo_huBAD468.
Human blood samples were collected into EDTA tubes and processed within 24 h of collection using standard techniques. Human plasma samples were stored at -80°C. Before use in neutralization assays, plasma samples were diluted 1 to 10 in DMEM + 1 mM MgCl2, heated 30 min at 56°C, to inactivate complement proteins, and frozen as single-use aliquots. Serial two-fold dilutions of plasma samples were incubated with SFV isolates for 2 h at 37°C before quantification of residual viral infectivity using the SFV microtitration assay in 96-well plates described above. The same multiplicity of infection (100 IU/well) was used for all viral strains. Assays were performed in triplicate. Neutralization titers were defined as the inverse of the plasma dilution required to reduce viral infectivity by half.
The lowest plasma dilution routinely tested was 1:20 to spare samples and avoid nonspecific reactivity. Titration curves were considered valid if there was a steady rise in the number of infectious units per well with increasing dilution of the sample and if a plateau was observed at the highest plasma dilutions. Neutralization titers were defined as the inverse of the dilution required to reduce viral infectivity by half. For their calculation, the number of infected cells/well were plotted (y) as a function of the dilution factor (x). The slope (a) and the intersection (b) with the y axis were calculated using the linear portion of the neutralization curve. Ymax was defined as the mean of infectious units at the plateau. The neutralization titer was calculated as [(Ymax/2)-b]/a. Calculation of the linear regression was possible only when infectivity was reduced by more than 50% for at least two dilutions, giving a quantification threshold of 1:40. For some samples, a reduction of infectivity ≥ 50% was observed at the 1:20 dilution only, which was confirmed at lower dilutions. Therefore, plasma samples exhibiting a reduction of infectivity ≥ 50% at the 1:20 dilution only were arbitrarily defined as having a neutralization titer of 1:20. Plasma samples exhibiting a reduction of infectivity < 50% at the 1:20 dilution were considered non-neutralizing. The value corresponding to half the detection threshold (1:10) was used when a quantitative expression of the results was necessary (on graphs and for statistical assays).
Study subjects were infected by SFV strains belonging to two genotypes (GI and GII), previously determined by direct sequencing of the env gene amplified from the buffy coat [33]. Here, env genotype-specific primers were used to detect coinfection by strains from the two genotypes. Env fragments were amplified from buffy coat genomic DNA using specific primers (S5 Table). Nested PCR was performed by mixing 500 ng genomic DNA in the enzyme buffer with the external primers (0.25 μM each), MgCl2 (3.5 mM), deoxynucleoside triphosphates (dNTPs, 200 μM each), and 0.5 μl HotStar-Taq polymerase (Qiagen) in a final volume of 50 μl. External PCR consisted of a 15-min-long denaturation step at 95°C, followed by 40 amplification cycles (45 s at 95°C, 45 s at 50°C, and 1 min per kb at 72°C) and a 7-min-long extension step at 72°C. The product (5 μl) was then used as the template for a second internal PCR under the same conditions but using the internal primers.
The expression-optimized four-component PFV vector system comprised constructs expressing PFV Gag (pcoPG4), Env (pcoPE), Pol (pcoPP), and the EGFP-expressing transfer vector puc2MD9 [76, 77]. Expression-optimized gorilla SFV envelope (env) genes were synthesized by Genscript (Piscataway, NJ, USA) using sequences from the GI-D468 env and GII-K74 gp80SU genes. Restriction sites were inserted in both sequences, allowing the generation of chimeric genes through cloning (sequences are available upon request). Gorilla SFV env open reading frames (ORFs) were cloned into the pcoPE plasmid using the HindIII-ApaI restriction sites, replacing the PFV env gene. Foamy vector particles were produced by co-transfection of four plasmids. Polyethyleneimine (45 μl JetPEI, Polyplus # 101-10N, Ozyme, Montigny-le-Bretonneux, France) and 15 μg total DNA (gag:env:pol:transgene ratio of 8:2:3:32) were added to a 10-cm2 culture dish seeded with 4 x 106 HEK 293T cells. Supernatant was collected 48 h later, clarified by centrifugation at 1,500 x g for 10 min, and stored at -80°C as single-use aliquots. Titration was performed on GFAB cells and an S6 Ultimate UV Image analyzer (CTL Europe, Bonn, Germany) was used to count fluorescently stained cells. One infectious unit was defined as a fluorescent cell, resulting from EGFP expression, for which the gene is encoded by the puc2MD9 plasmid packaged into vector particles.
We studied associations between categorical and continuous variables with Fisher’s exact test and Spearman’s rank test, respectively. Analyses were conducted using GraphPad Prism v5.0 (GraphPad Software, San Diego California USA). P values < 0.05 defined statistical significance.
|
10.1371/journal.pbio.1002375 | PI(4)P Promotes Phosphorylation and Conformational Change of Smoothened through Interaction with Its C-terminal Tail | In Hedgehog (Hh) signaling, binding of Hh to the Patched-Interference Hh (Ptc-Ihog) receptor complex relieves Ptc inhibition on Smoothened (Smo). A longstanding question is how Ptc inhibits Smo and how such inhibition is relieved by Hh stimulation. In this study, we found that Hh elevates production of phosphatidylinositol 4-phosphate (PI(4)P). Increased levels of PI(4)P promote, whereas decreased levels of PI(4)P inhibit, Hh signaling activity. We further found that PI(4)P directly binds Smo through an arginine motif, which then triggers Smo phosphorylation and activation. Moreover, we identified the pleckstrin homology (PH) domain of G protein-coupled receptor kinase 2 (Gprk2) as an essential component for enriching PI(4)P and facilitating Smo activation. PI(4)P also binds mouse Smo (mSmo) and promotes its phosphorylation and ciliary accumulation. Finally, Hh treatment increases the interaction between Smo and PI(4)P but decreases the interaction between Ptc and PI(4)P, indicating that, in addition to promoting PI(4)P production, Hh regulates the pool of PI(4)P associated with Ptc and Smo.
| The Hedgehog (Hh) signaling pathway plays important roles in both embryonic development and adult tissue homeostasis. A critical step in Hh signal transduction is the inhibition of Smoothened (Smo), an atypical G protein-coupled receptor (GPCR), by the Hh receptor Patched (Ptc). It is a longstanding question how Ptc inhibits Smo and how Hh promotes Smo phosphorylation and activation. It is unlikely that Ptc inhibits Smo by direct interaction. Here, we uncover that phosphatidylinositol 4-phosphate (PI(4)P), a specific phospholipid, directly interacts with Smo through an arginine motif in the Smo C-terminal tail and promotes Smo phosphorylation, activation, and ciliary localization. Ptc also interacts with PI(4)P, which is inhibited by Hh stimulation, indicating that Hh triggers the release of PI(4)P from Ptc. We further uncover that Hh stimulation induces the overall production of PI(4)P, likely by regulating PI(4)P kinase and phosphatase. Finally, in addition to the direct role in regulating Smo phosphorylation, G protein-coupled receptor kinase 2 (Gprk2) facilitates PI(4)P interaction with Smo. This study suggest that PI(4)P acts as a special small molecule shuttling between Ptc and Smo to modulate Hh responses.
| The Hedgehog (Hh) signaling pathway plays important roles in both embryonic development and adult tissue homeostasis [1–3]. In Drosophila, the Hh signal is transduced through a receptor system at the plasma membrane, which includes the receptor complex Patched-Interference Hh (Ptc-Ihog) and the signal transducer Smo [4–6]. Binding of Hh to Ptc-Ihog relieves the Ptc-mediated inhibition of Smo, which allows Smo to activate the cubitus interruptus (Ci)/Gli family of zinc finger transcription factors and thereby induce the expression of Hh target genes such as decapentaplegic (dpp), ptc, and engrailed (en) [7,8]. Over the last 30 years, many Hh pathway components have been identified, including those that control transmission, propagation, receipt, and transduction of the Hh signal. However, it is still unclear how Ptc inhibits Smo to block the activation of the Hh pathway and how Ptc inhibition of Smo is relieved by Hh stimulation. It is unlikely that Ptc inhibits Smo by direct association [9,10], as the inhibition occurs even when Smo is present in 50-fold molar excess of Ptc, and substochiometric levels of Ptc can repress Smo activation [10,11]. These findings suggest that the inhibition process is catalytic [10]. The involvement of small molecules, rather than a protein ligand, has been proposed: Ptc may inhibit the production of positive regulators or promote the synthesis of inhibitory molecules [10].
Smo, an atypical G protein-coupled receptor (GPCR), is essential in both insects and mammals for transduction of the Hh signal [8,12,13]. The activation of Smo appears to be one of the most important events in Hh signaling. Hh induces cell surface accumulation and phosphorylation of Smo [9] by multiple kinases, including protein kinase A (PKA), casein kinase 1 (CK1) [14–16], casein kinase 2 (CK2) [17], G protein-coupled receptor kinase 2 (Gprk2) [18], and atypical PKC (aPKC) [19]. These phosphorylation events activate Smo by inducing a conformational change [20] to promote Smo interaction with the Costal2-Fused (Cos2-Fu) protein complex [21–23]. It is believed that Hh-induced phosphorylation counteracts the autoinhibition imposed by arginine clusters in the Smo C-terminal tail (C-tail), which induces an open conformation that promotes the dimerization of Smo proteins [1,20]. Similar to other GPCRs, Smo cell surface accumulation is controlled by endocytic trafficking mediated by ubiquitination [24,25]. In mammals, Hh signal transduction depends on the primary cilium, and Smo accumulation in the cilium is required for Smo activation [26–28]. Therefore, the cilium represents a signaling center for the Hh pathway in mammals [29]. Phosphorylation by multiple kinases promotes the ciliary localization of mammalian Smo [30], but it remains unclear how Smo cell surface or ciliary accumulation and intracellular trafficking are controlled. A previous study has shown that mutation in INPP5E, a lipid 5-phosphotase, results in signaling defects in primary cilium [31], indicating a role for phospholipids to regulate the function of cilium. In the preparation of this manuscript, two studies found that phospholipids regulate ciliary protein trafficking [32,33]; however, it is unknown whether and how phospholipids directly regulate Smo.
A very well characterized system for studying Hh signaling is the Drosophila wing disc. Hh proteins expressed and secreted from the posterior (P) compartment cells act on neighboring anterior (A) compartment cells located adjacent to the A/P boundary to induce the expression of Dpp [34,35]. As a morphogen, Dpp diffuses bidirectionally into both the A and P compartments to control the growth and patterning of cells in the entire wing [36–38]. Other genes, including en and ptc, are also induced by Hh to specify cell patterning at the A/P boundary [39,40]. Expression of dpp monitors the low levels of Hh activity, and ptc expression indicates higher levels of Hh activity, whereas en induction appears to be an indicator of the highest doses of Hh signaling activity [39]. The transcription factor Ci is only expressed in A compartment cells that receive the Hh signal.
In this study, we found that Hh stimulation increases the levels of phosphatidylinositol 4-phosphate (PI(4)P) in both wing discs and cultured cells. We further found that PI(4)P activates Smo by promoting Smo phosphorylation. Mechanistically, we identified an arginine motif in the Smo C-tail that is responsible for the interaction of Smo with PI(4)P. Arginine to alanine mutation abolishes, whereas arginine to glutamic acid mutation elevates, Smo activity. We also found that, in addition to the kinase activity of Gprk2, its pleckstrin homology (PH) domain increases PI(4)P in wing discs and is required for Gprk2 to fully function in Hh signaling. The findings that Ptc interacts with PI(4)P and that Ptc inactivation increases the levels of PI(4)P indicate that PI(4)P acts downstream of Ptc to activate Smo in the Hh signaling cascade. Finally, we show that PI(4)P promotes phosphorylation and ciliary accumulation of mouse Smo (mSmo) in mammalian cells, and that PI(4)P prevents the ciliary accumulation of mouse Ptc1 and Ptc2. Taken together, our findings suggest that PI(4)P acts as a special small molecule shuttling between Ptc and Smo to modulate Hh responses.
In an effort to identify novel regulators in Hh signaling, we collected RNAi lines from the Vienna Drosophila Resource Center (VDRC) when the library was available and screened for kinases, phosphatases, and E3 ubiquitin-protein ligases using the wing-specific MS1096-Gal4. We tested selected RNAi lines for the ability to modify the phenotype of SmoPKA12, a weak dominant negative form of Smo that results in a reproducible wing phenotype with partial fusion between Vein 3 and Vein 4 when combined with the C765-Gal4 (S1D Fig), which represents a very sensitive genetic background for screening Smo regulators [19,25,41]. One of the “hits” was Stt4 kinase, the yeast homolog of PI4KIIIalpha required for the generation of PI(4)P. We found that, although knockdown of Stt4 alone did not produce any change in the wild-type wing (S1B Fig), Stt4 RNAi combined with SmoPKA12 expression enhanced the fusion of Vein 3 and Vein 4 (S1E Fig). In contrast, inactivation of Sac1 phosphatase, which dephosphorylates PI(4)P to phosphatidylinositol (PI), attenuated the fusion phenotype (S1F Fig). These results suggest that Stt4 and Sac1 regulate the activity of Smo in the wing. Consistently, Sac1 RNAi partially rescued the abdominal cuticle loss caused by Hh RNAi, although Sac1 RNAi alone did not show any cuticle phenotype (S1G–S1J Fig). A recent study established a genetic link between Smo and Stt4-Sac1 [42]; however, the molecular mechanisms are unclear.
Smo accumulates in P compartment cells as well as A compartment cells near the A/P boundary (Fig 1A) [9,15]. We found that the level of PI(4)P is elevated in the A compartment cells that abut the A/P border (Fig 1B), suggesting that Hh induces the accumulation of both Smo and PI(4)P in these cells. Consistently, the accumulation of Smo in Drosophila embryo correlated with the accumulation of PI(4)P (S2A–S2C Fig). We further found that PI(4)P levels were increased by the expression of Ci-3P and SmoSD123, the constitutively active forms of Ci and Smo, respectively (S2D and S2E Fig). To accurately measure and quantify the absolute concentration of PIP in cells, we established a mass spectrometry-based multiple reaction monitor (MRM) assay and examined whether Hh indeed induces the production of PIP. Based on a detailed method published recently for quantifying PIP2 (PI(3,4)P2, PI(3,5)P2, PI(4,5)P2), and PIP3 (PI(3,4,5)P3) [43], we optimized the conditions to examine PIP lipids. Trimethylsilyl diazomethane was used to protect the phosphate groups, which allowed for more efficient ionization of the methylated PIP species and a marked improvement in the sensitivity of the assay. We found that treatment of S2 cells with 60% Hh-conditioned medium [44] induced the formation of PIP in a timely manner (Fig 1C, left panel). Consistent with this, treatment of NIH3T3 mouse fibroblasts with mouse sonic Hh N-terminus (ShhNp) [30] stimulated the production of PIP (Fig 1C, right panel). Total PIP was quantified, since this assay was unable to distinguish PI(4)P from PI(3)P and PI(5)P.
To further characterize the regulation of PI(4)P by Hh, we used an enzyme-linked immunosorbent assay (ELISA) and found that Hh stimulated the production of PI(4)P in S2 cells in a concentration-dependent manner (Fig 1D, left panel). In addition, knockdown of Stt4 downregulated, whereas knockdown of Sac1 upregulated, the production of PI(4)P (Fig 1D, left panel), suggesting that the Hh-regulated formation of PI(4)P was mediated by Stt4 and Sac1. We further found that overexpression of Ptc prevented the production of PI(4)P, whereas RNAi-mediated knockdown of Ptc elevated production (Fig 1D, right panel), suggesting that Ptc regulates the levels of PI(4)P. To delineate the involvement of PI(4)P in Hh signaling, we used a ptc-luciferase (ptc-luc) reporter assay to monitor the activity of Hh signaling [44] and found that Hh-induced ptc-luc activity was suppressed by RNAi of Stt4 but elevated by RNAi of Sac1 (Fig 1E). Furthermore, treatment with PI(4)P and the expression of Inp54p (a PI(4,5)P2-specific phosphatase to produce PI(4)P) enhanced, whereas IpgD (converts PI(4,5)P2 into PI(5)P) suppressed, the basal and Hh-induced ptc-luc activity (Figs 1F and S3A). As a control, Inp54pD281A phosphatase-dead mutant had no effect on ptc-luc activity (Fig 1F). These data suggest that PI(4)P is a specific phospholipid that regulates Hh signaling in cultured S2 cells.
The PH domain is a known phosphoinositide-binding module that is important for signal transduction by sensing alterations in the membrane lipid composition. To visualize PI(4)P pools in wing discs, we used an RFP-PHOSBP reporter that contains two copies of the PH domain from the oxysterol binding protein (OSBP), which is known to specifically bind PI(4)P [45]. In wing discs, expression of RFP-PHOSBP accumulated PI(4)P (Fig 2B) and Smo (Fig 2C), compared to the expression of RFP alone (Fig 2A). In cultured S2 cells, treatment with PI(4)P enhanced Smo activity, indicated by an elevated ptc-luc reporter activity (S3B Fig), thus prompting the question of whether PI(4)P regulates Smo phosphorylation, since phosphorylation promotes Smo activation. Indeed, we found that PI(4)P, but not other phospholipid forms, increased the levels of basal and Hh-induced Smo phosphorylation detected by a phospho-specific antibody (SmoP) [44] recognizing phosphorylation within the second PKA/CK1 cluster (Fig 2D). In addition, PI(4)P treatment induced Smo phosphorylation to a lesser extent compared to Hh treatment, and the combination of Hh and PI(4)P induced hyperphosphorylation of Smo (Fig 2E). Consistently, treatment with PI(4)P induced mSmo phosphorylation in cultured NIH3T3 cells, which was detected by a phospho-specific antibody (PS1) [30] recognizing mSmo phosphorylation at the first CK1/GRK cluster (Fig 2F). In an in vitro kinase assay using glutathione S-transferase (GST)-Smo fusion protein containing Smo amino acids 656–755, we found that Smo phosphorylation by PKA and CK1 kinases was enhanced by the addition of PI(4)P, but not PI(4,5)P2 or PIP3 (Fig 2G), suggesting that PI(4)P directly regulates the phosphorylation of Smo. In support of this notion, we found that Smo interacted with the PH domain from OSBP, and that this interaction was enhanced by the treatment with either PI(4)P or Hh (Fig 2H and 2I).
It is possible that PI(4)P directly interacts with Smo and facilitates Smo interaction with the PH domain of OSBP. To test this, we used a solid-phase lipid-binding assay and found that purified full-length Myc-SmoWT strongly associated with PI(4)P, and weaker binding to PI(5)P was detected as well (Fig 3A, left column). Because the level of PI(5)P is much lower than that of PI(4)P in cells [46], PI(4)P is likely the primary lipid that binds Smo. We further found that Myc-SmoΔC (Smo lacking the C-tail) did not interact with PI(4)P (Fig 3A, right column), suggesting that the C-tail of Smo is required for binding to PI(4)P.
To identify the residues in the Smo C-tail interacting with PI(4)P, we used an in vivo approach to examine whether the activity of Smo is regulated by the PH domain expression. Membrane-tethered Smo C-terminal truncations by the myristoylation signal (Myr-SmoCT) possess the activity to induce ectopic dpp-lacZ expression [21]. We found that the membrane-tethered PH domain of OSBP (Myr-PHOSBP) increased the ectopic dpp-lacZ expression induced by Myr-Smo730-1035 but did not change dpp-lacZ expression induced by Myr-Smo764-1035 (compare Fig 3E to 3D and 3G to 3F), although Myr-PHOSBP itself had no effect on dpp-lacZ expression in the wing. This suggests that PHOSBP likely regulates Smo activity through aa 730–764. This domain of Smo contains one of the four positively charged arginine clusters, which are known to negatively regulate Smo activity by counteracting phosphorylation (Fig 3B) [20]. Surprisingly, using GST-Smo656-755 in the solid-phase lipid-binding assay, we found that SmoWT strongly interacted with PI(4)P; however, an Arg to Ala mutation (GST-SmoRA4) abolished this interaction (Fig 3H), suggesting PI(4)P interacts with only the fourth arginine cluster. In support of this finding, mutation in the fourth arginine cluster (R4) was sufficient to block PI(4)P binding in a PI(4)P beads pull-down assay (Fig 3I). To further characterize R4, we generated the Arg to Ala mutation in Smo full-length (Myc-SmoRA4) and found that Myc-SmoRA4 lost both the ability to bind PI(4)P (Fig 3J) and the interaction with Cos2-Fu complex (Fig 3K). In addition, phosphorylation of SmoRA4 was no longer regulated by Hh and PI(4)P (Fig 3K). SmoRA4 also had no responsiveness to PI(4)P stimulation in the ptc-luc assay (S4E Fig). Using the fluorescence resonance energy transfer (FRET) assay to test C-terminal CFP/YFP dimerization [20,44], we found that SmoRA4 had a much lower FRET signal and was much less responsive to both Hh and PI(4)P stimulation compared to SmoWT (S4F Fig). RA4 mutation did not cause protein misfolding because there was no difference between SmoRA4 and SmoWT regarding the expression and subcellular distribution in S2 cells. Taken together, our findings suggest that SmoRA4 is inactive. Indeed, driven by the tubulinα promoter that expresses Smo at a level close to endogenous gene expression [17], the expression of SmoRA4 did not rescue ptc and en expression in smo mutant cells (Figs 3L, 3M and S4G). Furthermore, we found that overexpression of SmoRA4 by MS1096-Gal4 did not induce ectopic expression of dpp-lacZ, as noted with SmoWT (Fig 3N and 3O), and that an R>A mutation in the constitutively active form of Smo (SmoSD123RA4) had lower ptc-luc activity compared to SmoSD123 (S4E Fig) and was unable to induce en expression in ventral A compartment cells (Fig 3R, compared to Fig 3Q). These data suggest that Smo activity is compromised by R>A mutation in the fourth arginine motif. In contrast, we found that R>E mutation (SmoRE4), which mimics negative charges caused by PI(4)P binding, elevated Smo activity to induce higher levels of dpp-lacZ expression in the wing disc (Fig 3P) and higher levels of ptc-luc activity, but had no responsiveness to PI(4)P stimulation (S4E Fig). Taken together, our findings suggest that the fourth arginine motif is required for Smo activation.
The binding position of PI(4)P in Smo is very critical, because fusion of the PH domain from OSBP to either the third intracellular loop (SmoL3PH) or the C-tail (SmoPH) retained PI(4)P with Smo (S4A Fig) but compromised Smo activity in the wing (S4C and S4D Fig, compared to S4B Fig) and resulted in loss of responsiveness to PI(4)P stimulation (S4E Fig). In comparison, fusion of CFP and YPF to the third intracellular loop and C-tail, respectively, did not change the activity of Smo [20]. Our findings suggest that PI(4)P binds Smo in a position-dependent manner.
Considering that the PH domain of OSBP interacts with and activates Smo, we wondered whether a PI(4)P transport protein (PITP) facilitates the interaction between PI(4)P and Smo, since Smo itself does not contain a PH domain. We used RNAi lines from the VDRC to screen a total of 15 typical PH domain-containing PITPs in the fly genome for their ability to modulate Hh phenotypes; inactivation of these proteins by RNAi did not affect Smo accumulation in wing discs, although RNAi of some candidate PITPs modified the wing phenotype of C765-SmoPKA12 (S1 Table). Interestingly, all Gprks contain a PH domain in their C-terminus, and this domain contributes to agonist-dependent translocation by facilitating interaction with lipids and other membrane proteins [47,48]. We next investigated whether the C-terminus PH domain of Gprk2 was important for its role in Hh signal transduction. Wild-type Gprk2 fully rescued en expression in gprk2 mutant cells (Fig 4A). However, deletion of the PH domain in the Gprk2 C-tail (Grpk2ΔC) abolished its ability to rescue en expression (Fig 4B), whereas replacing the PH domain in Gprk2 with the PH domain from OSBP (Gprk2-PHOSBP) restored this ability (Fig 4C). This is consistent with our previous finding that Gprk2KM, a kinase-dead form of Gprk2, has a kinase activity-independent role in regulating Smo [18]. These findings suggest that the PH domain is required for Gprk2 to fully function in transducing the Hh signal. In support of these results, Gprk2, Gprk2ΔC, and Gprk2-PHOSBP, but not Gprk2KM, were able to phosphorylate mSmo in vitro (Fig 5A), indicating that the removal or replacement of the PH domain does not affect the kinase activity. Thus, the function of the Gprk2 PH domain likely accounts for the kinase-independent role of Gprk2 in Smo regulation.
Because both Gprk2 transcription and Gprk2 protein expression are upregulated by Hh signaling, and Gprk2 is enriched at the A/P boundary [18,49], we hypothesize that, in addition to promoting the production of PI(4)P, Hh may regulate PI(4)P accumulation by enhancing the expression of Gprk2 as the endogenous carrier for PI(4)P. To examine the ability of Gprk2 to enrich PI(4)P in vivo, we knocked down Gprk2 in the wing disc and found that the levels of PI(4)P were decreased (S5A Fig). We also overexpressed Gprk2 or Gprk2ΔC and found that the expression of Gprk2 elevated the levels of both Smo and PI(4)P (Fig 4D and 4E), whereas the expression of Gprk2ΔC had no effect (Fig 4G and 4H). Similar to RFP-PHOSBP (Fig 2B and 2C), overexpression of PHGprk2 resulted in increased PI(4)P and Smo accumulation (S5B and S5C Fig). In addition, Gprk2 and PI(4)P were largely localized at the cell surface (Fig 4F and 4F′), whereas Gprk2ΔC was cytosolic (Fig 4I). These results suggest that the PH domain of Gprk2 is required for the enrichment of PI(4)P in vivo by localizing Gprk2 at the cell surface.
To further characterize the kinase activity-independent role of Gprk2 in regulating Smo, we examined Gprk2-regulated Smo phosphorylation in cultured S2 cells. We found that RNAi targeting the coding region of Gprk2, but not OSBP, attenuated PI(4)P-induced Smo phosphorylation detected by the anti-SmoP antibody (Fig 5C). RNAi targeting the 3′-UTR region of Gprk2 consistently inhibited Smo phosphorylation (Fig 5D, lane 4, top panel). We found that the expression of HA-Gprk2 or HA-Gprk2-PHOSBP rescued Smo phosphorylation inhibited by RNAi of Gprk2 3′-UTR but the expression of HA-Gprk2ΔC did not (Fig 5D), suggesting that the PH domain is responsible for Gprk2 to promote Smo phosphorylation increased by PI(4)P. We also found that deletion of the PH domain decreased the Gprk2-PI(4)P interaction in the PI(4)P beads pull-down assay (Fig 5E). Moreover, the PH domain of Gprk2 (PHGprk2) interacted directly with PI(4)P; mutation of arginine (PHGprk2RA) or phenylalanine (PHGprk2FA) abolished this interaction (Fig 5F). Finally, similar to the PHOSBP interaction with Smo (Fig 2H and 2I), the PHGprk2 interaction with Myc-SmoWT was increased by Hh and PI(4)P treatments in cultured S2 cells (Fig 5G and 5H). Taken together with the observation that deletion of the PH domain does not alter the kinase activity of Gprk2 in vitro (Fig 5A) and in cultured NIH3T3 cells (Fig 5B), our findings suggest that the Gprk2 PH domain plays a positive role in mediating Smo regulation by PI(4)P.
The finding that expression of the PH domain from OSBP accumulated PI(4)P (Fig 2B) prompted the notion that an endogenous protein may attract PI(4)P away from Smo in the absence of Hh. Ptc contains a sterol-sensing domain (SSD) and has structural similarity to the resistance, nodulation, division (RND) family of bacterial proton gradient-driven transmembrane molecular transporter [50]. SSD was first identified in proteins implicated in cholesterol metabolism but is now more broadly associated with vesicle trafficking. The Ptc SSD is essential for suppression of Smo activity [51], and mutations of SSD abrogate the Ptc-mediated repression of Smo, although these mutations do not compromise either binding or internalization of Hh [10,52]. It is possibe that the Ptc SSD controls the influx or the efflux of PI(4)P or attracts PI(4)P away from Smo. To test this hypothesis, we generated three vectors: HA-tagged wild-type full-length Ptc (HA-PtcWT), HA-tagged Ptc lacking its SSD domain (HA-PtcΔSSD), and HA-tagged SSD domain (HA-SSD). We transfected S2 cells with these constructs and evaluated the ability of each to interact with PI(4)P. When expressed in S2 cells, all proteins were expressed at low levels, detected only after immunoprecipitation (Fig 6A, top panel). We found that HA-PtcWT and HA-SSD strongly bound PI(4)P, whereas HA-PtcΔSSD did not bind (Fig 6A, lower panel).
To further determine whether the SSD domain from Ptc directly interacts with PI(4)P, we used the solid phase lipid-binding assay similar to that used for detecting Smo binding. We found that the SSD fragment protein purified from bacteria strongly associated with PI(4)P, but not with PIP2 or PIP3 phospholipids (Fig 6B), suggesting that the interaction between SSD and PI(4)P in the lipid beads protein pull-down assay is direct. The SSD association with PI(3)P or PI(5)P (Fig 6B) suggests that the expression of a single SSD domain may lose specificity for interaction, or, alternatively, that such interaction may also promote Ptc regulation of PI(3)P and PI(5)P. We also found a very strong interaction between SSD and phosphatidic acid (PA) or phosphatidylserine (PS) (Fig 6B); these may be nonspecific, as PA and PS binding to short protein fragments has often been considered questionable [53]. Our findings in cultured cells led us to examine the correlation of Ptc and PI(4)P in the wing. We found that mutation of ptc or knockdown of Ptc by RNAi increased PI(4)P levels in the wing disc (Fig 6C and 6D), similar to the observation that Ptc inactivation elevates PI(4)P in the salivary gland [42]. These indicate that the activation of Hh signaling by the inactivation of Ptc elevated the production of PI(4)P. Moreover, we found that the overexpression of PtcWT also increased the level of PI(4)P (Fig 6E), which is likely due to the ability of Ptc to accumulate PI(4)P. In support of these findings, PtcΔSSD overexpression had no effect on regulating the accumulation of Smo, Ci, and PI(4)P in wing discs.
In addition to promoting the production of PI(4)P, Hh may also regulate the pools of PI(4)P between Smo and Ptc. To test this hypothesis, we purified Smo and Ptc proteins from S2 cells treated with Hh-conditioned medium or control medium and accessed the protein interaction with PI(4)P. As shown in Fig 6E, the level of PI(4)P-bound Smo was increased by the treatment of Hh (Fig 6F, left panel). In contrast, the level of PI(4)P-bound Ptc was decreased by Hh treatment (Fig 6F, right panel). These data indicate that Hh treatment releases PI(4)P from Ptc, suggesting an additional layer of regulation beyond Hh promoting PI(4)P production.
It would be interesting to understand how Hh regulates PI(4)P. However, we found that Hh treatment did not significantly change the mRNA levels of Stt4 and Sac1 (S6A Fig). In addition, Hh did not change the protein levels of the overexpressed Stt4 and Sac1 in P compartment cells of the wing disc (S6B and S6C Fig). Hh also did not regulate the accessibility of the Stt4/Sac1 to Smo or Ptc in an immunoprecipitation assay with S2 cells. To examine whether Hh regulates the activity of Stt4 or Sac1, or both, we carried out in vitro kinase/phosphatase assays using purified Stt4 and Sac1 combined with PI substrate. We found that the phosphorylation of PI was enhanced when using Stt4 from cells treated with Hh (Fig 6G, lane 3, compared to lane 2, left panel). In addition, Sac1 from cells treated with Hh had less activity to dephosphorylate the constitutive PI phosphorylation (Fig 6G, lane 3, compared to lane 2, right panel). These data suggest that Hh elevates the activity of Stt4 and inhibits the activity of Sac1.
Next, we wondered whether PI(4)P plays a role in regulating mSmo, because PI(4)P induces mSmo phosphorylation (Fig 2F). We first tested different phospholipids for their effects in activating mSmo and found that, similar to Shh, PI(4)P treatment elevated mSmo activity as monitored by a Gli-luc reporter (Fig 7B). In contrast, PIP2 and PIP3 treatment had no effect on mSmo activity. In addition, the activity of the constitutively active form of mSmo (mSmoSD), which mimics mSmo phosphorylation by GRK2 and CK1 [30], was further increased by PI(4)P (Fig 7B). Consistently, Drosophila SmoSD123 activity was increased by PI(4)P (S4E Fig).
Similar to Drosophila Smo, mSmo contains arginine clusters in its C-tail (Fig 7A) [20]. We next examined whether the arginine motif(s) were responsible for regulation of mSmo by PI(4)P. As shown in Fig 7C, PI(4)P treatment increased Gli-luc reporter activity, although to a lesser extent compared to Shh treatment in the control group of NIH3T3 cells. PI(4)P and Shh treatment consistently increased Gli-luc activity when cells were transfected with mSmoWT (Fig 7C). However, R>A mutations in R3 and R4 arginine clusters (mSmoRA3 and mSmoRA4, respectively) attenuated the increased activity noted with PI(4)P, and mutations in both R3 and R4 (mSmoRA34) completely blocked the effect of PI(4)P on mSmo activation (Fig 7C). These data suggest that R3 and R4 are responsible for the regulation of mSmo by PI(4)P.
Phosphorylation promotes ciliary accumulation of mSmo, which correlates with pathway activation [30], but the molecular mechanisms that control Smo ciliary accumulation are poorly understood. mSmoWT was found in about 5% of cilia, and Shh treatment increased SmoWT accumulation in 75% of cilia (Fig 7D and 7E) [30]. Treatment with PI(4)P induced mSmoWT accumulation in 47% of cilia (Fig 7D and 7E), which was correlated with changes in mSmoWT phosphorylation (Fig 2F) and activity (Fig 7B) induced by PI(4)P. In contrast, mSmoRA34 had no response to PI(4)P treatment (4% of ciliary accumulation by PI(4)P treatment) (Fig 7D and 7E), although it had a low response to Shh stimulation (from 3% to 21% of ciliary localization). Consistent with these findings, mSmoRA34 had much lower activity and much less responsiveness to Shh stimulation in a previous study [20]. Our findings suggest that R3 and R4 clusters are responsible for PI(4)P-associated binding and activation of mSmo.
To determine whether Hh regulates the production of PI(4)P in vertebrate systems, using the ELISA assay combined with the anti-PI(4)P antibody, we examined the levels of PI(4)P in ptc1 mutant mouse embryonic fibroblasts (MEFs) and found that, compared to control MEFs, ptc1 MEFs had significantly increased PI(4)P (Fig 7F). Consistently, Shh treatment increased, whereas Smo inhibitor decreased, the levels of PI(4)P (S7A Fig), indicating that Hh signaling activity promotes PI(4)P production in cultured cells. We further investigated the ciliary localization of Ptc1 and Ptc2 and found that the ciliary localization of both Ptc1 and Ptc2 was decreased by Shh treatment or PI(4)P treatment (Fig 7G), which was consistent with the previous study that Hh inhibits the ciliary localization of Ptc1 [54]. Similar to Drosophila Ptc interaction with PI(4)P, we found that both Ptc1 and Ptc2 interacted with PI(4)P in the lipid beads protein pull-down assay (Fig 7H). These observations indicate a consistent regulation of Smo and Ptc by PI(4)P in Drosophila and mammalian systems. To incorporate the findings in this study and the findings published recently [32,33,42], we proposed a model in which Smo phosphorylation and ciliary accumulation is regulated by PI(4)P (Fig 8).
Hh signal transduction has been widely studied; however, the longstanding questions of how Ptc inhibits Smo activity and how Hh promotes Smo phosphorylation and activation remain. These issues constitute a primary focus of this study. A genetic analysis indicates that inactivation of Stt4 downregulates Smo accumulation, whereas knockdown of Sac1 by RNAi elevates Smo levels [42], suggesting the involvement of phospholipids in Hh signal transduction. Nevertheless, the mechanisms by which phospholipids regulate Hh signaling remain unknown. In this study, we identified and characterized a direct role for PI(4)P in binding Smo and promoting Smo phosphorylation and activation. The mechanism by which PI(4)P interacts with arginine motifs in Smo is likely conserved between Drosophila Smo and mouse Smo, as such motifs have been mapped in both species.
To explore the possible involvement of a PITP protein to facilitate the interaction between Smo and PI(4)P, we unexpectedly found that the PH domain of Gprk2 is responsible for the accumulation of PI(4)P that activates Smo. It has been shown that Gprk2 is positively involved in Hh signaling by directly phosphorylating Smo C-tail [18]. In addition, Gprk2 forms a dimer/oligomer and binds Smo C-tail in a kinase activity-independent manner to promote Smo dimerization and activation [18]. However, how Gprk2 promotes Smo dimerization and activation is unclear. In this study, we found that the function for Grpk2 PH domain to activate Smo is independent of its kinase activity (Fig 5A and 5B) and that the PH domain of Gprk2 is responsible for enriching PI(4)P that promotes Smo phosphorylation and dimerization. There are instances in which the binding of lipids to the PH domain promotes dimerization of the protein [55,56], raising the possibility that PI(4)P interaction with the PH domain of Gprk2 promotes its dimerization. Taken together, our findings suggest that the function of PH domain in Gprk2 accounts, at least in part, for the kinase activity-independent role of Gprk2 in Hh signaling, a deeper mechanism for Smo activation by Gprk2.
In the absence of Hh, Ptc inhibits Smo activity by promoting Smo endocytosis and turnover in intracellular compartments [9]. Ptc likely inhibits Smo catalytically [10], because substochiometric levels of Ptc are able to repress Smo activation [10,11]. Here, we found that Hh promotes the activity of Stt4 and inhibits the activity of Sac1 (Fig 6G), which, at least in part, explains the catalytic regulation.
A previous study proposed a model in which Ptc represses Smo by regulating lipid trafficking; Ptc recruits lipoproteins to endosomes, changing their lipid composition, in order to regulate Smo degradation [57], but the class of lipids remains unidentified. In the presence of Hh, Smo is phosphorylated and accumulates at the cell surface, resulting in protein activation [1,8]. However, it is unknown how Ptc inhibition on Smo is relieved by Hh stimulation. The ability of the Ptc SSD domain to interact with PI(4)P (Fig 6A and 6B) raises the possibility that Ptc may control the pool of phospholipids regulating the accessibility of Smo to PI(4)P. Our finding that Hh treatment decreased the interaction between Ptc and PI(4)P (Fig 6F) suggests the possibility that binding of Hh to Ptc results in a conformational change in Ptc and releases phospholipids. Thus, this study uncovers an additional layer of regulation by indicating the release of PI(4)P from Ptc, which may account for the optimal regulation of Smo.
The structure of the Smo N-terminal, including the extracellular cysteine rich domain (CRD), has been characterized [58,59]. Unlike other GPCRs, no ligand-binding function has been identified. It has been shown that Smo-mediated signal transduction is sensitive to sterols and oxysterol derivatives of cholesterol [60–62]. However, unlike vertebrate Smo, Drosophila Smo CRD does not interact with oxysterols [63]. In this study, we found that phospholipids activate both vertebrate and Drosophila Smo through binding to the arginine motif in the Smo C-terminus, although the C-tails have sequence divergence among species. Using the protein:lipid overlay assay, we found that PI(4)P directly binds Smo (Fig 3A) and that mutation in the R4 arginine motif abolishes this direct interaction (Fig 3H–3J). Importantly, R>A mutation abolished the activity of Smo (Figs 3L, 3M and S4G). It is likely that binding of PI(4)P to the arginine motif changes Smo conformation, thus allowing kinases to phosphorylate and activate Smo. In support of this notion, PI(4)P binding to specific arginine residues in specific locations is critical for Smo conformational change, because fusion of the PH domain to either the third intracellular loop or the C-terminus attracts PI(4)P to different locations in Smo, thus blocking Smo activation by PI(4)P (S4 Fig).
In this study, we focused on the regulation of Smo by PI(4)P and found that Hh regulates the accessibility of Smo to PI(4)P, evidenced by the Hh-promoted interaction between Smo and PH domain from either OSBP or Gprk2 (Figs 2I and 5E) and by the Hh-enhanced interaction between Smo and PI(4)P (Fig 6F). Binding of PI(4)P likely changes the conformation of Smo, leading to Smo phosphorylation by kinases. It should be noted that phosphomimetic Smo mutations (SmoSD123 and mSmoSD) are still regulated by PI(4)P (Figs 7B and S4E), suggesting that PI(4)P either promotes phosphorylation at other residues or has additional role(s) in activating Smo.
Hh signaling activity promoted the production of PI(4)P that was detected by the mass-spec assay (Fig 1C), which was a very sensitive approach. However, in a previous study, the overexpression of full-length wild-type Ci did not elevate the accumulation of PI(4)P in wing discs [42]. It is possible that low levels of Hh signaling activity induced by the expression of wild-type Ci are unable to induce detectable changes in PI(4)P accumulation in wing discs. The disc immunostaining with the anti-PI(4)P antibody might not be as sensitive as the mass-spec method. In support, by expressing SmoSD123 or the constitutively active Ci-3P, in which three PKA sites in the phosphorylation clusters were mutated to block Ci processing [64], we found that PI(4)P was accumulated in wing discs (S2A Fig), likely due to the high levels of Hh signaling activity induced by the active forms of Smo and Ci.
Generation of the Myc-SmoWT, Myc-SmoΔC, Myr-Smo730-1035, and Myr-Smo764-1035 constructs and transgenes was previously described [21]. Myc-SmoPH and Myc-SmoL3PH were constructed by fusing one copy of the OSBP PH domain (aa 17–123) to the Smo C-terminus and two copies of the OSBP PH domain after aa 451 of Smo intracellular loop 3, respectively. Generation of the Myc-SmoSD123 and Myc-SmoRA1234 was previously described [15,20]. Myc-SmoSD123RA4 was generated by the combination of SD123 and RA4. Myc-tagged SmoRA1, SmoRA2, SmoRA3, SmoRA4, SmoRA12, and SmoRA34 were generated by PCR. To construct tub-SmoRA4, the previously described tubulinα promoter [17] was inserted upstream to the SmoRA4 sequence. Transgenic lines were generated using the VK5 attP locus to ensure Smo protein expression at the same levels without positional effects. Genotypes for examining the activity of Smo transgene in smo clones include: yw hsp-flp/+ or Y, smo3 FRT40/hs-GFP FRT40, and tub-SmoRA4/+. RFP-PHOSBP and RFP-PHGprk2 were constructed by an in-frame fusion of two copies of the PH domain from OSBP or Gprk2 at the RFP C-terminus in the attB-UAST backbone [41]. RFP-PHOSBP and RFP-PHGprk2 transgenes were generated by insertion at the VK5 attP locus. As a control, the UAS-RFP transgenic line was generated with the same approach. Flag-Gprk2WT and Flag-Gprk2KM constructs and transgenes have been described [18]. Flag-tagged Gprk2ΔC (containing Gprk2 aa 1–666) and Gprk2-PHOSBP (aa 667–714 replaced by the PH domain of OSBP) were inserted in-frame in the Flag-UAST vector, and their transgenic lines were generated using standard P-element-mediated transformation. Multiple independent transgenes were generated, and those on the second chromosomes were used for rescue experiments. Gprk2 mutant clones were generated with yw 122 and FRT82 Gprk2/FRT82 hs-Myc-GFP. ptc mutant clones were generated with yw hsp-flp/+ or Y and ptc[wII] FRT42D/hs-GFP FRT42D. Inp54p (from Saccharomyces cerevisiae, Addgene 20155) and IpgD (from Shigella flexneri, a gift from Dr. Frederique Gaits-Iacovoni) were subcloned into HA-UAST backbones. A phosphatase-dead version of Inp54p with a D281A mutation (Inp54pD281A) was generated by site-directed mutagenesis. GST-SmoWT containing aa 656–755 has been described [15]. GST-tagged SmoRA12, SmoRA34, and SmoRA4 were generated by PCR on the backbone of GST-SmoWT. GST-PHGprk2, GST-PHGprk2RA, and GST-PHGprk2FA were constructed by the same approach. HA-PtcWT, HA-PtcΔSSD, and HA-SSD were generated by PCR with Ptc or its fragments inserted into HA-UAST backbones, and transgenic lines were generated using the VK5 attP locus. His-SSD was generated by in-frame fusion of Ptc SSD domain (aa421-589) to pET30a backbone. Myc-tagged mSmoWT, mSmoRA mutants, and mSmoSD (previous mSmoSD0-5) have been described [20,30]. GFP-tagged mSmoWT and mSmoRA34 were generated by subcloning each cDNA into pEGFP-N1 backbone. Flag-tagged Stt4 (Flag-Stt4) was generated by combining seven RT-PCR fragments with unique cloning sites into the SK+ backbone and, finally, subcloning into Flag-UAST backbone. HA-tagged Sac1 (HA-Sac1) was constructed by inserting the RT-PCR fragment into the HA-UAST backbone. pGE-mPtch1-Myc and pGE-mPtch2-Myc were generated by subclone of the open reading frame from pEGFP-mPtch1 and pEGFP-mPtch2 (Gift from Dr. Chi-Chung Hui), respectively. MS1096-Gal4, ap-Gal4, C765-Gal4, and Gprk2 deletion mutants have been described [17,18]. Stt4 RNAi lines (v15993 and v105614), Sac1 RNAi lines (v44376, v37216, and #56013), Ptc RNAi line (#28795), Hh RNAi line (v1402), GFP-Sac1 (#57356 and #57357), and prd-Gal4 (#1947) were obtained from VDRC or Bloomington Stock Center (BSC). Generation of Gprk2 RNAi lines have been described [18]. The RNAi lines that targeted each PITP were obtained from either VDRC or BSC (S1 Table).
PtdIns lipid extraction and quantification by mass spectrometry were carried out as previously described [43]. Briefly, 3 × 107 S2 cells (or 4 × 106 NIH3T3 cells) from a 100 mm dish were harvested and washed once with ice-cold PBS and suspended in 340 μL H2O and 1500 μL quench mix with 25 ng of 17:0–20:4 PI(4)P (Avanti Polar Lipids, Inc.) in 10 μL methanol as the internal standard. Lipids were extracted with 1450 μL CHCl3 and 340 μL 2 M HCl and subsequently derivatized with trimethylsilyl diazomethane (Sigma), as previously described [43]. After washing and drying under a stream of nitrogen at room temperature, samples were dissolved in 200 μL methanol. We applied 10 μl of each sample to LC-MS/MS analysis using a Shimadzu LC-20 HPLC and TSQ Vantage triple quadrupole mass spectrometer (ThermoFisher). A Jupiter 5μ C4 300A (50 × 1.0 mm) column (Phenomenex) was used with the multiple reaction monitoring (MRM) transitions described [43]. PIP concentrations were calculated from MRM peak areas and the internal standard and were subsequently normalized to cell number.
For PI(4)P quantification by ELISA assay, similar PtdIns lipid extraction procedures were used without adding the PI(4)P internal control. After washing and drying with nitrogen stream, lipid extracts were dissolved in ethanol and loaded into a microplate, dried under a vacuum, and incubated with 2% BSA in PBS at room temperature for 30 min. Mouse anti-PI(4)P or anti-PI(4,5)P2 monoclonal antibodies (Echelon Biosciences) were added for 1 h, followed by goat anti-Mouse IgG-HRP (Jackson ImmunoResearch) for 30 min, with 3 PBS washes after each inculation. Finally, chemiluminescence substrate (SuperSignal West Pico, Pierce) was added to the microplate, and luminescence intensity was determined by a luminometer.
GST-Smo fusion proteins expressed in bacteria were pooled by GST beads (GE Healthcare), then eluted with elution buffer (10mM Glutathione pH 8.0 in 50 mM Tris) at 4°C overnight. Myc-tagged Smo proteins expressed in S2 cells were immunoprecipitated with anti-Myc antibody combined with beads of protein A ultralink resin, followed by two sequential elutions with Myc peptide (Roche, 100mM KCl, 20% glycerol, 20 mM HEPES KOH, pH 7.9, 0.2 mM EDTA, 0.1% NP-40, 5 mM DTT, and 0.5 mM PMSF). The eluted purified GST-fusion proteins and Myc-tagged Smo proteins were concentrated by the Centrifugal filter units (Millipore) and incubated with lipid-dotted strips according to manufacturer’s instruction (Echelon Biosciences), followed by western blot with the anti-GST (Santa Cruz), anti-Myc (Santa Cruz), or anti-His (Millipore) antibodies.
For the PI(4)P beads pull-down experiments, HA-tagged Gprk2 proteins were expressed in S2 cells, immunoprecipitated with mouse anti-HA antibody (F7, Santa Cruz), eluted with HA peptide (Sigma, in 500 mM NaCl), and concentrated by the Centrifugal filter units (Millipore). GST-PHGprk2, GST-PHGprk2RA, and GST-PHGprk2FA proteins were expressed in bacteria and purified using the protocol employed for GST-Smo purification. The purified and concentrated GST-fusion proteins or epitope-tagged proteins were incubated with PI(4)P beads (Echelon Biosciences) with wash/binding buffer (10 mM HEPES, pH7.4; 0.25% NP-40; 150 mM NaCl), and subjected to western blot to detect PI(4)P bound proteins.
Wing discs from third instar larvae were dissected in PBS and then fixed with 4% formaldehyde in PBS for 20 min. After permeabilization with 1% PBST, discs were incubated with primary antibodies for 3 h and appropriate secondary antibodies for 1 h, and washed three times with PBST after each incubation. Affinity-purified secondary antibodies (Jackson ImmunoResearch) for multiple labeling were used. It was a challenge for disc staining with the mouse anti-PI(4)P antibody. We have adopted/modified a critical method for PI(4)P immunostaining from previous publications [46,65]. Discs were fixed in 4% formaldehyde in PBS and permeabilized in 1 M sucrose by freezing at -80°C for 1 h followed by thawing at room temperature. Then discs were washed with PBS and incubated with 50 mM NH4Cl for 15 min, followed by incubation with the anti-PI(4)P antibody at 4°C overnight. For Drosophila embryo primary antibody staining, stage 11 fly embryos with specific genotypes were dechorionated, fixed with Heptane solution, and immunostained with similar procedures. To examine mSmo ciliary localization, NIH3T3 cells were transfected with mSmo-GFP variants, treated with Shh or PI(4)P, and immunostained for mSmo localization in the cilium. Primary antibodies in this study were: mouse anti-Myc (9E10, Santa Cruz), anti-Flag (M2, Sigma), anti-SmoN (DSHB), anti-En (DSHB), and anti-PI(4)P (Z-P004, Echelon Biosciences); rabbit anti-β-Gal (Cappel), anti-GFP (Clontech), anti-Acetylated tubulin (Sigma), anti-PS1 [30], and rat anti-Ci (2A1, DSHB). Affinity-purified secondary antibodies (Jackson ImmunoResearch) for multiple labeling were used. Fluorescence signals were acquired on an Olympus confocal microscope and images processed with Olympus Fluoview Ver.1.7c. About 15 imaginal discs were screened and three to five disc images were taken for each genotype.
|
10.1371/journal.pgen.1002194 | LGI2 Truncation Causes a Remitting Focal Epilepsy in Dogs | One quadrillion synapses are laid in the first two years of postnatal construction of the human brain, which are then pruned until age 10 to 500 trillion synapses composing the final network. Genetic epilepsies are the most common neurological diseases with onset during pruning, affecting 0.5% of 2–10-year-old children, and these epilepsies are often characterized by spontaneous remission. We previously described a remitting epilepsy in the Lagotto romagnolo canine breed. Here, we identify the gene defect and affected neurochemical pathway. We reconstructed a large Lagotto pedigree of around 34 affected animals. Using genome-wide association in 11 discordant sib-pairs from this pedigree, we mapped the disease locus to a 1.7 Mb region of homozygosity in chromosome 3 where we identified a protein-truncating mutation in the Lgi2 gene, a homologue of the human epilepsy gene LGI1. We show that LGI2, like LGI1, is neuronally secreted and acts on metalloproteinase-lacking members of the ADAM family of neuronal receptors, which function in synapse remodeling, and that LGI2 truncation, like LGI1 truncations, prevents secretion and ADAM interaction. The resulting epilepsy onsets at around seven weeks (equivalent to human two years), and remits by four months (human eight years), versus onset after age eight in the majority of human patients with LGI1 mutations. Finally, we show that Lgi2 is expressed highly in the immediate post-natal period until halfway through pruning, unlike Lgi1, which is expressed in the latter part of pruning and beyond. LGI2 acts at least in part through the same ADAM receptors as LGI1, but earlier, ensuring electrical stability (absence of epilepsy) during pruning years, preceding this same function performed by LGI1 in later years. LGI2 should be considered a candidate gene for common remitting childhood epilepsies, and LGI2-to-LGI1 transition for mechanisms of childhood epilepsy remission.
| Major remodeling of the neuronal synaptic network occurs during childhood. The quadrillion synapses formed till the end of age two are trimmed to 500 trillion by age 10 through a selective process of strengthening of ideal connections, removal of redundant ones, and formation of new contacts. Very little is known about the basic mechanisms that direct this massive reorganization that leads to the adult brain. The most common epilepsies of humans occur in childhood and are characterized by remission prior to adulthood. Not much is known about their genetics and basic remission mechanisms. We describe here a canine equivalent disease and identify the defective gene, Lgi2. We show that the gene product is a secreted protein and interacts with neuronal ADAM receptors known to be involved in the regulation of synaptic remodeling in the developing brain. Our work sheds important light on the basic mechanisms of the most common neurological disease of children and discloses processes of epilepsy remission. The identification of the first focal epilepsy gene in dogs has also enabled the development of a genetic test to identify carriers for breeding purposes.
| Postnatal mammalian brain development proceeds in three phases the first of which is construction of the primary neural network (ages zero to two years in humans, zero to one week in mice, and estimated zero to one to two months in dogs). In humans, this phase generates a network of approximately one quadrillion synapses. The second phase, pruning (ages two to 10 years in humans, seven to 17 days in mice, and estimated two to four months in dogs), is chiefly characterized by massive removal of unneeded or otherwise inappropriate synapses, almost half the original synapses. The third and final phase is the remainder of life, during which synapse numbers remain stable [1]–[3].
Epilepsies are by far the most common neurological diseases in children two to 10 years of age, the three most common of which are Rolandic Epilepsy, Panayiotopoulos syndrome, and Childhood Absence Epilepsy (CAE). The first two of these three syndromes are focal-onset epilepsies where seizures start from defined brain regions, while CAE is a generalized epilepsy where seizures appear to start simultaneously from all brain regions. All three syndromes share a remarkable feature of remission after age 10, i.e. after network pruning is complete [4]. All three are genetically complex syndromes, and paucity of gene information has impeded their understanding, including how and why they remit. To date, a few ion channel mutations (e.g. in GABRG2, CACNA1H) have been found in CAE, accounting for far less than 1% of patients with this syndrome [5].
While the above three syndromes begin and end during the pruning phase of neurodevelopment in the vast majority of cases, other genetic epilepsies begin near or after the end of this phase, i.e. after age eight in most cases. These include the generalized Juvenile Myoclonic Epilepsy (JME) (to date with mutations in the EFHC1 or GABRA1 genes; penetrance ∼50%) [6], [7] and the focal-onset Autosomal Dominant Lateral Temporal Lobe Epilepsy (ADLTE) (also called Autosomal Dominant Partial Epilepsy with Auditory Features) with mutations in the LGI1 gene [8] (penetrance 67%) [9]. JME is generally a non-remitting lifelong epilepsy [10]. Remission rate in ADLTE has not been determined, although the literature indicates that most cases remain on seizure medications, unlike Rolandic epilepsy, e.g., where the vast majority do not [10], [11].
In the present work, we show that mutation of the Lgi2 gene, a gene closely related to Lgi1, causes remitting focal-onset epilepsy in dogs between ages one and four months, which is equivalent to human two to eight years. LGI2 belongs to a family of four closely related neuronal proteins including the well-studied LGI1. We report functional and expression studies of LGI2, which, combined with previous LGI1 studies, suggest a novel concept of the basis of remission common in childhood epilepsy.
The Lagotto Romagnolo is an ancient curly-haired water dog (water dogs, or water spaniels, originally served to retrieve game falling in water), which was selected in Italy to become an excellent truffle hunter. The popularity of the breed fluctuated with the truffle industry and in the early 1970s underwent a strong genetic bottleneck to near extinction, when a group of dog lovers decided to save it. The breed has since gained popularity for reasons unrelated to truffle or water hunting, and its numbers are in the thousands spread across most developed countries (http://www.lagottoromagnolo.org/).
The breed is affected by an epilepsy, Benign Familial Juvenile Epilepsy (BFJE), described in detail in reference [12]. Onset is at five to nine weeks of age, and the epilepsy invariably completely remits by four months of age. Remission is so reliable that the epilepsy is considered by many breeders as an unfortunate particularity of the breed and often disregarded. The seizures consist of whole-body tremors sometimes associated with alteration of consciousness. Electroencephalography (EEG) reveals unilateral epileptic discharges in central-parietal and occipital lobes, and magnetic resonance imaging (MRI) is normal. During the months with epilepsy the animals are often ataxic, but this resolves completely as the seizures disappear [12].
Towards the goal of mapping and identifying the BFJE gene we first reconstructed a large multinational Lagotto pedigree from which an example with 212 Finnish dogs including 34 cases is shown in Figure S1. The dogs live in homes as private pets, often in different countries in Europe, or were still with their breeders. Disease segregation suggested autosomal recessive inheritance (Figure S1). Next, we performed a single nucleotide polymorphism (SNP) genome-wide association study with DNA from 11 of the affected dogs and 11 unaffected littermates (discordant sib-pairs) (Figure S1) and found very strong association in a region of chromosome 3 (CFA3), peaking at the marker at base-pair 89159216 (Praw 0.000035; Pgenomewide 0.08) (Figure 1A and 1B). There was no significant association at any other genomic locus, the next best association being over 100-fold less significant (Figure 1A). Genotype analysis around the 89159216 SNP revealed a 1.7 Mb block of homozygous SNPs between markers at 87.3 Mb and 89.0 Mb in the 11 cases and none of the controls (Figure 1B). This region contains nine genes, including Lgi2. Sequencing Lgi2 revealed an exonic homozygous protein-truncating sequence change, c.1552A>T (p.K518X), in all 11 affected and none of the 11 unaffected animals (Figure 1C). Genotyping a cohort of 140 dogs for the 89159216 SNP, for the Lgi2 c.1552 sequence change, and for three additional SNPs from the homozygous region revealed extremely high associations including Praw 4.47×10−16 at 89159216 and Praw 1.05×10−23 (the highest association) at Lgi2 c.1552 (Table 1). These results strongly suggested that Lgi2 c.1552A>T (p.K518X) is the BFJE mutation.
Next we studied segregation of the sequence change in the pedigree. Of the 28 affected dogs from which we had samples, 26 (93%) were homozygous Lgi2 c.1552T (p.518X) (i.e. homozygous for the nonsense codon), two were heterozygous (7%), and none was homozygous for the wild-type (wt) A nucleotide. The two affected dogs that were heterozygous were also heterozygous for the 13 SNP haplotype around the Lgi2 locus, and we found no evidence for compound heterozygosity as all other variants in the gene were synonymous (Table S1). These results suggested that if the Lgi2 c.1552A>T (p.K518X) change is the BFJE mutation, it can, in a minority of cases, cause the epilepsy heterozygously. To explore this further, we screened an independent set of 36 sporadic Lagottos and found three homozygous for c.1552T, 14 heterozygous (39%), and 19 wild-type (wt). All three dogs homozygous for c.1552T had the syndrome, as did one of the carriers (7%), appearing to confirm the 7% rate of disease through heterozygosity, assuming that Lgi2 c.1552A>T (p.K518X) is causative.
Among 112 unaffecteds of the genotyped 140 dogs, 69 were homozygous for the wt A nucleotide, 41 were heterozygous, and two, 1.8%, were homozygous for c.1552T (OR = 532, 95%CI: 95.0-5747.1 and p = 1.05×10−23). The latter two may be mis-specified as unaffected - clinical information on many of the dogs in the pedigree was obtained through retrospective questionnaires, and it is possible that a breeder missed seizures, as the epilepsy in some cases is mild and short-lived [2]. Alternatively, these two cases may represent incomplete penetrance, assuming, again, that the sequence change we identified is causative. Similarly, other recent recessive gene discoveries indicate incomplete penetrance including canine lens luxation [13], degenerative myoelopathy [14] and a form of neuronal ceroid lipofuscinosis [15].
At this point, there were two possibilities. Either Lgi2 c.1552A>T (p.K518X) is the BFJE mutation with an incomplete penetrance in a minority of cases, or it is not the causative variant. To gather more data we proceeded with functional studies of the consequences of the truncating sequence change on the LGI2 protein.
We first determined whether the c.1552A>T sequence change prevents Lgi2 mRNA expression, e.g. through mRNA instability. RT-PCR experiments showed no mRNA reduction (Figure 2).
LGI2 belongs to a family of neuronally secreted proteins (LGI1 to LGI4) conserved across mammals and composed of N-terminal leucine-rich repeats (LRR) and C-termini containing seven EPTP repeats [3]. K518X truncates LGI2 within the seventh EPTP repeat (exon 8 of the gene) (Figure S2). Similar mutations truncating LGI1 in the EPTP repeats in humans, including in the seventh repeat, cause ADLTE, the human epilepsy with most commonly onset after age eight and persistence through adulthood. Where studied, the vast majority of LGI1 mutations, truncating or otherwise, prevent secretion of the protein encoded by the mutant allele, and ADLTE is therefore usually a disease due to lack of neuronal secretion of half the required amount of LGI [11], [16], [17]. We asked whether the LGI2 K518X truncation prevents LGI2 secretion. We performed western blot experiments with V5-tagged wt and mutant LGI2 transfected in HEK293 cells and found that while both proteins were present in cell lysates only wt LGI2 was found in the culture medium (Figure 3), indicating that the truncation prevents secretion.
Following secretion, LGI1 interacts with a subfamily of the ADAM (a-disintegrin-and-metalloproteinase) family of neuronal membrane proteins [18], [19]. Members of this subfamily, ADAM11, ADAM22 (post-synaptic), and ADAM23 (pre-synaptic), lack the metalloproteinase domain that other ADAMs use to convey extracellular signals intracellularly [19]. To determine whether LGI2 also binds ADAM22, ADAM23 and ADAM11 following secretion, we performed immunofluorescent cell surface-binding assays [18] in permeabilized and non-permeabilized cells by co-expressing wt or truncated LGI2 with different ADAMs. Wt LGI2 was secreted and then bound ADAM22, ADAM23 and ADAM11 expressed on the cell surface (ADAM11 result not shown). Truncated LGI2 was not secreted and did not bind the ADAMs (Figure 4A–4B). We also performed co-immunoprecipitation in rat brain and found that both Adam22 and Adam23 antibodies co-precipitated Lgi2 (Figure 5). In summary, wt LGI2 binds the same ADAM substrates of LGI1 following secretion, and the Lagotto K518X mutation prevents secretion and ADAM interaction, in the same fashion as the well-characterized truncating LGI1 epilepsy mutations.
Summarizing the results to this point, the genome-wide association study revealed a highly significant association in the vicinity of Lgi2, an extremely strong association (p = 1.05×10−23) with the protein-truncating c.1552A>T (p.K518X) sequence change in the gene, and no significant association with any other locus. Lgi2 is a close homologue of the epilepsy (ADLTE) gene LGI1, and the Lgi2 truncating mutation is closely similar to the most common type of epilepsy-causing mutations in LGI1. The consequence of the truncation on Lgi2 is identical to the consequence of truncation on LGI1, prevention of neuronal secretion and binding to ADAM receptors, which is presently the most favored mechanism of epileptogenesis in ADLTE. Finally, LGI1 mutations, including truncation mutations, are non-penetrant in 33% of individuals, compared to 1.8% non-penetrance in the case of the canine Lgi2 truncation. Considering all the above, we believe the data meet the burden of proof that Lgi2 c.1552A>T (p.K518X) is the BFJE mutation. BFJE is transmitted in imperfect Mendelian fashion. In the vast majority of cases, 93%, homozygous mutation is required for the disease to manifest. In a minority, 7%, heterozygosity suffices. Conversely, 1.8% of dogs may be resistant to seizing despite homozygous mutation. Finally, we found no dog with homozygous wt genotype at Lgi2 c.1552 that has BFJE.
Mouse studies show that LGI1 starts being expressed midway through the synapse pruning phase of brain development (after postnatal day 13 (p13)), and gradually increases to reach high and stable adult levels by the end of this phase (after p17) [20], [21]. Not surprisingly, the mice lacking LGI1 develop seizures after mid-phase pruning [22] and the great majority of human patients with ADLTE have onset of their epilepsy after age 10 years, the end of the pruning phase in humans, with the remaining few having initial seizures in the latter half of this phase [11], [16], [17]. However, it is worth noting that the very first seizures in ADLTE are often auditory seizures, which might not initially be appreciated to be seizures and might have occurred earlier than what is currently documented in the literature. Because BFJE occurs only in ages equivalent to human two to eight years, we sought to determine whether expression of LGI2 differs from that of LGI1. We examined the expression profile of all four LGI genes in adult tissues using the human GeneSapiens expression database [4] and found that the amount of LGI2 in adult brain is much lower than that of the other three (Figure S3). We next chased Lgi2 expression levels in mouse forebrain and cerebellum by performing quantitative RT-PCR every other day from birth till 27 days. Lgi2 expression in the cerebellum did not change appreciably over this time (Figure 6A). Expression in the forebrain, on the other hand, was highest at birth and through phase one of postnatal development (neural network construction phase), and declined to half the original amount by midway through the pruning phase (Figure 6B). Considering that BFJE occurs only during the pruning phase, these results suggest that LGI2's main functions take place in the developmental phase preceding the phase in which the epilepsy occurs.
Epilepsy is a common symptom of insult to the brain from various causes including tumors, trauma, stroke, and neurodegenerative disease. For example, the most common cause of epilepsy in the elderly is stroke and in neonates hypoxic ischemic encephalopathy. However, epilepsy can also be a disease onto itself, where seizures are the only or preponderant neurological symptom, i.e. the brain is normal except for its propensity to seize. Resolving the basic mechanisms of this type of ‘pure’ epilepsy is expected to provide the clearest insights into epileptogenesis. As mentioned, these pure epilepsy syndromes (sometimes called idiopathic, Greek for ‘disease onto itself’) are the commonest neurological diseases with onset in two to 10 year-old children, and in this age group most are genetic, commonly polygenic, and often characterized by remission in adolescence [5].
Genetic-idiopathic epilepsy syndromes are the most common neurological diseases of dogs, in some breeds 10 times more common than in humans [23]. In the present work we identify the first of the canine idiopathic epilepsy genes, in a remitting syndrome with onset and offset equivalent to human childhood two to 10 years. This epilepsy can now be eliminated from the Lagotto through selective breeding. Carrier frequency of the mutation is very high. We tested 576 Lagottos from three different countries and found a carrier rate of 32% (Table S2). On the other hand, the mutation appears restricted to this particular epilepsy in this particular breed. We tested 121 epileptic dogs from 40 different breeds, including Barbets, a Lagotto-related French water spaniel breed afflicted with a separate epilepsy, and none carries the BFJE mutation (Table S3).
Genetic epilepsies of various types, as simple or complex traits, are highly enriched in various canine breeds, including Miniature Wirehaired Dacshunds [24], Finnish Spitzs [25] and Belgian Shepherds [23] due to pure-breeding. Each of these traits is in genetic isolation within its corresponding breed. This vastly improves signal to noise ratio in genetic studies compared to human populations [26], which should facilitate mapping epilepsy genes. The Lagotto themselves segregate a second epilepsy with onset in adulthood completely distinct from BFJE [12]. We have established that this second epilepsy is not associated with the BFJE mutation (Table S1), and are presently mapping its gene(s). Five out of the six adult-onset cases with persistent seizures in our pedigree were genotyped and only one of them was homozygous for the BFJE mutation. However, the puppyhood history of this case is unknown and it was impossible to confirm retrospectively whether this case has also had BFJE. This case has an affected littermate with classical BFJE who is homozygous for the mutation. On the other hand, all the other genotyped adult-onset dogs were wildtypes strongly suggesting that this form of epilepsy has its own genetic cause, and that this single homozygous case may have suffered from both BFJE and the adult-onset form of epilepsy.
The BFJE gene is a homolog of the human epilepsy gene LGI1. LGI1 is neuronally secreted and binds three metalloproteinase-lacking ADAM receptors. Significant progress has started to be made in elucidating LGI1's functions at these receptors. LGI1 interaction with post-synaptic ADAM22 strengthens and stabilizes ADAM22-containing synapses [18], [21], [22]. Interaction with pre-synaptic ADAM23 enhances neurite outgrowth from ADAM23-containing axons [27]. Through its seven-bladed β-propeller structure (encoded by the EPTP repeats), LGI1 simultaneously binds ADAM23 and ADAM22, pulling pre and post-synaptic membranes together, physically stabilizing synapses containing these two proteins and strengthening neurotransmission in these synapses [22]. Importantly, LGI1 regulates neuronal terminal pruning and maturation, again through a combined pre and post-synaptic action [21].
Humans not expressing or secreting LGI1 from one allele develop epilepsy starting in the vast majority of cases after age eight and seeming to persist in adulthood in many cases. Mice completely lacking LGI1 are normal until midway through the pruning phase of brain development (∼P13), when LGI1 would normally have started being expressed, after which they develop seizures that progressively worsen as LGI1's amounts would normally have progressively increased, and die of violent convulsions by four weeks of life [22], [28], [29]. These results show that LGI1 is a vital protein, vital specifically in protecting the brain against seizures. In humans its partial loss results in epilepsy, and only epilepsy, and in the mice its complete loss leads to death from epilepsy prior to the presence of any other neurological symptom. This vital anti-epileptic role is mediated at least in part through the above three ADAM receptors suggesting that the LGI1-ADAM complexes and their related pathways are essential components of neural network electrical stability in the maturing and mature brain.
We show in our study that lack of secretion of LGI2 is also associated with epilepsy, at an earlier stage of development, and secreted LGI2 interacts with the same ADAM receptors as LGI1, suggesting that LGI2 participates in protecting the brain against seizures during the pruning phase of neurodevelopment at least in part through the same system utilized by LGI1 in the subsequent phase. Importantly, LGI2 expression is highest in the phase preceding pruning and epilepsy. This suggests that the LGI2 anti-epileptic activity anticipates the pruning phase, i.e. LGI2 acts during the network construction phase to help prepare a network that will not seize during the pruning phase. To date, there has been no compelling evidence-based theory of why so many epilepsies of childhood begin and end with the start and end of the pruning phase. Our results, combined with the body of LGI1 work, suggest the following. Construction of the initial network includes mechanisms, in which LGI2 participates, that ensure that the network will not seize during the pruning phase. Defects in these anticipatory anti-epileptic processes result in epilepsy as the massive changes of the pruning phase commence. The pruning phase itself encompasses mechanisms, in which LGI1 participates, that ensure that the pruned and remodeled network to serve the rest of the animal's life is electrically stable. These mechanisms are able to correct or compensate for earlier instabilities, e.g. those introduced by LGI2 deficiency, resulting in the remission that characterizes so many childhood idiopathic epilepsies.
Of the remaining two LGI proteins, LGI4 appears not to have a major role in the central nervous system. Instead, its chief function appears to be in regulating neuron-Schwann cell interaction, its secretion defect resulting in inability of Schwann cells to correctly myelinate peripheral nerves, resulting in peripheral nervous system hypomyelination and the murine claw-paw phenotype [30]. LGI3's function, on the other hand, appears to be similar to that of LGI1, as there is evidence that like LGI1 it regulates neurite outgrowth [31]. LGI3 starts being expressed at p7 in mouse, i.e. at the very start of the pruning phase, has steady and high expression throughout the brain in adulthood, and interacts with ADAM22 and ADAM23, as well as presynaptic SNARE complexes [31]–[34]. However, LGI3 does not rescue the LGI1 knockout mouse epilepsy [22], and therefore the two proteins are at least not interchangeable.
Four phenotypes have been associated with LGI1. The first is normalcy in the up to 33% of patients with heterozygous mutations in ADLTE families. Second is ADLTE in the remaining patients with heterozygous LGI1 mutations. The third is the recent realization that the acquired autoimmune epilepsy syndrome Limbic Encephalitis, long thought to be due to auto-antibodies against a potassium channel, is in fact due to an auto-antibody against LGI1. As expected, all patients with this condition are over 10 years of age [35]. The final phenotype is the intractable and fatal mid-pruning phase-onset murine epilepsy caused by complete LGI1 deficiency. In the clinic, we not infrequently encounter previously normal children with onset of explosive catastrophic epilepsy. Homozygous mutations in LGI1 (and possibly LGI3) should be considered as a possible cause of this presentation.
The phenotype associated with LGI2 in the present study is of a remitting epilepsy with focal onset and centrotemporal and occipital spikes on EEG, occurring within the age range equivalent to human two to 10 years. Two of the most common human epilepsy syndromes occur in this age range, Rolandic Epilepsy, which is focal in onset with centrotemporal spikes on EEG, and Panayiotopoulos Syndrome, again focal-onset, with both centrotemporal and occipital spikes on EEG. LGI2 should be considered a candidate gene in these common epilepsies.
We have collected blood samples from privately owned pets for our genetic studies and have a valid ethical permission for the proposed blood sampling in the study, ESLH-2009-07827/Ym-23 (Oct 2009–Oct 2012) from the Animal Ethic Committee, The State Provincial Office of Southern Finland, P.O. B150, 13101 Hämeenlinna.
Mapping of the benign focal juvenile epilepsy (BFJE) locus in Lagotto Romagnolo was based on clinically studied litters from Finland [12], German and Switzerland including 25 epileptic puppies, 17 healthy littermates and 12 parents. Furthermore, based on the retrospective questionnaire-based phenotype information, we expanded our study cohort to a total of 112 healthy LR dogs and 28 BFJE cases and collected also 36 sporadic dogs. Additionally, the study population included also five adult-onset epilepsy LR cases [12] and our clinically diagnosed juvenile epilepsy cases from other breeds including Barbets, Collies and German Shepherds (Table S1). Population-based allele and genotype frequencies were estimated from a population of 576 Lagotto samples from three different countries (Table S2). EDTA-blood samples were collected and genomic DNA was extracted using a commercially available kit (Puregene, Gentra Systems, Minneapolis, MN). The Finnish Kennel Club's breeding database, Koiranet, was utilized for pedigrees.
Altogether 22 dogs including seven discordant full sibs and four half-sibs were selected for GWAS. Genotyping was performed with Affymetrix's Canine SNP Array version 1 containing 26,578 markers (Affymetrix, Santa Clara, CA). The SNP association analysis was performed with PLINK software [36] with the criteria of MAF <0.05, call rate >75% and <25% of missing genotypes in individual dogs. After applying these filters, 17,273 SNPs remained in the analysis for all dogs. Genome-wide significance was ascertained through 10 000 random permutations of epilepsy phenotype.
Exons and splice junctions were amplified by PCR with primers listed in . The PCR products were purified with ExoSAP-IT kit (USB Corporation, Cleveland, Ohio) and sequenced with an ABI Prism 3730xl DNA analyzer (Applied Biosystems, Foster City, CA). To confirm that the Lgi2 is the causative gene in the associated 1.7 Mb region we sequenced four SNPs around the associated homozygozity region together with the mutation in 112 healthy and 28 epileptic Lagottos. Odds ratio was calculated using conditional maximum likelihood estimation and corresponding 95% CI was calculated from Fisher exact test. The calculations were done with R statistical software package. Absence of the mutation in other breeds was studied by sequencing epileptic cases from altogether 40 different breeds (Table S3).
To study the effect of the nonsense mutation on the stability of Lgi2 transcript, total RNA was isolated from an affected and a healthy Lagotto dog from peripheral blood using PAXgene Blood RNA Kit (PreAnalytix, Hombrechtikon, Switzerland). Total RNA isolated from the cerebellum of a Saluki puppy euthanized due to hernia diaphragmatica was used as an amplification control. cDNA synthesis was performed using the High Capacity RNA-to-cDNA kit (Applied Biosystems, Foster City, CA), and exon 4- and exon 8-specific primers (Table S4) were used to amplify Lgi2 by PCR. Transcriptional profiling of the LGI2 mRNA expression levels across a large number of human tissues was retrieved from the public GeneSapiens (PMID: 18803840) database containing data from a meta-analysis of 9873 samples analyzed using the Affymetrix gene expression microarrays [37].
To study the effect of the mutation on the expression and secretion of LGI2 in cell culture, we obtained the human LGI2 clone from GeneScript Corporation (Piscataway, NJ). The mutant LGI2 clone including the premature stop codon (p.K534X corresponding to canine p.K518X) was prepared from the wt clone and both were cloned into the pcDNA3.1D/V5-His vector in frame with the C-terminal V5-tag using pcDNA3.1 Directional TOPO Expression Kit (Invitrogen, Carlsbad, CA). The recombinant constructs were confirmed by sequencing. HEK293 cells were grown in DMEM-GLUTAMAX medium (Gibco Laboratories, North Andover, MA) supplemented with 10% fetal calf serum (FCS), 100 IU/ml penicillin, 100 µg/ml streptomycin and 1 mM Sodium Pyruvate and transiently transfected with the FuGENE 6 reagent (Roche Diagnostics, Indianapolis, IN) according to the manufacturer's instructions. Expression of the wt and mutant LGI2 were analyzed 48 hours after transfection by immunostaining on Western blots with anti-V5 antibodies from cell lysates and culture media. Media samples were concentrated 100-fold with Sentricon-10k concentrator (Millipore, Billerica, MA) before loading onto gel. GADPH was used as internal loading control using anti-GADPH antibodies. Proteins were visualized using the enhanced chemiluminescence method.
COS7 cells were co-transfected using human V5-tagged wt LGI2B or LGI2B p.K534X mutant or rat FLAG-tagged wt LGI1 with mouse HA-tagged Adam22 or Adam23. LGI1 and ADAM clones were described previously [18]. 24 hours after transfection, cells were fixed with 2% paraformaldehyde at RT for 10 min, blocked with PBS containing 10 mg/ml BSA and stained with anti-Flag or anti-V5 antibodies followed by Cy3-conjugated secondary antibody without permeabilization to visualize only the cell-surface bound LGIs. Then, the cells were permeabilized with 0.1% Triton X-100 for 10 min, blocked with PBS containing 10 mg/ml BSA, and stained with anti-HA polyclonal antibody, followed by Alexa488-conjugated secondary antibody. Fluorescent images were taken with a confocal laser microscopy system (Carl Zeiss LSM 510; Carl Zeis, Oberkochen, Germany).
To study the developmental expression of the Lgi2 transcript a colony of C57/BL6 mouse was established for tissue and RNA harvesting. Every other day after birth (except days 5 and 17) one mouse was sacrificed and the forebrain of the cerebrum and the cerebellum were harvested and deep-frozen in liquid nitrogen before total RNA isolation by the QIAGEN RNeasy mini kit. The isolated RNA was DNase I-treated before RT-PCR by the SuperScript First-Strand Synthesis System (Invitrogen, Carlsbad, CA). Quantitative PCR was performed using a SYBR Green method with MxPro-3005P multiplex Quantitative PCR systems. Lgi2-specific forward, atgtgtacgtggccatcgctca, and reverse, caaacttggtccagctctcgtcgta, primers were used for amplification.
|
10.1371/journal.pcbi.1005192 | Malaria Elimination Campaigns in the Lake Kariba Region of Zambia: A Spatial Dynamical Model | As more regions approach malaria elimination, understanding how different interventions interact to reduce transmission becomes critical. The Lake Kariba area of Southern Province, Zambia, is part of a multi-country elimination effort and presents a particular challenge as it is an interconnected region of variable transmission intensities. In 2012–13, six rounds of mass test-and-treat drug campaigns were carried out in the Lake Kariba region. A spatial dynamical model of malaria transmission in the Lake Kariba area, with transmission and climate modeled at the village scale, was calibrated to the 2012–13 prevalence survey data, with case management rates, insecticide-treated net usage, and drug campaign coverage informed by surveillance. The model captured the spatio-temporal trends of decline and rebound in malaria prevalence in 2012–13 at the village scale. Various interventions implemented between 2016–22 were simulated to compare their effects on reducing regional transmission and achieving and maintaining elimination through 2030. Simulations predict that elimination requires sustaining high coverage with vector control over several years. When vector control measures are well-implemented, targeted mass drug campaigns in high-burden areas further increase the likelihood of elimination, although drug campaigns cannot compensate for insufficient vector control. If infections are regularly imported from outside the region into highly receptive areas, vector control must be maintained within the region until importations cease. Elimination in the Lake Kariba region is possible, although human movement both within and from outside the region risk damaging the success of elimination programs.
| What would it take to eliminate malaria in southern Africa? While the past century has seen many countries eliminate malaria, and many regions have dramatically reduced malaria burden in the last fifteen years, no sub-Saharan African country has yet eliminated malaria. In the Lake Kariba region of Southern Province, Zambia, villages with high and low malaria burden are interconnected through human travel, making elimination potentially very challenging. We used detailed spatial surveillance data—including household locations, climate, clinical malaria incidence, prevalence of malaria infections, and bednet usage rates—to construct a model of interconnected villages in the Lake Kariba region, then tested a variety of intervention scenarios to see which ones could lead to elimination. We found that elimination is indeed possible and requires high, yet realistic, levels of bednet usage among area residents. Mass drug campaigns deployed to kill parasites within the human population can boost the chances of achieving elimination as long as bednet usage is already high.
| Malaria is a vector-borne parasitic disease affecting millions of people worldwide, with Plasmodium falciparum still causing over 400,000 deaths per year [1]. Recent escalation in vector control has greatly reduced global burden and brought many regions close to elimination [2]. In some settings, mass drug campaigns have been an effective tool for depleting the human infectious reservoir and breaking the cycle of transmission, although the effectiveness of such campaigns has been mixed [3].
The ultimate goal is global malaria eradication, but the extreme heterogeneity in transmission levels, vector bionomics, health systems, and population densities limit the applicability of any single strategy [4]. Elimination at a single country level could be challenging to maintain in the presence of high cross-border movements of infected individuals, depending on the country and its regional context [4]. As a result, the concept and implementation of regional malaria eliminations provides a useful staging for progress towards eventual global eradication [5].
Southern Africa is one region where programs are planning operational strategies for elimination [6]. Elimination in southern Africa requires elimination in the Lake Kariba region of Southern Province, Zambia, where areas of high- and low-intensity transmission are interconnected. Understanding how to achieve elimination in the microcosm of the Lake Kariba area would provide a solution to how to achieve elimination in Southern Africa and possibly in a number of other challenging settings.
The Zambia National Malaria Control Centre has successfully scaled up recommended malaria control interventions over the past decade and is pursuing alternative methods to further reduce malaria transmission, including community-targeted parasite reservoir reduction strategies [7,8], with a target elimination date of 2020. Beginning in 2012, mass drug campaigns have been carried out annually in the Lake Kariba region of Southern Province, Zambia (Fig 1), where transmission is seasonal and spatially variable.
Understanding the small-scale variation in the interconnected Lake Kariba region through mathematical modeling provides important insights into the critical thresholds for successful outcomes. Previous modeling efforts have provided guidance for conducting successful campaigns in generic settings [10–12], yet limited work has been done on understanding how a specific geography's individual features can also affect campaign outcomes [13,14] or how spatial variation in vectorial capacity can sustain transmission in interconnected areas. In this work, the most detailed spatial model of a specific geography to date is constructed and used to predict how various factors such as variation in transmission intensity and human migration patterns interact to influence the success of control and elimination efforts.
Local transmission dynamics were reconstructed within a mechanistic model of malaria transmission using the high-resolution surveillance data collected during the mass test-and-treat (MTAT) campaigns of 2012–13 in the Lake Kariba region, which included infection status, age, GPS coordinates of households, recent symptoms, recent treatment, and insecticide-treated net (ITN) usage. Village-scale biting rates were selected to calibrate local malaria prevalence and incidence to longitudinal surveillance data and seasonal patterns of clinical case counts reported at health facilities respectively.
The simulation framework was then used to assess the outcome of a variety of post-2016 intervention scenarios. Simulations predict that high coverage with vector control is a necessary condition of achieving elimination in this region, and mass drug administrations (MDAs) are effective at increasing the likelihood of elimination only if excellent vector control is already in place. Importations of infections into the Lake Kariba area from outside the region present a particular challenge to maintaining elimination if the importations occur in highly receptive areas, and vector control should be sustained in those areas to prevent outbreaks and reestablishment of endemic malaria as long as importations continue.
A high-resolution spatial model of twelve health facility catchment areas (HFCAs) in the Lake Kariba region was configured based on village-scale clusters of households and ITN usage, MTAT coverage, case management, and human migration rates derived from the surveys conducted in 2012–13 (Fig 1 and S1–S7 Figs, Methods). Malaria transmission was modeled within each cluster, where vector populations were driven by cluster-specific climate data and local abundance of larval habitats.
Preliminary entomological data indicated that both Anopheles arabiensis and Anopheles funestus are present in the study area (personal communication with Javan Chanda), with arabiensis biting rates highest between January and April during the warm rainy season while funestus peaks in September at the beginning of the hot dry season (Fig 2A). Relative abundances of arabiensis and funestus govern the seasonality of malaria transmission, while absolute abundances determine the intensity. For each village-scale cluster, combinations of larval habitat availabilities for arabiensis and funestus were simulated with appropriate patterns of ITN usage, case management rates, and MTAT coverage. The resulting simulated prevalence and clinical case counts were compared with surveillance data to select the combination of habitats yielding the best fit to field observations (Fig 3 and S8 Fig, Methods).
Calibration to surveillance data resulted in the expected gradient of higher habitat availability of both vector species closer to Lake Kariba and lower availability in the higher-altitude HFCAs more distant from the lake (Fig 2B). Chiyabi HFCA was predicted to have the largest amount of funestus in this region, suggesting that ITNs may be particularly effective there as funestus is indoor-feeding and highly anthropophilic [15].
The calibrated simulations captured both fine-scale cluster-level variation in malaria prevalence and large-scale spatio-temporal trends of temporary reduction in prevalence observed in the study area following MTAT rounds (Fig 4 and S9 Fig). Surveillance in 2012–13 reported stalled reduction in prevalence by late 2012 and rapid rebound after the following rainy season. This observation was replicated in the spatial model, where significant re-infection between rounds two and three in lakeside clusters drives rebound throughout the study area. The model was able to predict cluster-level prevalence in December 2014 with reasonable accuracy (S10 Fig, r2 = 0.46) considering that local variation in 2014 vector control implementation, location-specific migration patterns, and location-specific changes in health-seeking behavior as a result of the community health worker and case investigation programs that began in mid-2014 were not part of our simplified forward-simulation scenarios.
The observed seasonality of weekly clinical cases confirmed by rapid diagnostic test (RDT) was well-captured by simulations, including a characteristic pattern of high case counts between December and June and a small rise in cases in Chipepo and Sinamalima HFCAs as temperatures rise in October (Fig 5). Seven HFCAs had strong Spearman’s rank correlation between simulation and observed clinical cases (rank correlation values between 0.62 and 0.83), Luumbo and Sinamalima HFCAs had moderate correlation (values 0.45 and 0.51), Chiyabi HFCA had weak correlation (0.23), Gwembe HFCA had very weak correlation (-0.03), and Nyanga Chaamwe HFCA had a weak negative correlation (-0.27) likely due to inadequate data. With the exception of Gwembe HFCA, model fits to clinical cases were better for HFCAs with consistent reporting. Individual clusters also occasionally saw discrepancies between observed seasonality of clinical cases at the local health facility and simulated seasonality at the cluster (Fig 3C). On the modeling side, these differences could be due to limited knowledge of how local climate drives mosquito abundances, including the rates at which temporary larval habitats appear and disappear, or how infection, clinical symptoms, and health-seeking are related in this region.
Many factors drive uncertainty in both magnitude and seasonality of observed case counts. While simulations estimate a constant proportion of clinical cases seek care at health facilities, several factors could contribute to discrepancies between true and observed clinical incidence in the field. Gwembe HFCA, which otherwise has very low transmission, contains the district hospital. Clinical cases reported from Gwembe HFCA may have traveled from elsewhere within or outside the study area expressly to seek care. Case management rates in simulation were estimated from survey responses during the dry season, while individuals with fever during the wet season may show different health-seeking behavior based on road conditions and personal assessment of whether the fever is malarial. A strong distance-dependence on health-seeking means households closer to the health facility contribute disproportionately more to recorded case counts than to true clinical incidence, yet this data shows significant variation (S4B Fig). Clinics report any RDT-positive individual presenting with fever as a clinical case of malaria. However, a substantial portion of individuals in malaria-endemic areas with non-malarial fevers will be RDT-positive due to low-density infection or recent clearance of malaria [16,17]. Finally, the population denominator in this area is unknown, and preliminary analysis from cross-referencing individuals across all surveillance rounds suggests that the true population may be nearly twice as high as the simulation population, which was chosen based on estimates from operations teams working in the region.
Incomplete coverage and imperfect diagnostics in MTAT campaigns result in a significant portion of the parasite reservoir being left untreated [18–21]. This untreated reservoir will resume the cycle of transmission during the next rainy season. However, a mass-distribution mode that can treat undetected infections and a drug formulation such as dihydroartemisinin-piperaquine (DP), which has longer prophylactic protection against re-infection than artemether-lumefantrine (AL), the drug used in the MTATs, may be more successful at interrupting transmission [11].
Operations teams in the Lake Kariba region continued mass treatment following the 2012–13 MTAT by administering MDA and focal MDA to a randomized group of HFCAs. Simulations approximated operational activities in 2014–15 by administering a mass distribution of ITNs in June 2014 and MDA with DP to all HFCAs in December 2014 and July 2015, in line with operational schedules.
To compare how case management, ITN usage, and MDA coverage contribute to reducing malaria burden separately and together, a variety of post-2015 intervention scenarios were simulated (Figs 6–10). For each simulation, the fraction of total study area population living in clusters where no local transmission has occurred over a month-long period was measured for each month between January 2012 and January 2030. Elimination was counted to have been achieved if no infection events occurred anywhere in the simulated study area over a continuous three-year period.
As expected, clusters in high-burden HFCAs were more likely to retain transmission in all non-eliminating scenarios (Fig 6). Increasing passive case management rates without also distributing more ITNs or implementing any MDA campaigns resulted in nearly all low-burden clusters maintaining very low prevalence through January 2030 and likely interrupting local transmission in many low-burden HFCAs (scenario 2). Elimination within the Lake Kariba region thus becomes a question of which interventions are effective at reducing transmission in the high-burden clusters.
Discontinuing MDAs after the 2015 rounds while maintaining current levels of ITN usage did not result in long-term reduction in regional malaria transmission (Fig 7A, scenario 1). Increasing passive case management rates (scenario 2) established a new baseline of lower transmission during the decades following the end of MDAs in 2015 but could not achieve region-wide elimination. Combining the 2014–15 MDAs with ramp-ups in case management and increased ITN usage resulted in elimination under highly aggressive ITN distribution campaigns but not if ITN ramp-ups followed historical rates of increase (Fig 7B and S11 Fig). Under the ramp-up scenario (scenario 3), transmission rebounded after the last ITN distribution in 2022 to a new baseline largely determined by case management rate. In the aggressive scenario (scenario 4), 87% of simulations resulted in elimination.
The aggressive ITN distribution schedule consists of mass distributions with 80% coverage every two years between 2016 and 2022. As a result, 75–90% of the entire study area population is sleeping under an effective net every night over a 9-year period (S11F Fig). Under the ramp-up schedule, ITN usage with effective nets peaks at only 50–80%. This difference in maximum coverage achieved is enough to prevent elimination from occurring in the simulated scenarios.
Can MDA compensate for insufficient vector control? Extending annual dry-season MDAs through 2020 increased the probability of elimination under the aggressive ITN distribution schedule from 0.87 to 0.99 (Fig 8A, scenario 7), but if ITN usage was maintained at current levels or ramped up gradually, elimination was still never observed (scenarios 5 and 6).
Restricting post-2015 MDAs to high-burden HFCAs resulted in outcomes comparable to cases where MDAs were distributed to all HFCAs, with no elimination observed under the ramped ITN distribution and probability of elimination falling slightly from 0.99 to 0.94 under the aggressive ITN distribution (Fig 8B, scenarios 8 and 9), although the boosting effect of MDA on top of the aggressive ITN distribution campaigns was not as large as when MDA was implemented in all HFCAs: adding MDA to high-burden areas only increased probability of elimination from 0.87 to 0.94 while MDA in all areas increased probability of elimination from 0.87 to 0.99. In the small fraction of cases where the combination of aggressive ITN distributions and MDAs did not result in elimination, residual transmission remaining in the densely populated Sinamalima and Chiyabi HFCAs was able to reestablish transmission more broadly in the study area (S12 Fig).
In all scenarios, MDA coverage is simulated at levels observed during the 2012–13 MTAT campaigns. It is possible that MDA coverage could increase in later rounds as campaign logistics are improved. Scenario 10 in Fig 9 shows results of higher-coverage MDAs given on top of the ramped ITN distributions, similar to scenario 6 but with MDA coverage increased to 70%. Elimination is never observed in these simulations. In contrast, even two years of MDA at 70% coverage in high-burden clusters (Fig 9, scenario 11) boosts probability of elimination from 0.87 to 0.99 if high coverage vector control is already present.
Human movement within the Lake Kariba region connects areas of high and low transmission. In scenario 12 (Fig 10), migration rates were increased ten-fold from the base level used in all other scenario projections. Under these conditions, intervention packages that previously resulted in elimination (scenario 7) now have much smaller effects on transmission within the region: elimination is never observed. While the migration rate in scenario 12 is likely to be much higher than the true migration rate, this scenario demonstrates that outcome of an elimination program is highly dependent on the nature and extent of regional human movement. While efforts are beginning to be made to characterize relevant human migration patterns [22–24], understanding local geographies and conditions will ultimately be crucial for predicting whether an elimination program will succeed.
None of scenarios 1–12 include importation from outside the modeled region. While the Lake Kariba districts as a whole represent an isolated area of higher prevalence surrounded by lower-transmission areas [2], it is possible that travelers from further afield may bring infections into the area. In scenario 13, a total of 100 infections are imported twice a year into three highly connected, high-population clusters in Gwembe, Munyumbwe, and Sinamalima HFCAs. This importation rate is likely to be higher than the true importation rate, especially during later years of simulation if prevalence follows recent trends and continues to decline throughout southern Africa, and thus represents an upper bound on the effect of imported infections. Including these importations in a simulation with the same intervention mix as the highly successful scenario 7 resulted in no simulations achieving elimination. Maintaining routine vector control operations such as annual spraying with a long-lasting insecticide is sufficient to prevent widespread resurgence, although elimination remains unlikely as long as importations continue (S13 Fig). Elimination programs for a given area should not be designed in isolation but rather in the context of the broader region, as importation pressure from outside the elimination area can potentially disrupt desired outcomes inside the area.
The complex spatial dynamics of malaria transmission in the Lake Kariba region of Southern Zambia were captured in a mathematical model, which was then extended to compare the effects of future interventions. The simulation framework was informed by household surveillance in 2012–13, clinical case counts reported by local clinics, and preliminary entomological data. This work is the first to examine how village-scale variation in transmission intensity can drive transmission throughout an interconnected region.
While the 2012–13 MTAT surveys were a rich source of data for model-building, they presented an incomplete picture of local malaria transmission because measurements were taken during the dry season. The seasonality of transmission was thus difficult to characterize based solely on these measurements. Weekly clinical case counts reported from health facilities were broadly informative of transmission during the remaining months of the year, but ascertaining the proportion of RDT-positive fevers that are true malarial fevers is difficult. The seasonality of cases observed at the health facility may not reflect seasonality of transmission in every village within the HFCA as each HFCA may encompass a range of transmission intensities. Nonetheless, simulations were able to recapture both region- and village-scale spatio-temporal features of transmission observed during the 2012–13 operations.
Local entomology drives the seasonality of transmission. Attempting to model the study area with only arabiensis vectors requires non-seasonal year-round transmission in many lakeside areas in order to capture substantial prevalence during the November surveys, which would result in a mismatch with observed seasonality of clinical case counts. Incorporating funestus allowed the model to capture both dry season prevalence and seasonal patterns of clinical cases. More entomological data on vector species abundance and behavior will guide continued model refinements.
All simulations assumed full compliance with drug regimens. While perfect adherence to treatment is unlikely [25,26], elimination is possible even when compliance is extremely poor (S14 Fig), and previous modeling work has shown that compliance is not crucial in mass campaigns in settings without drug resistance [11]. Most individuals who harbor dry season infections have low parasite densities [16,18], and a single dose of antimalarial drug is often sufficient to clear those infections.
Excellent vector control is the single most important intervention for achieving elimination. Without sustained high coverage with ITNs, elimination is never observed in any of the scenarios tested, and when ITN usage is maintained at high levels over many years, elimination is observed in 87% of simulation runs. In this study, vector control has been modeled exclusively as ITN usage, but indoor residual spraying (IRS) would have similar effects if local vectors are susceptible. Combining IRS and ITNs likely represents a more attainable scenario for aggressive vector control given that ITN usage has an unpredictable behavioral component while IRS remains in place unless walls are washed or re-plastered. The simulations of the Lake Kariba area presented here are dominated by arabiensis vectors, which have been modeled with a 50% indoor biting rate. Regions where indoor biting predominates may require slightly less intense vector control interventions to achieve the same elimination outcomes.
Unless vector control is also sustained at high coverage, implementing MDA campaigns will not result in elimination and is not a good use of valuable resources. High coverage in drug campaigns is difficult to achieve, and simulations predicted that drug campaigns with coverage as high as 70% could not compensate for insufficient underlying vector control. As MDAs are expensive to administer, we encourage elimination programs to consider whether their operational region meets the preconditions we have discussed here for effective MDA before deciding to implement this intervention.
Simulations predict that if MDAs are implemented on top of a very aggressive vector control program, elimination becomes even more likely. As predicted by generic models [27], however, even in this context transmission can occasionally rebound after vector control ceases (Fig 8 and S12 Fig). Rebound in transmission depends on the extent of human movement as travelers from high-transmission areas re-seed transmission in low-transmission areas. If vector control is well-implemented at high coverage, targeting MDAs to high-burden areas is potentially a reasonable approach to increasing the likelihood of local elimination while saving resources, and could even result in improved outcomes if better coverage with MDA can be achieved in targeted areas.
As long as passive case management improves, simulations predict that malaria transmission will die out in low-burden areas. Human movement inside the study area and importation of infection from outside the study area decrease the likelihood of elimination. If certain areas are known to be hot spots of both importation and receptivity, vector control should be implemented as a preventative measure until elimination is achieved in the wider region and imported infections cease. Surveillance must be vigilant in quickly identifying local outbreaks and responding with high-coverage vector control and possibly focal MDA around index cases. If outbreaks are not adequately contained and endemic malaria is reestablished, sustained high coverage with vector control must be re-implemented and MDA should be considered to rapidly deplete the parasite reservoir before other areas also experience resurgence. Elimination in the face of repeated importations is a major challenge and victory should not be declared too soon, as relaxing surveillance prematurely can easily lead to resurgence in areas with a history of high transmission.
While generic models have predicted that targeting transmission foci would be effective [28], this study demonstrates for the first time that elimination strategies in interconnected regions are well-served by focusing directly on reducing transmission in regional hotspots rather than attempting to maintain an “elimination front” and tackling the most difficult areas last. Programs will need to identify where regional hotspots are located, with the understanding that hotspot locations and intensities can change dynamically as interventions are deployed and changes in vector species composition, human movement patterns, housing conditions, and climatic cycles alter the local landscape of transmission.
As more regions reduce transmission and approach malaria elimination, it becomes crucial to understand how to set up malaria operations for successful and lasting elimination at local, national, and regional levels. Southern Africa is a particular challenge as vectorial capacity is substantial in some areas and the region has become increasingly interconnected. This study suggests that elimination in the southern Africa region will require several years of sustained high coverage with vector control interventions, and mass drug campaigns are unlikely to be effective if vector control is insufficient. In areas with both high receptivity and high importation rates, vector control must be maintained until importation ceases. Ultimately, it follows that regional malaria elimination in southern Africa is nevertheless within reach with current tools, provided the efficacy and operational efficiency attained in the Lake Kariba operational area can be extended and targeted to other key areas.
During each of the 2012 and 2013 dry seasons (June-November) in Southern Province, Zambia, three large-scale MTAT rounds were undertaken [7]. Individuals were visited at their homes in a full community census and, following consent, administered RDTs. Test-positive individuals were treated with the antimalarial drug artemether-lumefantrine (AL), of which the first dose was directly observed. During each MTAT round, RDT results were geo-tagged by household location. Information was collected on household demographics, ITN usage, recent fevers, and recent drug treatments.
In this analysis, focus was restricted to a contiguous block of twelve HFCAs in Gwembe and Sinazongwe Districts along Lake Kariba (Fig 1). These HFCAs cover approximately 80,000 people living in a geographic area spanning a range of endemic malaria transmission intensities. Demographics, migration, ITN usage, treatment-seeking, and drug campaign coverage in simulations were informed by survey data collected during the MTAT rounds. Spatial variation in transmission intensity was captured by selecting local vector larval habitat availabilities to match observed seasonal and spatial patterns in RDT prevalence and clinical incidence.
Given best-fit habitat parameters for each cluster along with cluster-specific climate, population, historical ITN usage, and drug campaign coverage, the set of 115 clusters was simulated together in a spatial model. Cluster climate was simulated using climate data at cluster’s centroid coordinates, and cluster population was set to the median population size of aggregated households within the cluster across the 2012–13 MTAT rounds. Total population over all 115 clusters was around 52,600 individuals, corresponding to the average number of individuals surveilled in each round rather than the total number of distinct individuals surveilled over all rounds. To account for population scaling, each cluster’s calibrated larval habitat parameters were proportionally adjusted to scale the magnitude of vector populations. A cluster of population p in the spatial simulation had larval habitat parameters adjusted to p1000sc*. Importations of malaria from outside the study area were not included unless specifically indicated.
Simulations modeled a period of 24 years beginning in 2006, and climate data after 2013 was inferred from the preceding years. The baseline simulation up to and including the 2012–13 MTAT activities was extended to include the 2014–15 MDA interventions in the Lake Kariba region. Simulated MDA campaigns with dihydroartemisinin-piperaquine (DP) were administered to all clusters in two campaigns beginning in December 2014 and July 2015. Each MDA campaign consisted of two rounds separated by 60 days. During MDA, individuals received treatment regardless of their RDT result. The coverage of the MDA campaigns in each cluster was set to the coverage level of the 2012–13 MTAT campaign for that cluster. DP was administered with age-dependent dosing and full compliance with all treatment courses was assumed unless specifically indicated.
Potential post-2015 intervention mixes were simulated to evaluate their ability to reach and maintain malaria elimination in the Lake Kariba region by 2030 (Figs 6–10), in line with southern Africa regional elimination goals. Combinations of ramp-ups in case management, increases in ITN usage, extending drug campaigns for an additional five years, targeting MDAs at high-burden HFCAs, coverage achieved by drug campaigns, impact of human migration rates, and impact of importation into the region were simulated and compared.
Case management ramp-up (scenarios 2–13) was modeled as gradual increase of both the percentage of people with access to malaria treatment and the rate at which symptomatic people seek treatment given treatment is available. Ramp-ups in case management were modeled to begin in January 2012 and plateau in January 2019 (S4C Fig), by which point 93% of malaria clinical cases in children under 5 received treatment, 89% of malaria clinical cases in people over age 5 received treatment and 98% of severe malaria cases across the entire population received treatment.
Three ITN coverage ramp-up options over the period 2014–22 were considered (S3B Fig and S11 Fig). Under “maintain current” (scenarios 1, 2, and 5), ITN coverage is maintained at 2015 levels via new ITNs distributed to individuals at birth; under “ramp-up” (scenarios 3, 6, 8, 10), ITN coverage is gradually increased between 2016 and 2022, when all distributions cease, extrapolating the historical and present ITN distribution coverage trajectory for each cluster post-2013; and under “aggressive” (scenarios 4, 7, 9, 11–13), ITN distributions covering 80% of the population administered every other year between 2016 and 2022.
For simulations where MDAs were extended through 2020 (scenarios 5–10, 12–13), five annual MDA campaigns with DP were administered beginning each July for the five years from 2016–20. Each campaign consisted of two rounds separated by 60 days. Campaign coverage at each cluster was set at the cluster’s 2012–13 MTAT coverage and did not vary from year to year.
In addition to MDAs over the entire Lake Kariba region, targeted MDA campaigns were simulated (scenarios 8, 9, 11). The twelve HFCAs were divided into high- and low-burden groups where high-burden HFCAs were those with RDT prevalence above 10% in children during surveillance in April 2014. In simulations of targeted MDAs, clusters in high-burden HFCAs received five additional annual MDAs with DP beginning in July 2016 as described above, while clusters in low-burden HFCAs did not receive MDA after the 2014–15 rounds.
All the intervention mix scenarios were simulated in the context of human migration across clusters. Scenarios were simulated at two migration settings. For scenarios 1–11 and 13, low migration, 7,440 round trips and 220 permanent relocations occurred each year. For scenario 12, high migration, 74,400 round trips and 2,200 permanent relocations occurred each year. In all scenarios, migration was spread evenly throughout the year and was independent of age.
Other than scenario 13, no intervention scenarios included importation of infections from outside the study area. For scenario 13, a total of 100 infections were imported annually into the entire study area. Three highly connected, high-population clusters, one from each of Gwembe, Munyumbwe, and Sinamalima HFCAs, were selected to be the loci of importation.
|
10.1371/journal.pgen.1005713 | Hnrnph1 Is A Quantitative Trait Gene for Methamphetamine Sensitivity | Psychostimulant addiction is a heritable substance use disorder; however its genetic basis is almost entirely unknown. Quantitative trait locus (QTL) mapping in mice offers a complementary approach to human genome-wide association studies and can facilitate environment control, statistical power, novel gene discovery, and neurobiological mechanisms. We used interval-specific congenic mouse lines carrying various segments of chromosome 11 from the DBA/2J strain on an isogenic C57BL/6J background to positionally clone a 206 kb QTL (50,185,512–50,391,845 bp) that was causally associated with a reduction in the locomotor stimulant response to methamphetamine (2 mg/kg, i.p.; DBA/2J < C57BL/6J)—a non-contingent, drug-induced behavior that is associated with stimulation of the dopaminergic reward circuitry. This chromosomal region contained only two protein coding genes—heterogeneous nuclear ribonucleoprotein, H1 (Hnrnph1) and RUN and FYVE domain-containing 1 (Rufy1). Transcriptome analysis via mRNA sequencing in the striatum implicated a neurobiological mechanism involving a reduction in mesolimbic innervation and striatal neurotransmission. For instance, Nr4a2 (nuclear receptor subfamily 4, group A, member 2), a transcription factor crucial for midbrain dopaminergic neuron development, exhibited a 2.1-fold decrease in expression (DBA/2J < C57BL/6J; p 4.2 x 10−15). Transcription activator-like effector nucleases (TALENs)-mediated introduction of frameshift deletions in the first coding exon of Hnrnph1, but not Rufy1, recapitulated the reduced methamphetamine behavioral response, thus identifying Hnrnph1 as a quantitative trait gene for methamphetamine sensitivity. These results define a novel contribution of Hnrnph1 to neurobehavioral dysfunction associated with dopaminergic neurotransmission. These findings could have implications for understanding the genetic basis of methamphetamine addiction in humans and the development of novel therapeutics for prevention and treatment of substance abuse and possibly other psychiatric disorders.
| Both genetic and environmental factors can powerfully modulate susceptibility to substance use disorders. Quantitative trait locus (QTL) mapping is an unbiased discovery-based approach that is used to identify novel genetic factors and provide new mechanistic insight into phenotypic variation associated with disease. In this study, we focused on the genetic basis of variation in sensitivity to the acute locomotor stimulant response to methamphetamine which is a behavioral phenotype in rodents that is associated with stimulated dopamine release and activation of the brain reward circuitry involved in addiction. Using brute force monitoring of recombination events associated with changes in behavior, we fortuitously narrowed the genotype-phenotype association down to just two genes that we subsequently targeted using a contemporary genome editing approach. The gene that we validated–Hnrnph1 –is an RNA binding protein that did not have any previously known function in psychostimulant behavior or psychostimulant addiction. Our behavioral data combined with our gene expression results provide a compelling rationale for a new line of investigation regarding Hnrnph1 and its role in neural development and plasticity associated with the addictions and perhaps other dopamine-dependent psychiatric disorders.
| Substance use disorders (SUDs) involving psychostimulants such as cocaine and methamphetamine (MA) are heritable; however, their major genetic determinants remain poorly defined [1–4]. In particular, genome-wide association studies (GWAS) of psychostimulant abuse have yet to discover the underlying genetic factors or causal sequence variants. SUDs involve multiple discrete steps including initial use, escalation, withdrawal, and relapse, each of which is believed to have a distinct genetic architecture. Therefore, we and others have used model organisms to explore the genetic basis of intermediate phenotypes, including initial drug sensitivity [5]. Model systems have great potential for studying addiction-relevant intermediate phenotypes [6] because they provide exquisite control over environmental conditions, including exposure to psychostimulants.
Psychostimulants activate the mesocorticolimbic reward circuitry in humans [7] and stimulate locomotor activity in mice [8]. The primary molecular targets of psychostimulants are the membrane-spanning monoaminergic transporters. Amphetamines act as substrates and cause reverse transport and synaptic efflux of dopamine, norepinephrine, and serotonin [9–11]. Sensitivity to the locomotor stimulant response to MA is heritable and may share a genetic basis with the addictive, neurotoxic, and therapeutic properties of amphetamines [8, 12–15]. More broadly, determining the genetic basis of sensitivity to amphetamines may provide insight into the neurobiology of other conditions involving perturbations in dopaminergic signaling, including attention deficit hyperactive disorder (ADHD), schizophrenia, and Parkinson’s disease [16]. This hypothesis is supported by our recent identification of a genetic correlation between alleles that increased amphetamine-induced euphoria and alleles that decreased risk of schizophrenia and ADHD [17].
We and others have reported several quantitative trait loci (QTLs) in mice that influence MA sensitivity [12, 18–24]. A distinct advantage of QTL analysis is that chromosomal regions can eventually be mapped to their causal polymorphisms. However, obtaining gene-level and nucleotide-level resolution can be extremely challenging when beginning with a lowly recombinant population such as an F2 cross. A classical approach is to fine map QTLs derived from an F2 cross using successively smaller congenic strains. Whereas this approach is efficient for Mendelian alleles, there are only a few examples in which this approach has been successful in identifying alleles for more complex, polygenic traits, such as histocompatibility [25], substance abuse [26] and depressive-like behavior [27].
In the present study, we fine mapped a QTL on chromosome 11 that modulates methamphetamine sensitivity and that segregates between C57BL/6J (B6) and DBA/2J (D2) inbred strains [12, 20]. We used interval-specific congenic lines in which successively smaller D2-derived segments were introgressed onto a B6 background [28]. We also conducted transcriptome analysis of brain tissue from a congenic line that captured the QTL for reduced MA sensitivity. Our transcriptome analysis focused on the striatum, which is a brain region important for psychostimulant-induced locomotor activity and reward [29]. We used GeneNetwork [30] and in silico expression QTL (eQTL) analysis of several brain regions to identify cis- and trans-eQTLs that may explain changes in the transcriptome caused by this QTL. Finally, to identify the quantitative trait gene responsible for reduced MA sensitivity, we used transcription activator-like effector nucleases (TALENs) to introduce frameshift deletions in the first coding exon of each positional candidate gene [31].
Several genome-wide significant QTLs that influenced MA sensitivity were previously reported in this B6 x D2-F2 cross, including QTLs on chromosomes 1, 8, 9, 11, 15, and 16 [20]. Here, we further dissected the chromosome 11 QTL (peak = 50 Mb; D2 < B6) into 5 min bins and identified a peak LOD score at 25 min post-MA administration (Fig 1). We then produced interval-specific congenic lines to fine map this QTL. The genomic intervals (Mb) for the congenic lines and the peak F2-derived QTL are illustrated in Fig 2A and the SNP markers that defined the congenic intervals for Lines 1–6 are listed in S1 Table. As shown in Fig 2B–2E, some of the congenic lines captured a QTL that reduced MA sensitivity whereas others did not (see also S2A and S2B Fig and S1 Text). Whether or not a strain captured a QTL is indicated by a + or–sign in Fig 2A.
Congenic Line 4 was the smallest congenic that captured a QTL for reduced MA sensitivity. Therefore, we produced subcongenic lines from Line 4, as shown in Fig 3A. The SNP markers that defined the congenic intervals for Lines 4a-4h are listed in S2 Table. Production and analysis of these congenic lines was more efficient because the D2-derived allele was dominant. Therefore all lines shown in Fig 3 were heterozygous for the D2-derived congenic interval. Once again, some but not all of the congenic lines captured the QTL inherited from Line 4 (Figs 3B, 3C, 3D and S3, S3 Table and S1 Text). Based on the observation that Line 4b but not 4c captured the QTL, we were able to define a 206 kb critical interval (Fig 3E). The first proximal SNP in Lines 4b was rs29424921 and first proximal SNP in Line 4c was rs29442500. The physical location of these SNPs defined the boundaries of the critical interval (50,185,512–50,391,845 bp; S2 Table). This interval contains only two protein coding genes: Hnrnph1 (heterogeneous nuclear ribonucleoprotein) and Rufy1 (RUN and FYVE domain containing 1; Fig 3E and S4 Table).
Using Line 4c to define the distal boundary presumes that our analysis of Line 4c was powerful enough to detect the QTL if it were present. We used data generated from Line 4b to estimate the QTL effect size; based on this estimate, a sample size of N = 25 per group would be required to achieve 80% power to detect this QTL in Line 4c. We phenotyped an even larger number of mice from Line 4c (N = 30–40 per genotype), but did not detect the QTL (Fig 3D). Therefore, we can confidently interpret the negative results from Line 4c. Further negative results obtained from five additional subcongenic lines also support the critical interval as defined in Fig 3E (see S3 Fig and S3 Table).
Studies of congenic lines can be confounded by residual heterozygosity that lies outside of the congenic region. In order to address this concern, we genotyped individuals from Line 4 subcongenics at 882 SNPs using a SNP genotyping microarray. Although we did identify a single D2-derived SNP on chromosome 3, it was observed both in wild-type and heterozygous congenic mice and was not associated with the locomotor response to MA (see S4 Fig and S1 Text). Based on these results we rejected the possibility that the differences in the congenic lines were due to residual heterozygosity.
In an effort to understand the molecular impact of this QTL, we used RNA-seq to identify gene expression differences in the striatum of naïve Line 4a congenics versus their naïve B6 littermates. We identified between 91 differentially expressed genes with an FDR of 5% and 174 differentially expressed genes with and FDR of 20%. The majority of these genes were downregulated in Line 4a (S6 Table). Notably, Nr4a2 (Nurr1) was the most significant, demonstrating a 2.1-fold decrease in expression (p = 4.2 x 10−15; Fig 4). Decreased Nurr1 expression in Line 4a was confirmed using qPCR (S5A Fig and S7 Table).
We used the Ingenuity Pathway Analysis (IPA; Ingenuity Systems, Redwood City, CA, USA; www.qiagen.com/ingenuity) software in conjunction with the genes we identified with an FDR of 5% to explore pathways that were enriched for these genes. The top three canonical pathways that we identified included the neuronal functions Glutamate Receptor Signaling, Gαq Signaling, and G-Protein Coupled Receptor Signaling (S8 Table). Neither transcriptome nor qPCR analysis detected any significant difference in gene- or exon-level expression of Hnrnph1 or Rufy1 (S5B, S5C and S6 Figs). The most strongly implicated IPA network was, “Cellular Development, Nervous System Development and Function, Behavior”. This network consists of several downregulated genes involved in neural development, maintenance, and signaling (Fig 4), including Bdnf, which was downregulated and connected to several downregulated genes involved in synaptic transmission, including Malat1, the vesicular glutamate transporters VGLUT1 (Slc17a7) and VGLUT2 (Slc17a6), as well as the AMPA-4 receptor subunit (Gria4), alpha-1d adrenergic receptor (Adra1d), and calcium-dependent secretion activator 2 (Cadps2). The top “Diseases and Functions” annotations included Huntington’s disease, nervous system coordination, and disorder of basal ganglia (S9 Table), further supporting dysfunction in striatal innervation and signaling. Htt (huntingtin) was the top predicted upstream transcriptional regulator followed by Creb1 (cyclic AMP response element binding protein) which together accounted for 23 (25%) of the 91 differentially expressed genes (S7 Fig).
Gene Ontology (GO) pathways identified via WebGestalt [32, 33] complemented the IPA results and generally indicate neuronal dysfunction. The top biological process was synaptic transmission and signaling processes, the top molecular functions involved membrane proteins including transporters and g protein-coupled receptors and the top cellular components were associated with neuronal synapses (Table 1).
In order to identify genetic polymorphisms associated with changes in gene expression observed in the congenic region of Line 4a, we used GeneNetwork [30] to identify both cis- and trans-eQTLs that originated from B6/D2 polymorphisms within the Line 4a congenic region (FDR < 20%; S6 Table). We identified several trans-QTLs caused by SNPs within the Line 4a region, including a link between genetic variation in Hnrnph1 and differential expression of Ipcef1 (Tables 2 and S6) [30], a gene that lies within Oprm1 (mu opioid receptor) and is transcribed in the reverse direction. These observations support the gene expression differences we observed using RNA-seq and indicate that our QTL regulates the expression of numerous other genes outside of the QTL interval.
One of the major advantages of genetic analysis in model organisms is the ability to perform experimental manipulations to evaluate observed correlations between genotype and phenotype. We used TALENs to introduce frameshift deletions that resulted in premature stop codons into the first coding exon of each of the two protein coding genes within the 206 kb critical interval–Hnrnph1 and Rufy1. We identified two founders that were heterozygous for 11 bp and 16 bp frameshift deletions in the first coding exon of Hnrnph1 (Hnrnph1 +/-; Founders #28 and #22; Figs 5A and S8). We did not observe any off-target deletions in the highly homologous Hnrnph2 gene nor did we observe compensatory change in striatal Hnrnph2 expression (S9 Fig).
Hnrnph1 +/- mice showed reduced expression of Hnrnph1. When we used qPCR primers that hybridized to DNA sequences that were contained in both wild-type (Hnrnph1 +/+) and Hnrnph1 +/- mice, there was a significant upregulation of total Hnrnph1 transcript levels in Hnrnph1 +/- versus Hnrnph1 +/+ mice (Fig 5C and 5D). However, we also used qPCR primers that overlapped the deleted interval and in this case we observed a significant downregulation of Hnrnph1 +/+ transcript levels in Hnrnph1 +/- mice (Fig 5E). These observations provide functional evidence that the Hnrnph1 frameshift deletion disrupted gene transcription. Similar to Lines 4, 4a and 4b, Hnrnph1 +/- mice from Line #28 and Line #22 that were derived from Founders #28 and #22 both exhibited reduced MA sensitivity (Fig 5F and 5G), thus recapitulating the congenic QTL phenotype. Reduced MA sensitivity was also observed using 30 min behavioral sessions (S10 Fig).
In contrast to Hnrnph1 +/- mice, Rufy1 +/- mice carrying a frameshift deletion (S8 Fig) did not exhibit any difference in behavior (Fig 6). To further support the likelihood of reduced neurobehavioral function in Hnrnph1 +/- mice, Hnrnph1 expression is also clearly higher than Rufy1 in the adult brain (S6 Fig; S11 Fig) [34].
To summarize, we observed a significant reduction in MA sensitivity in Hnrnph1 +/- mice, but not Rufy1 +/- mice that recapitulated the congenic QTL phenotype, thus identifying Hnrnph1 as a quantitative trait gene for MA sensitivity.
We used positional cloning and gene targeting to identify Hnrnph1 as a novel quantitative trait gene for MA sensitivity. First, we identified a broad, time-dependent QTL on chromosome 11 using an F2 cross between two inbred strains (Fig 1). We then narrowed a QTL from the initial 40 Mb interval to approximately 10 Mb using interval-specific congenic lines (Figs 2, 3, S2 and S3). Further backcrossing yielded a fortuitous recombination event that narrowed a critical interval to just 206 Kb; this region contained only two protein coding genes: Hnrnph1 and Rufy1 (Fig 3E). Striatal transcriptome analysis identified potential neurobiological mechanisms, including a predicted deficit in midbrain dopaminergic neuron development and neurotransmission. The use of GeneNetwork [30] to identify eQTLs associated with our transcriptomic findings provided mechanistic insight, including a trans-QTL that maps to Hnrnph1 that could cause differential expression of Ipcef1 (Table 2; S6 Table). Finally, we took advantage of the power of mouse genetics to create mice heterozygous for a frameshift deletion in either Hnrnph1 or Rufy1. Hnrnph1 +/- mice but not Rufy1 +/- mice recapitulated the congenic QTL phenotype, providing direct evidence that Hnrnph1 is a quantitative trait gene for MA sensitivity (Figs 5 and 6).
QTL mapping studies of rodent behavior have rarely provided strong evidence for causal quantitative trait genes [26, 27, 35]. We began pursuing this QTL more than a decade ago, when the difficulty of such projects was widely underestimated. A key limitation of our initial mapping strategy was the use of an F2 cross, in which extensive linkage disequilibrium created large haplotype blocks, resulting in the identification of very broad QTLs. Combining low resolution and high resolution QTL mapping in congenic lines revealed a more complex genetic architecture, indicating that Hnrnph1 is not the only causal gene within the F2 interval that underlies the QTL. Inheritance of two copies of the D2 segment enhanced the heterozygous phenotype in Line 1, yet had no further effect once the size of the segment was reduced following the creation of Line 4 (Fig 2B and 2E). We interpret this observation to suggest that Line 1 contains an additional, recessive QTL within the 35–50 Mb region of Line 3 that could summate with the Line 4 QTL to produce the larger effect size. This 35–50 Mb region could be fine-mapped to the causal genetic factor by introducing additional recombination events into Line 3. This detailed level of insight into the genetic architecture of a single large-effect QTL could only be made possible by employing a sufficiently powered phenotypic analysis of interval-specific congenic lines. Thus, a key to our success in identifying a single gene was the fact that while the QTL originally identified in the F2 cross was likely the product of multiple smaller QTLs, we were able to capture one major QTL in Line 4 and in subcongenic lines which appears to correspond to a single quantitative trait gene that we have now identified as Hnrnph1.
Transcriptome analysis of Line 4a supports a neurodevelopmental mechanism by which the QTL regulates MA sensitivity. Nr4a2 (a.k.a. Nurr1) was the top downregulated gene and codes for a transcription factor that is crucial for midbrain dopaminergic neuron development, survival, and cellular maintenance of the synthesis, packaging, transport, and reuptake of dopamine [36]. Nurr1 was a core component of a top-ranked gene network composed of primarily downregulated genes important for neurogenesis, neural differentiation, and synaptogenesis (Nr4a2 / Nurr1, Bdnf, Tbr1, Neurod6, Ets2, Malat1, Elavl2; Fig 4). Accordingly, there was a downregulation of striatal signaling pathways, including glutamate (Slc17a7, Slc17a6, Gng2, and Gria4), Gαq (Gng2, Chrm1, Adra1b, Adra1d), and GPCR signaling (Pde1b, Rgs14, Chrm1, Adra1b, Adra1d) (S8 Table). With regard to Gαq signaling, MA acts as a substrate for NET, causing efflux of NE [9] which then binds to α-adrenergic receptors that are coded by Adra1b and Adra1d. Notably, knockout mice for either of these receptors exhibit reduced amphetamine-induced locomotor activity [37, 38].
Some of the differentially expressed genes in Line 4a were previously associated with variation in amphetamine reward and reinforcement, including Nr4a2 (Nurr1), Adora2a, and Slc17a7 (Vglut1) [39]. Furthermore, the top predicted upstream regulator—Htt (huntingtin; S7A Fig) is a master regulator of a network of genes in the extended amygdala associated with protracted abstinence from chronic exposure to opioids, cannabinoids, nicotine, and alcohol [40].
Inheritance of the Hnrnph1 locus caused downregulation of a smaller reverse-transcribed gene located within the middle of Oprm1 (mu opioid receptor) called Ipcef1 (p = 0.001; FDR = 12%; S6 Table). We also identified a trans-eQTL in Hnrnph1 that regulates Ipcef1 expression (Table 2 [30]). Hnrnph1 was previously shown to regulate the expression Oprm1 (mu opioid receptor gene) via 5’ UTR-mediated repression [41] and splicing [42]. Furthermore, the human intronic SNP rs9479757 in OPRM1 was associated with heroin addiction severity and decreased binding affinity of HNRNPH1, resulting in exon 2 skipping [43]. Thus, Hnrnph1 regulation of Ipcef1 expression could represent an additional mechanism of Oprm1 regulation [44].
The QTL that contains Hnrnph1 is predicted to perturb the neural development of the mesocorticolimbic circuitry that mediates MA behavior. Hnrnph1 (heterogeneous nuclear ribonucleoprotein) codes for an RNA binding protein (RBP) that is highly expressed throughout the brain, including the striatum, cortex, and hippocampus (S11 Fig) [34] and binds to G-rich elements to either enhance or silence splicing [45, 46]. hnRNPs such as Hnrnph1 form hnRNP-RNA complexes to coordinate splicing of thousands of genes [46]. In addition, HNRNPH1 regulates 3’ UTR cleavage and polyadenylation [47] and several hnRNPs export mRNAs to neuronal processes to regulate spatiotemporal translation and post-translational modifications [48]. Synaptic activity can increase protein abundance of hnRNPs at the post-synaptic density of primary neurons [49]. The hippocampus contains focal expression of over 15 hnRNPs, including H1 (S11 Fig [34]). Importantly, Hnrnph1 contains a glycine rich domain that permits nucleocytoplasmic shuttling via transportin 1 [50] and exhibits activity-dependent translocation to the cytoplasm [51]. Several hnRNPs exhibit activity-dependent localization at the synapse [49], suggesting additional neuronal functions of Hnrnph1 in addition to splicing.
We identified Hnrnph1 as a quantitative trait gene responsible for MA sensitivity. However, the quantitative trait nucleotide(s) remain obscure. Hnrnph1 contains 18 genetic variants within the gene, including 15 intronic SNPs, a SNP in the 5’ UTR, a synonymous coding SNP, and a single T insertion in the 3’ UTR (S4 Table [52, 53]) that could cause brain region-specific differential expression of Hnrnph1 and/or its ability to regulate splicing of its transcriptome-wide targets [46, 47]. We did not observe differential striatal expression of Hnrnph1 at the gene level or the exon level as a consequence of inheriting the Line 4a QTL (S5 and S6 Figs). Our focus was limited to the striatum which is a behaviorally relevant region [16, 29] that exhibits high Hnrnph1 expression during early adulthood (S11 Fig). Therefore, the QTL could influence Hnrnph1 expression at a different time period, in a different, behaviorally relevant brain region, or in a specific subpopulation of cells. Interestingly, striatal microarray datasets in BXD strains indicate an increase in Hnrnph1 expression from postnatal day 3 to postnatal day 14 as well as a change in the strain rank order of expression [30] which suggests that genotypic differences in Hnrnph1 expression could depend on the developmental time point. Finally, because excised introns can trans-regulate gene expression, an alternative explanation is that excised, SNP-containing introns from Hnrnph1 can function as polymorphic long noncoding RNAs to perturb their trans-regulation of the transcriptome [54].
To our knowledge, there are no GWAS studies reporting genome-wide significant associations of HNRNPH1 variants with complex diseases or traits (http://www.ebi.ac.uk/gwas/). Interestingly, HNRNPH1 binding affinity and splicing can be modulated by genome-wide significant SNPs associated with bipolar disorder, major depressive disorder, and schizophrenia, including rs1006737 (CACNA1C), rs2251219 (PBRM1), and rs1076560 (DRD2) [55]. Thus, HNRNPH1 splicing could profoundly impact the neurobiological mechanisms underlying these disorders. Additionally, HNRNPH1 and RBFOX1/2 coordinate splicing [56, 57] and knockdown RBFOX1 (an autism-associated RBP involved in neural development [58]) in human neural progenitor cells revealed over 200 alternatively spliced genes containing HNRNPH1 binding sites [56] and 524 genes containing binding sites for ELAVL2, a neurodevelopmental RBP [59] that was downregulated in Line 4a (Fig 4).
In summary, we identified Hnrnph1 as a quantitative trait gene for MA sensitivity. This is rarely accomplished in rodent forward genetic studies of behavior and will likely advance our understanding of the neurobiological basis of multiple neuropsychiatric disorders involving monoaminergic dysregulation. Identifying brain region- and cell type-specific splicing targets of Hnrnph1 could reveal therapeutic targets for these disorders, many of which have been associated with specific gene splicing events [55]. Furthermore, pharmacological perturbation of RBP function could one day serve as an effective therapeutic strategy. Recent findings in models of neurodegenerative disease show that targeting RBP signaling could be a promising treatment approach [60].
All procedures in mice were approved by the Boston University and the University of Chicago Institutional Animal Care and Use Committees and were conducted in strict accordance with National Institute of Health guidelines for the care and use of laboratory animals. Colony rooms were maintained on a 12:12 h light–dark cycle (lights on at 0600 h). Mice were housed in same-sex groups of two to five mice per cage with standard laboratory chow and water available ad libitum. Age-matched mice were 50–100 days old at the time of testing (0900–1600 h).
For Lines 1–6 and Lines 4a-4h, locomotor activity was assessed in the open field [19]. Briefly, congenics, subcongenics, and wild-type littermates were transported from the vivarium to the adjacent behavioral testing room where they habituated for at least 30 min prior to testing. Mice were then placed into clean holding cages with fresh bedding for approximately five min before receiving an injection of saline on Days 1 and 2 (10 μl/g, i.p) and an injection of methamphetamine on Day 3 (MA; 2 mg/kg, i.p.; Sigma-Aldrich, St. Louis, MO USA). Mice were placed into the center of the open field (37.5 cm x 37.5 cm x 35.7 cm; AccuScan Instruments, Columbus, OH USA) surrounded by a sound attenuating chamber (MedAssociates, St. Albans, VT USA) and the total distance traveled was recorded in six, 5 min bins over 30 min using VersaMax software (AccuScan).
Mice heterozygous for a frameshift deletion in Hnrnph1 (Hnrnph1 +/-) or Rufy1 (Rufy1 +/-) were engineered (http://www.bumc.bu.edu/transgenic/), bred, and phenotyped at Boston University School of Medicine. Mice were bred and phenotyped in a manner similar to the congenics at the University of Chicago, with the exception that the open field was a smaller size (43.2 cm long x 21.6 cm wide x 43.2 cm tall; Lafayette Instruments, Lafayette, IN USA) and mice were recorded daily for 1 h rather than 30 min to allow a more robust detection of the phenotype. Reduced MA sensitivity was also replicated in Hnrnph1 +/- mice using the 30 min protocol (Supplementary Information). Behavior was videotaped using a security camera system (Swann Communications, Melbourne, Australia) and data were collected and analyzed using video tracking (Anymaze, Stoelting, Wood Dale, IL USA).
Because our primary focus was on MA-induced locomotor activity on Day 3, we first ran a two-way repeated measures ANOVA for Day 3 using genotype and sex as factors and time as the repeated measure. Because sex did not interact with genotype or time for any of the lines on Day 3, we combined sexes for the analysis of Days 1–3 and used repeated measures ANOVA with genotype as the main factor. Main effects of genotype and genotype x time interactions were deconstructed using one-way ANOVAs and Fisher’s post-hoc test of each time bin or t-tests in cases where there were two genotypes. A p-value of less than 0.05 was considered significant.
B6 x D2-F2 mice (N = 676) were generated, maintained, genotyped, and analyzed as previously described [20, 22]. Genome-wide QTL analysis was performed in F2 mice using the R package QTLRel that contains a mixed model to account for relatedness among individuals [61]. We recently validated the use of permutation when estimating significance thresholds for mixed models [62]. Sex was included as an interactive covariate. For each analysis, significance thresholds (p < 0.05) were estimated using 1000 permutations. The F2 data and R code for are publicly available on github (https:/github.com/wevanjohnson/hnrnph1).
Lines 1 and 6 were obtained from Dr. Aldons Lusis’s laboratory at UCLA (Lines “11P” and “11M” [28]) and had previously been backcrossed to B6 for more than 10 generations. These lines contained homozygous, introgressed regions from D2 on an isogenic B6 background that spanned chromosome 11. Because Lines 1 and 6 contained such large congenic intervals, we first phenotyped non-littermate offspring derived from homozygous congenic breeders versus homozygous B6 wild-type breeders (The Jackson Laboratory, Bar Harbor, ME; Figs 2 and S2) rather than heterozygous-heterozygous breeders to avoid the otherwise high likelihood of introducing unmonitored recombination events. Thus, we ensured that each individual possessed an identical genotype within each congenic line. The same type of control group is typically employed in the initial screen of chromosome substitution strains [19, 63, 64] which are essentially very large congenic lines. We crossed Line 1 to B6 and phenotyped the F1 offspring alongside age-matched B6 mice. B6 cohorts were combined into a single group for the combined analysis of all three genotypes for Line 1 (homozygous for B6, homozygous for D2, and heterozygous; Fig 2).
Next, we backcrossed Line 1 heterozygotes to B6 to generate subcongenic Lines 2–5 (Figs 2 and S2). Recombination events were monitored using genomic DNA extracted from tail biopsies and a series of TaqMan SNP markers (Life Technologies; Carlsbad, CA; S1 Table). We then used heterozygous-heterozygous breeding in Lines 2–5 to produce littermates of all three genotypes for simultaneous phenotyping (Figs 2 and S2). Because the QTL in Line 4 represented the smallest congenic region and was dominantly inherited, we backcrossed Line 4 heterozygotes to B6 to generate heterozygotes and wild-type littermates for Lines 4a-4h (Figs 3 and S3). We used additional TaqManSNP markers (Life Technologies) to monitor recombination events and defined the precise congenic boundaries using PCR and Sanger sequencing of SNPs chosen from the Mouse Sanger SNP query database (http://www.sanger.ac.uk/cgi-bin/modelorgs/mousegenomes/snps.pl [52]). Genomic coordinates are based on mm9 (Build 37).
We assayed tail SNP DNA from one heterozygous congenic mouse and one B6 wildtype littermate from Lines 4a-4d (eight mice total) using services provided by the DartMouseSpeed Congenic Core Facility at the Geisel School of Medicine at Dartmouth College (http://dartmouse.org/). A total of 882 informative B6/D2 SNPs were analyzed on the GoldenGate Genotyping Assay (Illumina, Inc., San Diego, CA) using DartMouse’s SNaP-Mapand Map-Synth software to determine the allele at each SNP location. After detecting a single off-target locus on chromosome 3 (rs13477019; 23.7 Mb), we used a custom designed TaqMan SNP marker for rs13477019 (Life Technologies, Carlsbad, CA USA) to confirm the result and to genotype additional samples from Lines 4a-4h for which we had both DNA and behavioral phenotypes. Data from this SNP marker were then used to test for the effect of genotype at the chromosome 3 locus on MA-induced locomotor activity.
We harvested and pooled bilateral 2.5 mm diameter punches of the striatum for each individual sample from naïve, congenic mice and B6 wildtype littermates from Line 4a (N = 3 females and 5 males per genotype; 50–70 days old). Total RNA was extracted as previously described [23] and purified using the RNeasy kit (Qiagen, Valencia, CA, USA). RNA was shipped to the University of Chicago Genomics Core Facility where cDNA libraries were prepared for 50 bp single-end reads according to the manufacturer’s instructions using the Illumina TruSeqStranded mRNA LT Kit (Part# RS-122-2101). Purified DNA was captured on an Illumina flow cell for cluster generation and sample libraries were sequenced at eight samples per lane over two lanes (technical replicates) on the Illumina HiSeq 2500 machine according to the manufacturer’s protocols. FASTQ files were quality checked via FASTQC and possessed Phred quality scores > 30 (i.e. less than 0.1% sequencing error). Using the FastX-Trimmer from the FastX-Toolkit, the 51st base was trimmed to enhance read quality and prevent misalignment. FASTQ files were utilized in TopHat [65] to align reads to the reference genome (UCSC Genome Browser). Read counts per gene were quantified using the HTSeq Python package and the R Bioconductor package edgeR was used to analyze differential gene expression. EdgeR models read counts using a negative binomial distribution to account for variability in the number of reads via generalized linear models [66]. “Home cage” was included as a covariate in the statistical model to account for cage effects on gene expression. The p-values obtained for differential expression were then adjusted by applying a false discovery rate (FDR) method to correct for multiple hypothesis testing [67]. The transcriptome dataset and code for RNA-seq analysis are available via NCBI Gene Expression Omnibus (http://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?token=cxkdoeaudvyhlqt&acc=GSE66366).
Oligo-dT primers were used to synthesize cDNA from total RNA to examine mRNA expression. Primer efficiencies for real-time quantitative PCR (qPCR) experiments were calculated using cycle threshold (CT) values (SYBR® Green; Life Technologies) derived from five, 10-fold serial cDNA dilutions; efficiencies (E) ranged from 90–100% (R2 = 0.99–1). Each sample was run in triplicate and averaged. Differential gene expression was reported as the fold-change in congenic or frameshift-deleted mice relative to B6 wild-type littermates using the 2-(∆∆CT) method [68].
We used our differentially expressed gene list from the striatal transcriptome that contained both the log2 fold-change and p-values (FDR < 5%) and applied IPA (www.qiagen.com/ingenuity) to identify enriched molecular pathways, functional annotations, gene networks, upstream causes, and predicted neurobiological consequences caused by inheritance of the QTL. IPA utilizes an algorithm that assumes that an increase in the number of molecular interactions indicates an increase in the likelihood of an effect on biological function. IPA uses a manually curated database (IPA Knowledge Base) containing the published literature to extract gene networks containing equally treated edges that directly and indirectly connect biologically related genes (www.qiagen.com/ingenuity). IPA analyses were conducted in February 2015.
To identify published cis- and trans- eQTLs that could explain gene expression differences caused by inheritance of the Line 4a congenic interval, we queried differentially expressed genes (FDR < 20%; 174 genes total; S6 Table) in transcriptome datasets from several brain regions in GeneNetwork [30] involving BXD recombinant inbred strains (recombinant inbred strains derived from B6 and D2 strains). We considered cis- and trans-QTLs originating from SNPs located within the 50–60 Mb locus and employed an arbitrary cut-off of LRS ≥ 13.8 (LOD ≥ 3). We only included genes where there was an exact match of gene with the LRS location using the appropriate genome build coordinates for each dataset.
TALENs vectors encoded either the right or left arm of the TALE effector that targeted the first coding exons of Hnrnph1 or Rufy1 (Cellectis Bioresearch, Paris, France). Upon bacterial cloning and purification, TALENs vectors containing a T7 promoter were linearized and used as templates for in vitro mRNA synthesis (mMessage mMachine T7 transcription kit; Life Technologies), and purified using MEGAclear transcription clean-up kit (Life Technologies). Each mRNA cocktail was diluted in sterile buffer and injected into B6 single-cell embryos at the BUMC Transgenic Core facility (http://www.bumc.bu.edu/transgenic/). We developed a genotyping assay utilizing native restriction enzyme recognition sites within the TALENs FokI cleavage domain. Genomic DNA was extracted from mouse tail biopsies and PCR-amplified with primers targeting100 base pairs upstream and downstream of the TALENs binding domain. Amplicons were then exposed to restriction digest overnight, run on a 2% agarose Ethidium Bromide Tris-Borate-EDTA gel, and imaged with ultraviolet light. TALENs-targeted deletions were identified by the presence of undigested bands caused by a loss of the restriction site. To confirm base pair deletions in our founder lines, undigested restriction enzyme-exposed PCR amplicon bands were excised, gel-purified, and vector-ligated overnight at 4°C using the pGEM T-easy Vector Systems (Promega). The ligation reaction was transformed into MAX Efficiency DH5α Competent Cells (Invitrogen) and plated onto Ampicillin-IPTG/X-Gal LB agarose plates for blue-white selection. Following overnight incubation at 37°C, white colonies were picked, cultured in ampicillin-enriched LB medium, and amplified. The PCR product was purified using the QIAprep Miniprep kit (QIAGEN). We then sequenced the vectors for the deletions using the pGEM T7 site upstream of the insert.
An Hnrnph1 forward primer (GTTTTCTCAGACGCGTTCCT) and reverse primer (ACTGACAACTCCCGCCTCA) were designed to target upstream and downstream of the TALENs binding domain in exon 4 of Hnrnph1. Genomic DNA was used to amplify a 204 bp PCR product using DreamTaq Green PCR Mastermix (ThermoScientific). PCR products were treated with the BstNI restriction enzyme (New England Biolabs) or a control enzyme-free buffer solution and incubated overnight at 60°C to ensure complete digestion. Enzyme-treated PCR products and untreated controls were resolved in 2% agarose gel electrophoresis with 0.5 μg/mL ethidium bromide to visualize under UV light. There were two BstNI restriction sites within the Hnrnph1 amplicon that were located proximal and distal to the TALENs FokI cleavage zone. Mice heterozygous for the Hnrnph1 deletion showed two bands on the gel, while B6 controls showed a single band.
Similar to Hnrnph1, a Rufy1 forward primer (AATCGTACTTTCCCGAATGC) and reverse primer (GGACTCTAGGCCTGCTTGG) targeted upstream and downstream of the TALENs binding domain in the first coding exon (exon 1). The 230 bp PCR amplicon contained a SacII restriction site that was deleted in Rufy1 +/- mice. Thus, Rufy1 +/+ mice showed a single, smaller digested band whereas Rufy1 +/- mice showed both the digested band as well as a larger, undigested band.
To assess off-target activity in Hnrnph1-targeted mice, we used the UCSC genome browser to BLAT the TALENs binding domains and identified a single homologous region located within the first coding exon of Hnrnph2. We used the same PCR- and gel-based assay to test for the deletion in Hnrnph2 with the exception that we used forward (GCCACCAAGAGTCCATCAGT) and reverse primers (AATGCTTCACCACTCGGTCT) that uniquely amplified a homologous 197 bp sequence within Hnrnph2 that contained a single Bstn1 restriction site. Digestion at the Bstn1 site produced an 81 bp band and a 115 bp band.
|
10.1371/journal.pgen.1004688 | Spinster Homolog 2 (Spns2) Deficiency Causes Early Onset Progressive Hearing Loss | Spinster homolog 2 (Spns2) acts as a Sphingosine-1-phosphate (S1P) transporter in zebrafish and mice, regulating heart development and lymphocyte trafficking respectively. S1P is a biologically active lysophospholipid with multiple roles in signalling. The mechanism of action of Spns2 is still elusive in mammals. Here, we report that Spns2-deficient mice rapidly lost auditory sensitivity and endocochlear potential (EP) from 2 to 3 weeks old. We found progressive degeneration of sensory hair cells in the organ of Corti, but the earliest defect was a decline in the EP, suggesting that dysfunction of the lateral wall was the primary lesion. In the lateral wall of adult mutants, we observed structural changes of marginal cell boundaries and of strial capillaries, and reduced expression of several key proteins involved in the generation of the EP (Kcnj10, Kcnq1, Gjb2 and Gjb6), but these changes were likely to be secondary. Permeability of the boundaries of the stria vascularis and of the strial capillaries appeared normal. We also found focal retinal degeneration and anomalies of retinal capillaries together with anterior eye defects in Spns2 mutant mice. Targeted inactivation of Spns2 in red blood cells, platelets, or lymphatic or vascular endothelial cells did not affect hearing, but targeted ablation of Spns2 in the cochlea using a Sox10-Cre allele produced a similar auditory phenotype to the original mutation, suggesting that local Spns2 expression is critical for hearing in mammals. These findings indicate that Spns2 is required for normal maintenance of the EP and hence for normal auditory function, and support a role for S1P signalling in hearing.
| Progressive hearing loss is common in the human population but we know very little about the molecular mechanisms involved. Mutant mice are useful for investigating these mechanisms and have revealed a wide range of different abnormalities that can all lead to the same outcome: deafness. We report here our findings of a new mouse line with a mutation in the Spns2 gene, affecting the release of a lipid called sphingosine-1-phosphate, which has an important role in several processes in the body. For the first time, we report that this molecular pathway is required for normal hearing through a role in generating a voltage difference that acts like a battery, allowing the sensory hair cells of the cochlea to detect sounds at extremely low levels. Without the normal function of the Spns2 gene and release of sphingosine-1-phosphate locally in the inner ear, the voltage in the cochlea declines, leading to rapid loss of sensitivity to sound and ultimately to complete deafness. The human version of this gene, SPNS2, may be involved in human deafness, and understanding the underlying mechanism presents an opportunity to develop potential treatments for this form of hearing loss.
| Spinster homolog 2 (Spns2) is a multi-pass membrane protein belonging to the Spns family. Though the functions of Spns1 and Spns3 are largely unknown, Spns2 is known to act as a sphingosine-1-phosphate (S1P) transporter, based upon previous studies in zebrafish and mouse [1]–[6]. S1P is a vital lipid. It has diverse roles, functioning as a signalling molecule regulating cell growth [7], [8], programmed cell death [9], angiogenesis [10], [11], vascular maturation [12], [13], heart development [14] and immunity [15], [16] by binding specific G-protein-coupled S1P receptors. Red blood cells and endothelial cells are important sources of circulating S1P [17]–[19]. The role of Spns2 in regulating S1P signalling is still elusive.
Spns2-deficient mice were initially discovered to be deaf during a large-scale screen of new mouse mutants carried out by the Sanger Institute's Mouse Genetics Project (MGP). The MGP uses the KOMP/EUCOMM resource of over 15,000 genes targeted in embryonic stem (ES) cells and aims to generate new mutants and screen them for a wide range of diseases and traits to reveal the function of 160 mutant genes each year [20]. Screening of hearing using the Auditory Brainstem Response (ABR) is part of the standardised battery of primary phenotypic tests and is carried out at 14 weeks of age [20]. Mutants generated from the KOMP/EUCOMM ES cell resource normally carry LoxP and Frt sites (Fig. 1A) engineered to facilitate further genetic manipulation to generate the conditional allele and then to knock out gene expression selectively [21].
Spns2-deficient mice showed early onset of hearing loss that progressed rapidly to profound deafness. This was associated with declining endocochlear potential (EP), which appeared to be the primary physiological defect. At later stages we observed degeneration of sensory hair cells and decreased expression of several key genes required for normal generation of the EP in the lateral wall of the cochlea, but these appeared to be secondary effects. By producing and analysing different conditional knockouts, we established that Spns2 expression was required locally in the inner ear rather than systemically. Our study suggests a vital role for Spns2 and S1P signalling in hearing.
The introduction of a cassette with an additional splice acceptor site is predicted to interrupt normal transcription of the Spns2 gene (Fig. 1A) and generate a truncated non-functional transcript encoding the first 146 out of 549 amino acids of the Spns2 protein [4]. The Spns2tm1a/tm1a mice were fertile and can survive to adulthood, but were born at sub-Mendelian ratios (15.9% homozygotes among 747 offspring of heterozygous intercrosses; χ2 test, p<0.001). Quantitative real-time PCR revealed that residual transcript of Spns2 in cochleae, eyes and livers of the homozygous mice was substantially reduced compared to that of the heterozygous and wildtype mice (Fig. 2A). In order to completely inactivate expression of the Spns2 gene, we produced Spns2tm1b/tm1b mice by crossing Spns2tm1a/tm1a with mice ubiquitously expressing Cre recombinase to delete exon 3 in the germline. Spns2tm1b/tm1b mice were also fertile and can survive to adulthood with a birth rate at sub-Mendelian ratios (16.9% homozygotes among 266 offspring of heterozygous intercrosses; χ2 test, p = 0.0044). In other aspects of their phenotype, Spns2tm1b/tm1b mice were broadly similar to Spns2tm1a/tm1a mice (see http://www.sanger.ac.uk/mouseportal/search?query=spns2 for a comparison of the two lines). Spns2tm1a/tm1a mice were the first to be available and were used for most experiments in this study, and may be more relevant to human disease because most disease-causing mutations reduce rather than eliminate gene activity.
We found that Spns2tm1a/tm1a mice had profound hearing impairment during screening at 14 weeks old by auditory brainstem response (ABR) measurement (Fig. 1D). Spns2tm1a/tm1a mice displayed no overt signs of abnormal behaviour throughout life up to 12 months old (n = 20 homozygotes, 29 control littermates) suggesting normal vestibular function. ABR measurements were recorded at younger ages to find out the time of onset of hearing loss. Hearing impairment in Spns2tm1a/tm1a mice can be detected as early as 2 weeks of age at high frequencies (Fig. 1B, E, G), and thresholds were raised further at a wide range of frequencies by 3 weeks old (Fig. 1C, F, G). The mutant thresholds recorded at P21 were significantly elevated compared with mutant thresholds at P14, indicating progressive hearing loss (Kruskall-Wallis One-Way Analysis of Variance on ranks, H = 102.857, 17 degrees of freedom, p<0.001). There was no difference in ABR thresholds between the Spns2+/tm1a and Spns2+/+ mice, indicating recessive inheritance. Thresholds continue to improve in control mice from 2 to 3 weeks old as hearing function matures [22].
Quantitative real-time PCR showed that Spns2 is expressed in the cochlea, both in the lateral wall and organ of Corti, and in eyes and liver in P4 wildtype mice (Fig. 2A). X-gal staining was used to indicate expression domains of Spns2, benefiting from the knockout first design of the allele which includes a LacZ reporter gene (Fig. 1A). At P10, X-gal labeling was detected in the spiral prominence area (Fig. 2C), hair cells (Fig. 2D), Reissner's membrane (Fig. 2B), vessels of inner ear (Fig. 2E) including the spiral modiolar vessels (Fig. 2F), proximal auditory nerve and bony shell of the cochlea (Fig. 2B), and the stria vascularis close to where the Reissner's membrane attaches. There was a similar labelling pattern at P14. The X-gal staining was also detected in the maculae and cristae in the vestibular system (Fig. S1A,B). On the basis of the expression pattern, our further investigation focused on two key components of the inner ear: the organ of Corti, where the pressure wave is transduced into action potentials, and the lateral wall, which maintains the ionic homeostasis of the cochlear endolymph.
The Spns2tm1a/tm1a mice showed a normal gross morphology of the middle ear and ossicles assessed by dissection and gross inspection, and the cleared inner ears also showed no malformation (Fig. S1C,D). We performed scanning electron microscopy (SEM) of P4, P21, P28 and P56 Spns2tm1a/tm1a mice and littermate controls. The hair cells of Spns2tm1a/tm1a mice appeared normal at P4 (Fig. S1E,F) and at P21 (Fig. 3A,B). There was scattered or patchy degeneration of stereocilia of outer hair cells in the homozygous cochleae at P28 (Fig. 3C,D). Hair cell degeneration became more apparent over time and by P56, only a few outer hair cells remained at the apex with most of them missing in other turns, and inner hair cells showed signs of degeneration such as fused stereocilia (Fig. 3E,F).
We also used transmission electron microscopy to examine the cochlear duct at P28, and observed degeneration of the basal turn organ of Corti and an apparently reduced density of dendrites in Rosenthal's canal in that turn (see later).
The increase in ABR thresholds preceded degeneration of the organ of Corti suggesting that these were secondary changes rather than the primary cause of the hearing impairment.
The stria vascularis is responsible for pumping K+ into the endolymph and generation of the endocochlear potential (EP) [23]. The EP starts to develop at around P6 in the mouse reaching adult values around P16 [24] and plays a key role in sound transduction because it provides approximately half of the electrochemical gradient that drives cations from the endolymph, a K+-rich extracellular fluid, into the sensory hair cells through mechanoelectrical transduction channels. The lateral wall of the cochlea is composed of the stria vascularis, spiral prominence and the spiral ligament. A defect in the function of any of these components could interfere with the generation of the EP. Therefore, we measured the EP to evaluate the function of the lateral wall. The EP values of the control mice were normal, around 99 to 120 mV, which matched their normal hearing. Spns2tm1a/tm1a mice had abnormally low EP values of 2 to 41 mV at both P21 and P28 when they were profoundly deaf (Fig. 4). However, at P14 the EP was higher, ranging from 52 to 107 mV, with some measurements within the normal range, corresponding to the partially preserved hearing at that age (Fig. 1C). The EP appeared to develop to near-normal levels and then declined very rapidly between P14 and P21. These data indicate that the cause of hearing loss in Spns2tm1a/tm1a mice is a failure to maintain the normal level of the EP after it develops, suggesting that the primary lesion is more likely to be in the lateral wall than the organ of Corti.
In order to understand what causes the reduction in the EP, we investigated the lateral wall, where the EP is produced. Generation of a voltage difference requires efficient separation of different compartments within the cochlear duct with adequate electrical resistance. Therefore we examined the morphology of cell boundaries between adjacent marginal cells and between basal cells in whole-mount samples of the stria vascularis. Filamentous actin was stained by phalloidin to label the cell boundaries at different ages. At P14, both wildtype and homozygous mice showed a distinctive regular hexagonal pattern of the boundaries of marginal cells (Fig. 5A,B). At P28, a subtle change was seen in Spns2tm1a/tm1a mice (Fig. 5C) and it became worse with age with a patchy pattern of different layouts of cell boundaries (Fig. 5D,E). However, the boundaries were always continuous and intact in Spns2tm1a/tm1a mice without any sign of breakdown (Fig. 5C–E). We quantified the marginal cell numbers by using the labelled cell boundaries at P28. The density of marginal cells in Spns2tm1a/tm1a mice was reduced compared with controls (t test, p<0.05), associated with irregular layout of marginal cell boundaries (Fig. 5J). Noticeably, this irregularity of marginal cells was not detected in mutants at P14, a stage when hearing had started to deteriorate, but appeared at later stages. The boundaries between basal cells of the stria vascularis did not show any obvious anomalies in mutants (Fig. S2A,B).
The morphology of strial capillaries was examined as well. Some Spns2tm1a/tm1a mice showed slight dilation (2 out of 5 mice) in patches along the length of the cochlear duct, and apparently increased branching (Fig. 5H) at P14. At P28, these changes were detected in all five tested mice and were more severe than at P14 although were still patchy (Fig. S2C,D). The number of branch points per unit area in homozygotes was significantly more than that in the control mice (Mann-Whitney Rank Sum Test, p<0.05 at both P14 and P28; Fig. 5I).
We analysed the structure of the lateral wall of the cochlea, including the stria vascularis and spiral ligament, using semithin sections and transmission electron microscopy in P28 mice. The position of Reissner's membrane was normal in semi-thin sections of P28 cochleae (Fig. 6A,B), with no evidence of hydrops or collapse. No systematic differences in the appearance of fibrocytes of the spiral ligament were observed (Fig. 6C,D). The inner boundaries of marginal cells of the stria, facing the endolymph, have a typical scallop-shaped surface in wildtype mice with the junctions between adjacent cells raised, but this feature was not seen in the Spns2tm1a/tm1a mice and the luminal surface appeared flat (Fig. 6E,F). Nuclei of marginal and basal cells appeared more rounded in Spns2tm1a/tm1a mice than in wildtypes (Fig. 6E,F). There was also a marked difference in the appearance of endothelial cells and pericytes [25], [26] of strial capillaries, with the nuclei of mutant cells appearing more darkly-stained (Fig. 6G,H). However, this abnormality appeared to be limited to capillaries of the stria vascularis only, and was not seen in the spiral ligament capillaries (Fig. 6I), suggesting a specific effect of Spns2 deficiency on the capillaries of the stria vascularis.
Intermediate cells of the stria vascularis are derived from melanocytes and tend to accumulate pigment during ageing or under stress such as in mice with Pendrin deficiency [27], [28]. The pigmented cells may derive from migratory melanocytes that adopt macrophage-like features during development [27], [29] or may derive from macrophage invasion [28]. At 4 weeks of age, we did not observe any obvious difference in strial pigmentation between Spns2 mutants and control littermates, but by 7 months old the mutant strias appeared more strongly pigmented than controls (Fig. S2E,F). This timecourse, after the onset of raised thresholds, suggests the accumulation of pigment is likely to be a secondary effect, not a cause of cochlear dysfunction.
In view of the abnormal morphology of marginal cell boundaries, we asked whether the diffusion barrier of stria vascularis, for example between adjacent marginal cells, was affected because normal morphology of boundaries at P14 does not necessarily mean normal function. We used biotin as a tracer injected into the endolymphatic and perilymphatic compartments of 6 week old mice to test the barrier permeability of the stria vascularis. There was no evidence of biotin entry into the stria vascularis of Spns2tm1a/tm1a or control mice indicating a normal diffusion barrier of stria vascularis (Fig. 7A,B). As we observed dilated strial capillaries with abnormal endothelial cells and pericytes, we tested their permeability by injecting BSA-FITC into the tail vein. There were no signs of leakage of the tracer to the tissues surrounding the strial capillaries in Spns2tm1a/tm1a mice suggesting that they have normal permeability to BSA-FITC (Fig. 7C,D).
To further investigate the reasons underlying the reduced EP and deafness in the Spns2 tm1a/tm1a mice, we analysed expression of some key proteins involved in normal EP formation and maintenance by immunofluorescence labelling, including Kir4.1 (Kcnj10), Kv7.1 (Kcnq1), Cx26 (Gjb2), Cx30 (Gjb6), Na+, K+-ATPase (Atp1a1), NKCC1 (Slc12a2) and ZO-1 (Tjp1). In homozygotes aged P14 (Fig. 8F,G,J,K,N,O), the expression of these proteins appeared normal, apart from the expression of Kcnj10, which appeared normal in most mutants (5/8), with the remaining three mice showing reduced labelling in the basal turn only (Fig. 8A–C). At 5–6 weeks old, we observed similar expression of Na+/K+-ATPase, NKCC1 and ZO-1 in Spns2 tm1a/tm1a mice compared with control mice (Fig. S3A–F). However, the other four proteins showed reduced expression at this age and the expression of Kcnj10 was largely absent in the stria vascularis (Fig. 8D,E). In contrast, there was similar labelling intensity of Kcnj10 in the satellite cells of the spiral ganglion in mutants and control mice (Fig. S3G,H). The expression of Kcnq1 appeared to be evenly distributed on the luminal surface of marginal cells in both wildtype and homozygotes at P14 (Fig. 8F, G), but at 5–6 weeks, labelling of Kcnq1 in the homozygotes was absent in some of the marginal cells that had enlarged boundaries (Fig. 8H,I). There was extensive expression of Gjb2 and Gjb6 in fibrocytes type I, II, and V of the spiral ligament in the wildtype (Fig. 8J,L,N,P) and homozygotes (Fig. 8K,O), but expression appeared greatly reduced behind the spiral prominence area corresponding to the fibrocyte type II region [30] in the Spns2tm1a/tm1a mice (Fig. 8M,Q) aged 5–6 weeks old. As EP started to reduce at P14, while expression of Kcnj10 (in the majority of observed mice), Kcnq1, Gjb2 and Gjb6 appeared normal at this age, these findings suggest that the reduction in expression of these key proteins in adults is secondary to a primary dysfunction of EP generation.
We generated the conditional allele of Spns2 (Spns2tm1c) by crossing the Spns2tm1a allele to a line expressing Flp recombinase to excise the inserted cassette (Fig. 1A). The Spns2tm1c/tm1c mice have the same Spns2 allele as the wildtype except that exon 3 is flanked by two loxP sites. Spns2tm1c/tm1c mice showed normal ABR thresholds and normal morphology of hair cells (Fig. 9A, C). These observations confirmed that the inner ear defects we found in the Spns2tm1a/tm1a mice were due to the insertion of the cassette and its disruption of Spns2 gene function.
We then asked whether the hearing defects of Spns2tm1a/tm1a mice are caused by localised deficiency of Spns2 in the inner ear or are mediated systemically. S1P is known to be released from several other tissues that could affect cochlear function, including various blood cell types and endothelial cells [1], [2]. We generated conditional knockout mice carrying the Spns2tm1d allele in specific tissues by crossing mice carrying the Spns2tm1c allele with mice carrying Cre recombinase under the control of five different promoters: Tie1, Pf4, Lyve1, EpoR and Sox10, driving expression of Cre recombinase in blood vessel endothelial cells, platelets, lymphatic endothelial cells, red blood cells, and the inner ear with surrounding neural crest-derived mesenchyme respectively. Sox10-Cre transgenic mice have been successfully used to express Cre recombinase in the developing inner ear previously [31]. Homozygous Spns2tm1d mutants carrying the Tie1, Pf4, Lyve1 and EpoR Cre alleles all had normal ABR thresholds in young adults (Fig. 10). In contrast, no ABR response was detected in the young adult Spns2tm1d/tm1d;Sox10-Cre mice (Fig. 10). Spns2tm1d/tm1d; Sox10-Cre mice showed a similar pattern of progression of raised thresholds between 2 and 3 weeks old as observed in Spns2tm1a/tm1a mice (Fig. 11). Spns2tm1d/tm1d; Sox10-Cre mice also showed similar inner ear pathological changes as found in Spns2tm1a/tm1a mice, such as degeneration of hair cells (Fig. 9B, D) and irregular arrangement of marginal cell boundaries. Therefore, we propose that Spns2 plays an important role in mammalian hearing through its localised function in the inner ear.
The Spns2tm1b/tm1b mutants, with deletion of exon 3 of Spns2, showed a more severe increase in thresholds than the Spns2tm1a/tm1a and Spns2tm1d/tm1d;Sox10-Cre mice from 2 weeks old, the earliest stage studied, with a lack of detectable response at most frequencies in most mice up to the maximum stimulus intensity used, 95 dB SPL (Fig. 11, middle row).
Spns2tm1a/tm1a mice had other defects detected by the MGP phenotyping pipeline such as low white blood cell count and increased bone mineral density [4], [20], [32]. However, the eye defects were of particular interest because of the relatively common association of retinal defects with deafness, as in Usher syndrome for example, and our finding of Spns2 expression in the eye (Fig. 2A). We assessed the retina for features corresponding to those found in the organ of Corti, and found focal degeneration of the retina (Fig. 12D,E). As focal degeneration can be associated with a Crb1rd8 mutant allele found in some C57BL/6 lines [33], we sequenced this gene and confirmed that the Spns2 retinal phenotype was independent of the rd8 mutation. As the Spns2tm1a/tm1a stria vascularis capillaries showed abnormal morphology, we examined the retinal vasculature in whole mount preparations. Retinal vein morphology also appeared abnormal with some veins appearing thinner in mutants than in controls as well as veins of irregular caliber (Fig. 12F–I). This retinal vascular phenotype was first evident by P10 when the retinal vasculature was still undergoing development, and persisted into adulthood. We therefore undertook branchpoint analysis to quantify any differences between Spns2tm1a/tm1a and Spns2+/tm1a mice. This showed no difference between genotypes at P10 (Fig. 12J). We also analysed the pericyte coverage of the retinal vessels, as a reduction in pericyte coverage is associated with increased vascular permeability. We found no significant difference in pericyte coverage in peripheral vessels, and a small but significantly reduced coverage in central vessels (Fig. S4A). The vitreous and optic nerve appeared normal in the mutants (Fig. 12D,E).
Other eye defects included open eyelids at birth (Fig. S4B) resulting in corneal opacity, vascularization and ulceration (Fig. 12A,B,C). These corneal defects made retinal assessment by ophthalmoscopy impossible in vivo. Histological examination showed corresponding gross morphological defects. Eyes were smaller, the cornea was thickened with vascularisation, the anterior chamber was collapsed, and the lens was small and cataractous (Fig. 12D,E). We observed the anterior eye defects in the Spns2tm1a/tm1a and Spns2tm1b/tm1b mice, but these defects were not seen in Spns2tm1d/tm1d;Sox10-Cre mice or any of the other four conditional lines.
Here, we report that Spns2-deficient (Spns2tm1a/tm1a) mice have profound hearing loss and propose an underlying mechanism: a rapid decline in EP paralleling loss of auditory sensitivity and preceding degeneration of hair cells, suggesting that the primary lesion is in the cochlear lateral wall, the site of EP production and maintenance [23], [34], [35]. Reduced EP has been associated with ion transport defects [36]–[39]; defects of tight junctions [40] or gap junctions [41]–[43]; absence of melanocytes [24]; microvascular disease [44], [45]; abnormal spiral ligament development [35]; sphingomyelin metabolic disturbance [46], or the lateral wall can simply be a target in systemic diseases [47]. This suggests complexity underlying the strial/metabolic category of hearing loss described in humans [48]. EP is essential for hair cell function [49]. Degeneration of hair cells secondary to reduced EP has been reported in other mouse mutants [50], [51] and normal EP seems to be important for survival of the hair cells. However, as Spns2 is expressed in hair cells as well as in the lateral wall, we cannot exclude the possibility that disruption of Spns2 function in the organ of Corti also contributed to raised ABR thresholds and hair cell degeneration. Analysis of mice with conditional knockout of Spns2 in hair cells and other cochlear cell types will be useful in dissecting the role of Spns2 further.
Spns2 acts as a transporter of S1P [1]–[6]. S1P may modulate vascular tone [52] and has been shown to regulate the inner ear spiral modiolar artery tone in vitro [53], [54]. S1P-induced vasoconstriction is thought to be important to protect strial capillary beds from high pressure [54]. We found that Spns2 was expressed in blood vessels of the inner ear including spiral modiolar vessels. Any reduction in local S1P level due to Spns2 dysfunction may weaken vasoconstriction and explain the dilation of strial capillaries in Spns2tm1a/tm1a mice. The relationship between capillary size and EP value is not unidirectional; both smaller and larger strial capillaries have been reported in different mouse mutants with low EP [55], [56]. S1P signalling also can affect vascular permeability [57]–[59]. However, we did not see increased permeability of strial capillaries using BSA as a tracer. Recently, Mendoza and colleagues found little difference in lung vascular permeability between Spns2-deficient and control animals [2], similar to our finding in strial capillaries of Spns2tm1a/tm1a mice. BSA is a medium molecular mass tracer (66.4 kDa), so tracers of different sizes and properties such as Evans blue and cadaverine [29] may be useful for further investigation of the strial capillary barrier.
We found decreased expression of several proteins critical for normal EP production at 5–6 weeks of age in Spns2-deficient mice. However, the expression of most of these proteins appeared normal at the time when the EP has already started to decline at P14, suggesting that these are likely to be secondary effects. The morphological changes of marginal cell boundaries and reduction in marginal cell density together with a lack of expression of Kcnq1 in affected cells are also likely to be secondary changes because these features were normal when hearing started to deteriorate at P14 and they did not affect strial permeability. Loss of Kcnq1 expression in marginal cells with expanded luminal surfaces may be a common consequence of strial dysfunction because it has been reported in several different mutants with reduced EP [28], [46]. It has been suggested that these common changes in mutant marginal cells may occur because these cells are relatively sensitive to metabolic stress [60], [61]. The variable decrease in Kcnj10 labelling in the basal turn and slightly dilated strial capillaries in a few mutants at P14 were the earliest abnormalities we have seen, and may correlate with the variable reduction in EP values at the same age. Disrupted expression of Kcnj10 is another common consequence of strial dysfunction [37], [47], [62]. The causal direction between reduced EP and decreased Kcnj10 expression in Spns2-deficient mice needs further investigation.
The most robust labeling for Spns2 expression was consistently in the hair cells and spiral prominence. The function of the spiral prominence is unclear. Two types of voltage-dependent K+ currents are expressed in spiral prominence epithelial cells, which may play a role in the homeostasis of inner ear fluids [63]. Another gene expressed strongly in the spiral prominence is pendrin (Pds), and the Pds mutant mouse also shows reduced EP [64], loss of expression of Kcnj10 [37] and Kcnq1 [28], and increased accumulation of strial pigmentation compared with controls [28]. However, the Pds mutant shows a severe early developmental defect of the inner ear with extensive hydrops [65], which we do not find in Spns2 mutants. This indicates that Spns2-deficient and Pds-deficient mice may have different mechanisms underlying the reduced EP and hearing impairment. Dysfunction of the spiral prominence in Spns2-deficient mice may be the main trigger of reduction of the EP and a series of pathological changes in inner ears.
One later change we saw in the lateral wall was a localised decrease of Gjb2 and Gjb6 expression in the type II fibrocytes of the spiral ligament located adjacent to the spiral prominence. Type II fibrocytes are important for K+ recycling and are considered to mediate K+ translocation between the epithelial cell network of the organ of Corti and the fibrocyte network of the lateral wall, and to facilitate ion flow directed towards the stria vascularis [66]. Intriguingly, this reduction in expression was seen for Gjb2 and Gjb6 only, and no reduction in labelling was found of Atp1a1 and Slc12a2, two other proteins strongly expressed in type II fibrocytes [67], [68]. This may indicate a selective impact of Spns2 deficiency on Gjb2 and Gjb6 expression in nearby cells, or alternatively these two genes may be more sensitive to changes in homeostasis in the cochlear duct than Atp1a1 and Slc12a2.
S1P is a bioactive lipid and acts as a second messenger intracellularly and as a ligand for cell surface G protein-coupled receptors extracellularly [69]. Five different S1P receptors participate in cellular responses based on the cell type and available downstream effectors [70]. S1P signalling has been implicated in maintenance of hair cells via activation of S1P receptor 2 (S1PR2) [71]. S1pr2-null mice are deaf and share some pathological changes with Spns2-deficient mice, such as disorganized cell boundaries of marginal cells, dilated capillaries in the stria vascularis, and degeneration of the organ of Corti [54], [71], [72]. Unlike S1pr2-null or S1pr2/S1pr3 double null mice, no overt vestibular defects were found in Spns2-deficient mice. Thus, we propose that the Spns2-S1P-S1PR2 signalling axis is important for normal hearing function. A similar Spns2-S1P-S1PR2 signalling axis may exist in bones as both S1pr2-deficient [73] and Spns2-deficient [32] mice have strong but brittle bones with high bone mineral density. In contrast, the Spns2-S1P-S1PR1 signalling axis is more important for lymphocyte trafficking [1]–[5].
Systemic disruption of Spns2 function in blood vessel or lymphatic endothelial cells, platelets or red blood cells did not affect hearing, suggesting that systemic loss of Spns2 activity in these tissues does not mediate the hearing loss we see in the Spns2tm1a mutants. However, when we deleted Spns2 locally in the inner ear using the Sox10-Cre recombinase, the resulting mutants were deaf. Sox10 is expressed throughout the otic epithelium from an early stage of development as well as in cranial neural crest-derived cells, so can effectively drive deletion of exon 3 of the Spns2tm1c allele in the entire inner ear [74], [75]. These findings indicate that hearing loss in Spns2tm1a/tm1a mice is due to local loss of Spns2 function in the inner ear.
Defects of the anterior eye were only seen in the Spns2tm1a and Spns2tm1b homozygous mutants. We did not see anterior eye defects in any of the 5 conditional alleles, consistent with normal anterior eye development in another conditional Spns2;Tie2-Cre mutant mouse [1]. The anterior eye phenotype appears to be due to a developmental abnormality resulting in defective eyelid formation and subsequent corneal opacity and vascularisation. Spns2 also plays a role in retinal blood vessels. Our results showed that global Spns2 knockout resulted in a mild phenotype of the retinal vasculature (thin and irregular veins) with decreased pericyte coverage in the central retina which may be related to the widely known role of S1P signalling in angiogenesis [76], [77]. The milder vascular phenotype in the retina than in the cochlea may be due to differences in the requirement for Spns2 in these tissues. We also found focal retinal degeneration in these mutant eyes suggesting a role for Spns2 in the photoreceptor and/or retinal pigment epithelium. Taken together, these findings suggest that SPNS2 is not only a candidate gene for involvement in deafness, but also for deaf-blind syndromes.
In summary, we report here that Spns2-deficient mice displayed rapidly progressive hearing impairment associated with a rapid decline in the EP between P14 and P21. The mechanism by which Spns2 deficiency leads to decreased EP merits further investigation, but it most likely involves local S1P signalling. Following the early drop in the EP, later changes include reduced expression of key proteins involved in cochlear homeostasis and ultimately sensory hair cell loss. Our findings suggest that Spns2 is a promising candidate gene for human deafness. Furthermore, Spns2-deficient mice may serve as a model to learn more about the role of S1P signalling in auditory function and the mechanism underlying at least one form of strial hearing loss.
Mouse studies were carried out in accordance with UK Home Office regulations and the UK Animals (Scientific Procedures) Act of 1986 (ASPA) under a UK Home Office licence, and the study was approved by the Wellcome Trust Sanger Institute's Ethical Review Committee. Mice were culled using methods approved under this licence to minimize any possibility of suffering. The mice were maintained in individually-ventilated cages at a standard temperature and humidity and in specific pathogen-free conditions. Either sex was used for this study.
The Spns2 mutant allele we used carries a promoter-driven cassette designed to interrupt normal gene transcription but flanked by Frt sites to enable its removal and conversion to a conditional allele with a critical exon surrounded by LoxP sites (a knockout-first design; [21], [78]). The allele is designated Spns2tm1a(KOMP)Wtsi, abbreviated to Spns2tm1a in this study. A schematic of the knockout-first design of Spns2tm1a allele is shown in Fig. 1A. The mutant mice were generated by blastocyst injection of the targeted ES cell using standard techniques [20], [21] and germ line transmission of Spns2tm1a was confirmed by a series of genotyping PCR analyses [79].
The Spns2tm1a colony was maintained on a mixed genetic C57BL/6BrdTyrc-Brd;C57BL/6Dnk;C57BL/6N background. Spns2tm1a/tm1a mice were crossed to HprtTg(CMV-Cre)Brd/Wtsi transgenic mice (on a C57BL/6NTac background) with systemic expression of Cre recombinase to remove the cassette and produce mice carrying the Spns2tm1b allele (Fig. 1A). Mice showing the correct excision were mated to wildtype C57BL/6N mice and offspring carrying the Spns2tm1b allele were mated to breed out the Cre allele and expand the colony. The Spns2tm1c allele was produced by crossing Spns2tm1a/tm1a mice to Gt(ROSA)26Sortm1(FLP1)/Wtsi mice expressing Flp recombinase ubiquitously in which the promoter-driven cassette was excised and exon 3 was retained flanked by LoxP sites (Fig. 1A). All the genotyping PCR primers and product sizes are shown in Table 1. Lack of the rd8 mutant allele was confirmed by conventional sequencing of the Crb1 gene [33].
Spns2tm1c/tm1c mice were mated to Sox10-Cre mice (Tg(Sox10-cre)1Wdr) [31] to delete the floxed exon 3 and to generate a frameshift mutation of Spns2 in the inner ear and craniofacial neural crest-derived tissues [31]. Genotyping was carried out using genomic DNA extracted from pinna tissue, which was mosaic under this conditional knockout strategy, and the conditional Spns2tm1d allele was confirmed by co-presence of Spns2tm1c and Sox10-Cre allele PCR bands. Since S1P is released from different blood cells and endothelial cells, we used mice expressing Tie1-Cre [80], Pf4-Cre [81], Lyve1-Cre [82] and EpoR-Cre [83] to inactivate the Spns2 gene in blood vessel endothelial cells, platelets, lymphatic endothelial cells and red blood cells respectively by crossing with Spns2tm1c/tm1c mice to produce the conditional knockouts.
The organ of Corti, lateral wall (stria vascularis and spiral ligament), eyes and livers of postnatal day (P)4 homozygous, heterozygous and wildtype littermate mice were dissected in RNAlater (n = 3 for each genotype). Total RNA was isolated with QIAshredder columns (QIAgen, cat. no. 79654) and the RNAeasy mini kit (QIAgen, cat. no. 74104). RNA was normalized to the same concentration for cDNA synthesis using oligo dT and SuperScrip II (Invitrogen). Real-time PCR was performed in triplicate for each sample using a CFX connect real time PCR machine (BIO-RAD). The Spns2 probe was designed to cover the 3′ untranslated region (Applied Biosystem). Hypoxanthine-guanine phospharibosyltransferase (Hprt) was amplified simultaneously (Applied Biosystem, Mm01318747_g1) as an internal reference. The relative quantity of Spns2 was calculated using the 2−ΔΔCt method [84].
X-gal staining can be used to visualise the expression of Spns2 due to the LacZ gene inserted in the cassette of Spns2tm1a allele (Fig. 1A), downstream of the Spns2 promotor. Inner ears of P10 and P14 heterozygous and homozygous mice (at least three mice of each age group) were dissected out and fixed in 4% paraformaldehyde (PFA) for 45 minutes to 2 hrs. These were washed twice with PBS and decalcified in 10% EDTA until soft. After a PBS wash and immersing in 30% sucrose, inner ears were embedded in Agarose type VII (low gelling temperature, Sigma-Aldrich), then mounted using OCT compound ready for cryosectioning at 14 µm. Sections were treated with Solution A (2 mM MgCl2; 0.02% NP-40; 0.01% sodium deoxycholate; PBS) for 15 mins, then incubated with Solution B (Solution A plus 5 mM K3Fe(CN)6; 5 mM K4Fe(CN)6; 1 mg/ml X-Gal in DMSO) over night at 37°C. Sections were rinsed in water then counterstained in Fast Red to label nuclei, mounted and examined.
Mice were anaesthetised by ketamine hydrochloride (100 mg/Kg, Ketaset, Fort Dodge Animal Health) and xylazine hydrochloride (10 mg/Kg, Rompun, Bayer Animal Health) and subcutaneous needle electrodes were inserted on the vertex (active), and over the left (reference) and right (ground) bullae. A calibrated sound system was used to deliver free-field click (0.01 ms duration) and tone pip (various frequencies from 6–30 kHz of 5 ms duration, 1 ms rise/fall time) stimuli at a range of intensity levels in 5 dB steps. Averaged responses to 256 stimuli, presented at 42.2 per second, were analysed and thresholds established as the lowest sound intensity giving a visually-detectable ABR response [85]. For P14 and P21 mice, in order to achieve higher sound pressure levels, sound was delivered to the external auditory meatus via a parabolic cone loud speaker attachment for click (0.01 ms duration) and tone pip stimuli (frequencies from 3–42 kHz of 5 ms duration in 3 dB SPL steps). Separate cohorts of P14 and P21 Spns2tm1a mice were used at the standard maximum intensity of 95 dB SPL with free field delivery as shown in Fig. 11 and at the higher sound intensities delivered near field, directly to the external auditory meatus in Fig. 1. The median ABR thresholds recorded in homozygous mutants at P14 and P21 shown in Fig. 1 were compared using the Kruskall-Wallis One-Way Analysis of Variance on Ranks, as thresholds did not show a normal distribution.
The temporal bones were isolated. The inner ears were dissected out and fixed by 2.5% glutaraldehyde in 0.1M sodium cacodylate buffer with 3 mM calcium chloride at room temperature for 3 hours. Cochleae were finely dissected in PBS. This was followed by further processing using an osmium-thiocarbohydrazide-osmium (OTOTO) method [86]. The samples were dehydrated in increasing concentrations of ethanol, critical-point dried (Bal-Tec CPD030), mounted and examined under a HITACHI S-4800 scanning electron microscope. At least 3 wildtype, heterozygous and homozygous mice were examined for each age group (P4, P21, P28 and P56). Middle ears were dissected and examined. Inner ears were cleared by a standard glycerol clearing technique and examined for gross structural defects (control, n = 5; homozygotes, n = 5; aged P30–34).
Inner ears (wildtype, n = 2; heterozygotes, n = 2; homozygotes, n = 4, at P28) were dissected out and gently perfused with 2.5% glutaraldehyde, 1% paraformaldehyde in 0.1M sodium phosphate buffer with 0.8 mM calcium chloride through the round and oval windows and a small hole in the apex then fixed at room temperature for 2 hours. Secondary fixation was in 1% osmium tetroxide in sodium phosphate buffer for 1 hour. After 5 washes in 0.1 M sodium phosphate buffer, inner ears were decalcified in 0.1M EDTA at 4°C until soft. Then the samples were dehydrated through an ethanol series, staining in 2% uranyl acetate at the 30% ethanol stage, embedded in Epon resin mixed 1∶1 and then 3∶1 in propylene oxide for 30 min and infiltrated overnight under vacuum in neat resin. The samples were embedded at 60°C for 24–48 hours. 1 µm sections were cut through the modiolar plane and stained with toluidine blue for light microscope observation. 60 nm sections were cut on a Leica EM UC6 ultramicrotome, stained in 2% uranyl acetate and aqueous lead citrate and imaged on an FEI Spirit Biotwin 120 kV transmission electron microscope using a Tietz F4.15 CCD.
Mice were anaesthetized with 0.01 ml/g body weight of 20% urethane, a tracheal cannula was inserted and the bulla was opened to reveal the cochlea while the body temperature was kept at 37°C by a feedback-controlled heating pad. A small hole was made in the bony wall of the cochlea over the basal turn of scala media, and a micropipette electrode filled with 150 mM potassium chloride was advanced through the hole and through the lateral wall into the scala media. The potential difference between the scala media and a reference silver/silver chloride pellet under the dorsal skin was recorded [24].
The inner ears were rapidly dissected out and fixed in 4% paraformaldehyde at room temperature for 2 hours. The lateral walls were dissected out in PBS for surface preparation. Filamentous actin was visualized by rhodamine phalloidin (1∶200, Molecular Probe) at room temperature for 2 hours. Strial capillaries were visualized by Isolectin B4 (Vector Laboratories, 1∶50) at 4°C, overnight in PBS with 10% sheep serum). Samples were mounted with Vectashield Mounting medium (Vector, Cat. No: H-1000) and imaged by confocal microscopy (Carl Zeiss, LSM 510 META). The numbers of capillary branch points per field (220×220-µm fields) in the middle turn (40–70% of the distance along the cochlear duct from the base) of the stria vascularis (control, n = 4; homozygotes, n = 5, at P14. control, n = 3; homozygotes, n = 5, at P28) was quantified using image J. Data were presented as a density in a 100×100 µm field and statistics analysis was conducted using Mann-Whitney Rank Sum Test, SigmaPlot v12.0. Surface preparations were also used for analysis of Kcnq1 expression, using overnight incubation at 4°C with goat anti-Kcnq1 polyclonal antibody (Santa Cruz, 1∶200) followed by washing with PBS and incubation with donkey anti-goat secondary antibody (Invitrogen, 1∶500) prior to analysis by confocal microscopy. At least three homozygotes and three controls were used at P14, P28 and 6 months for phalloidin labeling to show marginal cell boundaries, and P14 and 5–6 weeks for Kcnq1 expression. The density of marginal cells was measured in the phalloidin-labelled whole mount preparations by counting the number of cells defined by their labeled boundaries in two areas each 100×100 µm from the middle turn (40–70%) of each cochlea (n = 4 homozygous mutants; 4 littermate controls).
The cochleae were dissected out and fixed in 4% PFA at room temperature for 2 hours. Cryosections were obtained as described above for X-gal staining. We used the following antibodies: rabbit anti-Kcnj10 polyclonal (Alomone labs, 1∶300), rabbit anti-Gjb2 polyclonal (from WH Evans, 1∶300), rabbit anti-Gjb6 polyclonal (Zymed, 1∶400), mouse anti- Na+, K+-ATPase (α1 subunit) monoclonal (Sigma, 1∶300), mouse anti- NKCC1 monoclonal (C. Lytle, Developmental Studies Hybridoma Bank, University of Iowa, 1∶300) and rabbit anti-ZO-1 polyclonal (Zymed, 1∶300). Sections were blocked by incubation with 10% sheep serum (with 0.1% TritonX-100 in PBS) for 40 mins. Sections were incubated with appropriate primary antibodies overnight at 4°C, washed with PBS and incubated with corresponding secondary antibodies at room temperature for 2 hours (donkey anti-rabbit, donkey anti-goat, Invitrogen, 1∶500). After washing with PBS, slides were imaged by confocal microscopy. Three mice of each genotype (Spns2tm1a/tm1a and Spns2+/+) were used for each antibody at P14 and 5–6 weeks old.
The inner ears (wildtype, n = 1; heterozygote, n = 3; homozygote, n = 4 at 6 weeks old) were dissected and round and oval windows were opened in PBS containing 1 mM calcium chloride. A hole was made in the basal turn leading to the scala media. The membranous labyrinth was perfused for 5 minutes with 400 µl Sulfo-NHS-LC-Biotin (Thermo Scientific, 10 mg/ml, in PBS with 1 mM calcium chloride) through the round and oval windows and the hole exposing the endolymphatic compartment. Following a PBS wash, the inner ears were fixed in 4% paraformaldehyde at room temperature for 2 hours, and processed for cryosectioning as described above. The biotin tracer was detected by fluorescein isothiocyanate (FITC)-conjugated streptavidin (Thermo Scientific, 1∶100) incubating at room temperature for 30 min. Samples perfused with PBS alone were used as negative controls.
Mice (wildtype, n = 2; heterozygote, n = 2; homozygote, n = 2; at 5–6 weeks old) were warmed in a 39°C incubator for 5 minutes and then held in a mouse restrainer so that the tail was accessible and the tail vein visible. A 50 µl aliquot of 5% (w/v) FITC-conjugated bovine serum albumin (BSA-FITC; Sigma cat. no. A9771; size 66.4 kDa), made up in sterile PBS, was injected into the tail vein. The mice were left at room temperature for 45–60 minutes to allow the BSA-FITC to permeate all capillaries and to allow for any vascular extravasation. The mice were sacrificed by CO2 inhalation and the auditory bullae dissected out. Whole cochleae were exposed and fixed by removing a small piece of bone at the apex and gently perfusing 4% PFA through the round and oval windows. Cochleae were then immersed in fixative and left on a rotator for 1.5 hours at room temperature. Whole-mounts of stria vascularis were dissected from fixed cochleae, covered with Vectashield Mounting Media (Vector, Cat. No: H-1000) in glass bottom culture dishes (MatTek Corp.) and imaged using confocal laser-scanning microscopy (Carl Zeiss, LSM 510 META).
Mice underwent ophthalmic screening at 15 weeks of age. They were assessed for gross morphological changes to the eye using a slit lamp (Zeiss SL130) and ophthalmoscope (Heine Omega 500). The eye was examined both undilated and dilated (topical tropicamide). Images using the slit lamp were collected using a Leica DFC420 camera. The mice were culled under terminal anaesthesia followed by cervical dislocation and both eyes from 3 male homozygous mutants and 3 wildtype mice were removed and fixed. Pupil-optic nerve sections were processed, stained with hematoxylin and eosin, and standard images were captured under light microscopy for review [87].
For whole mount retinal analysis, heterozygotes and homozygotes were used (n = 3 for each genotype at P10, n = 2 at 8 weeks old) and the eyes removed and fixed in 4% PFA. Retinae were prepared and stained as described [88] using flourescein-conjugated Griffonia simplicifolia Isolectin B4 (Vector Laboratories, UK) to label blood vessels, mouse anti-proteoglycan NG2 (Millipore UK Ltd., UK) to label pericytes, and donkey anti-rabbit secondary Alexa-594 (Molecular Probes, Life Technologies, UK). The homozygote and control mouse retinae were stained in the same well to control for changes in staining efficiency, distinguishing the retinae by different numbers of radial incisions. All tissues were mounted in Vectashield (Vector Laboratories Ltd., Peterborough, UK), imaged by confocal microscopy (Nikon A1R; Nikon Instruments, Inc., Melville, NY), and maximum intensity projections of z-stacks were created using NisElements AR Version 4.0 software (Nikon UK, Kingston Upon Thames, UK). The MetaMorph Angiogenesis Tube Formation application (Molecular Devices, Berkshire, UK) was used for quantification. Confocal images were used to determine the total area covered by vessels and by pericytes to calculate the percentage pericyte coverage of vessels and the total number of capillary branchpoints per unit area. We imaged three areas of each retina: three different regions of the central retina each encompassing an artery and vein, and two images each of peripheral arteries and peripheral veins, a total of five images/retina. Threshold values were kept the same for analysis of samples of the same stage.
|
10.1371/journal.pcbi.1002694 | Tissue-Specific Functional Networks for Prioritizing Phenotype and Disease Genes | Integrated analyses of functional genomics data have enormous potential for identifying phenotype-associated genes. Tissue-specificity is an important aspect of many genetic diseases, reflecting the potentially different roles of proteins and pathways in diverse cell lineages. Accounting for tissue specificity in global integration of functional genomics data is challenging, as “functionality” and “functional relationships” are often not resolved for specific tissue types. We address this challenge by generating tissue-specific functional networks, which can effectively represent the diversity of protein function for more accurate identification of phenotype-associated genes in the laboratory mouse. Specifically, we created 107 tissue-specific functional relationship networks through integration of genomic data utilizing knowledge of tissue-specific gene expression patterns. Cross-network comparison revealed significantly changed genes enriched for functions related to specific tissue development. We then utilized these tissue-specific networks to predict genes associated with different phenotypes. Our results demonstrate that prediction performance is significantly improved through using the tissue-specific networks as compared to the global functional network. We used a testis-specific functional relationship network to predict genes associated with male fertility and spermatogenesis phenotypes, and experimentally confirmed one top prediction, Mbyl1. We then focused on a less-common genetic disease, ataxia, and identified candidates uniquely predicted by the cerebellum network, which are supported by both literature and experimental evidence. Our systems-level, tissue-specific scheme advances over traditional global integration and analyses and establishes a prototype to address the tissue-specific effects of genetic perturbations, diseases and drugs.
| Tissue specificity is an important aspect of many genetic diseases, reflecting the potentially different roles of proteins and pathways in diverse cell lineages. We propose an effective strategy to model tissue-specific functional relationship networks in the laboratory mouse. We integrated large scale genomics datasets as well as low-throughput tissue-specific expression profiles to estimate the probability that two proteins are co-functioning in the tissue under study. These networks can accurately reflect the diversity of protein functions across different organs and tissue compartments. By computationally exploring the tissue-specific networks, we can accurately predict novel phenotype-related gene candidates. We experimentally confirmed a top candidate gene, Mybl1, to affect several male fertility phenotypes, predicted based on male-reproductive system-specific networks and we predicted candidates related to a rare genetic disease ataxia, which are supported by experimental and literature evidence. The above results demonstrate the power of modeling tissue-specific dynamics of co-functionality through computational approaches.
| Phenotypes caused by mutations in genes often show tissue-specific pathology, despite organism-wide presence of the same mutation [1], [2], [3], [4]. Therefore, a logical genomics approach to infer candidate genes and their functions is to integrate large-scale data in a tissue-specific manner. However, such efforts are hampered by the lack of adequate tissue-specific training and feature data and by the methodologies to model tissue-specificity systematically in human or other mammalian model organisms.
Functional relationship networks, representing the likelihood that two proteins participate in the same biological process, provide invaluable information for phenotype gene discovery, pathway analysis, and drug discovery [5], [6], [7], [8], [9], [10], [11]. In human and model mammalian organisms, these networks have been used to predict genes associated with genetic diseases or phenotypes through computational mining of the network structure [5], [6], [7], [8], [10], [11]. For example, we have previously generated a mouse functional relationship network and used it to identify that Timp2 and Abcg8 are bone-mineral density (BMD)-related genes [11], though neither of these were previously detected in quantitative genetics studies. So far, these analyses have been limited to global functional networks representing the overall relationships between proteins without accounting for tissue specificity. Analyses based on global functional relationship networks, while effective, ignore a critical aspect of biology that could significantly improve their utility: genetic diseases often target specific tissue(s) and thus perturbations of proteins or pathways may have differential effects among diverse tissues. For example, Timp2, which we have previously identified to be related to BMD [11], is also involved in the control and/or development of neurodegenerative disease [12]. Such multi-functionality is not directly reflected by the global network but would be revealed by different connections in tissue-specific networks. Therefore, computational modeling and analyses of tissue-specific networks are needed to identify phenotype-associated genes that exhibit tissue-specific behavior.
Current approaches to create functional relationship networks are difficult to apply in a tissue-specific manner. Typically, networks are constructed by integrating data sources that vary in terms of measurement accuracy as well as biological relevance for predicting protein functions. Machine learning methods, such as Bayesian networks, learn the relative accuracy and relevance of datasets when given a ‘gold standard’ training set, which consists of gene pairs that are known to work in the same biological process. Then probabilistic models are constructed to weigh and integrate diverse datasets based on how accurately they recover the ‘gold standard’ set. The networks generated by this approach lack tissue-specificity information, because systematic collections of large-scale data or ‘gold standard’ pairs with quantitative tissue-specific information are often not available.
Here, we address the tissue-specificity challenge by simulating the natural biological mechanism that defines tissue-specificity: co-functionality in most cases would require the presence of both proteins in the same tissue. Inspired by our previous efforts to establish biological process-specific networks, such as networks specifically related to the cell cycle or to mitochondrial biogenesis [13], [14], [15], we integrate low-throughput, highly-reliable tissue-specific gene expression information (e.g. RT-PCR, in situ hybridization, etc.) from the Mouse Gene Expression Database (GXD) into our probabilistic framework when learning the reliability of each data source. Such an approach is more intuitive for the tissue-specific network setting because it is relatively less likely that a non-expressed gene would collaborate with an expressed gene even though they are ‘functionally related’ in the global sense (i.e. co-annotated to either a GO term or a KEGG pathway). There are exceptions to this guideline, including signaling and hormonal pathways that traverse multiple organ systems. However, many cellular processes important for phenotypes are largely restricted to specific tissues. Therefore, by constraining the ‘gold standard’ to pairs of genes that are both expressed in a tissue, we are able to establish functional networks that are highly specific in capturing the dynamic properties of different tissues.
In addition to generating the first tissue-specific networks for the laboratory mouse, we also explicitly tested the potential of using such networks to predict phenotype-associated genes. To do so, we mapped diverse phenotypes to their respective tissues in the laboratory mouse, according to the terminology and description of the phenotypes. We show that the tissue-specific functional relationship networks can improve our prediction accuracy for phenotype-associated genes compared to a single global functional relationship network through computational analyses, and through experimentally confirmed predictions of novel fertility-related genes and visualization of their local networks. We further identified candidate genes specifically predicted by the cerebellum network to be related to ataxia, which are supported by both literature and experimental evidence. Our networks are publicly available at http://mouseMAP.princeton.edu, which features the ability to compare networks across tissues for analyzing the dynamics of functional relationships. Our current framework covers 107 major tissues in the laboratory mouse and focuses on cross-network comparison and phenotype-associated gene discovery. However, as more data become available, this approach will serve as a prototype for applications to pathway analyses and drug screening.
In this study, we develop and apply a novel algorithm that generates tissue-specific functional networks in the laboratory mouse by integrating diverse functional genomic data, and we demonstrate that our tissue-specific networks are more accurate in predicting phenotype-related genes than a single global functional network. In the following sections, we first outline the strategy used to generate tissue-specific networks by interrogating gene expression profiles across tissues and integrating different data sources using Bayesian statistics. Second, we developed a cross-network comparison metric for identifying significantly changed genes across networks which are enriched in tissue specification and development. Third, we quantitatively demonstrate that combining our tissue-specific networks with a state-of-the-art machine learning algorithm can produce improved predictions of genotype-phenotype relationships compared to previous single global networks [11]. Fourth, we identify candidate genes related to male fertility specifically predicted by our tissue-specific networks (but not by the global network), and verify a top prediction in an independent, unbiased mutant screen. Finally, we used cerebellum-specific network to predict genes associated to a less-studied disease, ataxia, which are supported by both literature and experimental evidence. The predictions made by our approach for all examined networks are available online at http://mouseMAP.princeton.edu.
A common mechanism resulting in tissue-specific protein functionality is the modulation of gene expression levels between tissues [16], [17], [18]. This observation is our theoretical foundation for establishing tissue-specific networks, in which links between proteins represent the probability that they are involved in the same biological processes within a specific tissue. To simulate such tissue-specificity, we developed a Bayesian approach (Figure 1) that incorporates highly-reliable, low-throughput measures of tissue-specific gene expression into training set, which we utilized to produce networks focused on the real functional relationships occurring within the tissue under consideration. This Bayesian framework essentially learns how informative each dataset is given a set of ‘gold standard’ training pairs, i.e. pairs of proteins known to be functional in the same biological process and both expressed in the tissue of interest.
In the global (non-tissue-specific) sense, following previous definitions [5], ‘gold standard positives’ are defined by co-annotation to specific Gene Ontology (GO) biological process terms [19], while ‘gold standard negatives’ are defined as pairs that both have specific GO annotations yet do not share any annotations. For each tissue-specific gold standard set, a positive pair has to meet two requirements: first, the pair must be ‘co-functional’ as defined in the global sense, and second, both genes must be expressed in the tissue under consideration as evident in highly reliable, low-throughput expression datasets, which, in most cases, is necessary for the pair to have a functional relationship in that tissue. These tissue-specific gold standards are then used to quantify how relevant each genomic dataset is in recovering tissue-specific functional relationships, regardless of the tissue of origin for each genomic dataset. This allows us to leverage the entire compendium of high-throughput genomic data to generate accurate tissue-specific networks, even for tissues which do not have existing tissue-specific whole-genome experiments, by relying on non-tissue-specific datasets, heterogeneous samples, and potentially related tissues and experiments. For example, biliary tract, which is not specifically represented in our current collection of high-throughput features used for classification, can still be accurately predicted by utilizing information from related, heterogeneous samples, such as gene expression microarrays of whole liver or the hepatic system, as well as non-tissue-specific information, such as sequence phylogeny and in vitro binding assays. Thus our approach can leverage the implicit relationships between a dataset and a tissue and therefore enables generation of tissue-specific networks even from feature data that is not resolved for a specific tissue type.
For tissue-specific expression information, our gold standards rely on the Gene Expression Database (GXD) of the Mouse Genome Informatics group (MGI). GXD provides an extensive, hierarchically structured dictionary of anatomical expression results for mouse to allow us to carry out our analysis [20]. The data in GXD are derived from traditional, “small-scale” expression experiments, such as in situ hybridization, RT-PCR, and immunohistochemistry, which simply reflect presence or absence of a gene within the tissue examined. No high-throughput expression data were used for our gold standard construction. In total, there are 107 tissues included in our analysis.
We pursue two main goals in this study: First, we generate tissue-specific networks that synthesize as much data as possible and provide these networks to the public through an online visualization interface at http://mouseMAP.princeton.edu. For this, we gathered diverse genomic data for mouse as inputs (Dataset S1) to support the functional relationships, including protein-protein physical interactions [21], [22], [23], [24], homologous functional relationship predictions from simpler organisms [9], phenotype and disease data [19], [25] and 960 expression datasets, totaling 13632 experimental conditions [26], [27], [28], [29]. The reliability of each dataset is learned through Bayesian network classifier training, using the tissue-specific gold standards described above. Essentially, a dataset deemed more relevant and accurate for the tissue under consideration will be given higher weight, and the final probability of pair-wise functional relationships is determined by updating the initial probability (prior) based on the weighted input of all genomic datasets. This procedure resulted in tissue-specific probabilistic functional relationship networks for the laboratory mouse that effectively summarize these diverse data sources and enable biology researchers to easily explore the resulting functional landscape. Second, we test the hypothesis that tissue-specific networks could assist us to predict phenotype-related genes more accurately. In this case, to prevent circularity in our methodology, phenotype and disease data were excluded from network generation, and the results were used to predict novel phenotype-associated candidate genes. We demonstrate that tissue-specific networks enhance biological clarity and result in more accurate predictions. Our resulting networks and predictions provide biology researchers with functional interactions specific to each tissue as well as phenotype hypotheses of genes.
One key application of tissue-specific networks is to identify novel genes and relationships between genes that may be specific to a particular tissue. To computationally evaluate our ability to identify novel relationships, we used cross-validation to test whether our tissue-specific Bayesian scheme is more accurate than the global network. Cross-validation was used to assess predictions by evaluating the accuracy of recovering subsets of known annotations withheld during the training process. Specifically, we performed 3-fold cross-validation, by holding out one third of the tissue-specific ‘gold standard’ pairs in each of the three iterations. We learned the parameters in the Bayesian networks, i.e., the reliability of each dataset, through the other two thirds of the ‘gold standard’, and then used these networks to predict the probabilities for the held-out one third of the protein pairs.
Compared to a single global functional relationship network, our approach significantly improved our ability to predict tissue-specific functional linkages. The mean AUC (area under the receiver operating characteristic curve, which represents the accuracy in recovering tissue-specific functional relationships) for the global network estimated through three-fold cross-validation was 0.68. Tissue-specific networks achieved median AUC of 0.72. With a random baseline of 0.5 in AUC, this represents a ∼20% improvement of the tissue-specific networks over the predictive power of the global network. This improvement is consistent over all 12 major organ systems defined by GXD [20]. (Figure 2A). Immune system-related networks acquired the most median improvement of 22.7% and digestive system-related networks achieved least median improvement of 14.3%. For example, for lymphoid system (MA:0002435), we improved our AUC from 0.65 to 0.72 and for ventricular zone, brain, we improved from 0.65 to 0.77. Such improvement is consistent across the entire precision-recall spaces (Figure 2B, Dataset S2 for all precision-recall curves). In all cases, tissue-specific networks performed better than the global network in predicting functional relationships specific to that tissue, which demonstrates the robustness of our integration approach across different systems and tissues in the laboratory mouse.
One important application of our tissue-specific networks is to identify functional relationships between genes that change significantly across tissues. This provides a platform for analyzing tissue-specific molecular interactions, as well as tissue-specific roles for genes that are ubiquitously expressed but play different roles in different tissues. For example, Wnt10b (wingless related MMTV integration site 10b) is expressed in many tissues throughout development and participates in many biological processes including bone trabecular formation [30] and cell differentiation involved in skeletal muscle development [31]. The interactors of Wnt10b in our muscle-specific and bone-specific functional networks reflect its differential roles in these two tissues. The top neighbors in the muscle-specific network consist of genes responsible for skeletal muscle development (Figure 3A). For example, BIN1 participate in the biological process muscle cell differentiation (GO:0042692) [32], PLAU is involved in the process skeletal muscle tissue regeneration (GO:0043403) in rat and MYF6 directly function in muscle cell biogenesis [33]. In fact, 8 out of the 19 top connected nodes of Wnt10b in the muscle-specific network are involved in skeletal muscle cell development, reflecting the functional role of Wnt6b. On the contrary, in the bone-specific functional network, the top neighbors of Wnt10b consist of genes involved in bone mineralization and bone structure formation (Figure 3B), representing 12 out of the 19 top connected nodes. This observation suggests that our networks can provide a resource for comparing the dynamic functions of a single gene across different tissues.
To quantify gene connectivity changes across networks, we developed a metric that captures how much the edges involving a gene differ across networks (see methods), and we implemented a web-based visualization interface (http://mouseMAP.princeton.edu) allowing users to query genes of interest and compare the local network between tissues. Essentially, connectivity change of a gene is defined by the sum of absolute values of fold changes (over prior) of connections between this gene to all other genes. Some genes vary greatly in their connectivity between tissues, potentially reflecting their tissue-specific roles. Of the top 100 altered genes, they were significantly enriched for “anatomical structure development” (GO:0048856) and “organ development” (GO:0048513). Additionally, genes with connectivity altered in specific tissues compared to the global network, tend to be enriched for GO terms related to the tissue under consideration. For example, when comparing the nervous system-specific network (MA:0000016) against the global network, the top changed genes are enriched in “central nervous system development” (GO:0007417), “diencephalon development” (GO:0021536), and “brain development” (GO:0007420) (Table 1). The full enrichment analysis is provided in Dataset S3.
A key hypothesis in this study is that analyzing tissue-specific networks may improve our ability to identify phenotype-related genes. To test this hypothesis, we regenerated tissue-specific networks using the same Bayesian approach as above, but excluded all phenotype and disease data as inputs to avoid circularity in our cross-validations. Then, we mapped 451 phenotypes to their most related tissue in the laboratory mouse according to the terminology and description of these phenotypes in the Mammalian Phenotype ontology [19]. For each phenotype, we compared novel predictions made using the appropriate tissue-specific network as compared to using the global network. This method is based on our previously developed machine learning scheme (network-based SVM) [11] that mines information in functional relationship networks to prioritize candidate genes according to their links to known genes related to a disease or phenotype.
To test whether our tissue-specific networks are more capable of identifying phenotype-associated genes than the global network, we used bootstrap bagging [34] to evaluate which network performs better. Bootstrap bagging is suitable for phenotype predictions, where positive examples (known phenotype-associated genes) and negative examples (random genes) are highly imbalanced [35]. Its stability and comparably good performance in estimating error rates has been tested in extensive simulations for positive example set sizes ranging from less than 20 [35] to >200 [36], which is the approximate range we are using in our evaluation. For the 451 mapped phenotypes, the median AUC when utilizing tissue-specific networks is 0.794, representing an improvement of 11.8% over utilizing the global functional network. For many phenotypes, using tissue-specific networks can improve our ability to extract potentially experimentally-verifiable predictions. For example, at one percent recall (the low recall end is where most of the follow-up experimental confirmations will focus on), we achieved a precision of 1.00 compared to 0.33 using global network for the phenotype abnormal spleen white pulp morphology (MP:0002357), and a precision of 0.5 compared to 0.28 for abnormal malpighian tuft morphology (MP:0005325). Additionally, the AUC for “abnormal osteogenesis” (MP:0000057) was 0.77 using the global network, but 0.81 for tissue-specific networks. The AUC for “abnormal nervous system electrophysiology” (MP:0002272) using the global network was 0.716, but was 0.763 using the nervous system-specific network (Figure 4C for example precision-recall curves). Such significant improvement demonstrates the potential of mining tissue-specific networks to prioritize phenotype-associated genes.
Performance improvements were consistent across phenotypes of different sizes (Figure 4A). For phenotypes with 300–1000 annotated genes (around 1.5% to 5% of genome), we achieved a median AUC of 0.814 (improvement of 8.7%); for phenotypes with 100–300 genes, the median AUC was 0.792 (improvement of 13.0%); and for phenotypes with 30–100 genes, the median AUC was 0.769 (improvement of 11.0%). At 10 percent recall for the 300–1000, 100–300, and 30–100 groups, we achieved precisions of 14.8, 17, and 20 fold over random, respectively. This consistency indicates the robustness of tissue-specific networks against the number of known genes in predicting phenotype-associated genes.
Performance improvements were also consistent across different major organ systems. Phenotypes involved in the endo/exocrine system achieved the most significant improvement in AUC (+35%, compared to global networks against baseline of 0.5) and those in cardiovascular system achieved 21.8% improvement in AUC. However, prediction accuracy was improved across all major systems, with the least improvement of 5.9% in renal/urinary phenotypes. Phenotypes related to musculoskeletal systems achieved the highest AUC of 0.82 and the group with lowest AUC was digestive system, which still achieved an average of 0.78. The consistency in improvements across different organ systems demonstrates the robustness of our modeling framework to predict phenotype-related genes in a tissue-specific manner.
We focused on two cases to illustrate how our tissue-specific networks can facilitate disease gene discovery. These two phenotypes represent two extremes of the phenotype/disease-associated gene prediction problem. The first, reduced male fertility, is a broadly defined, common phenotype with many causative genes already known. The second, ataxia, is a rare neurological disorder affecting ∼3–10/100,000 of the general population [37], [38], [39]. Roughly 40 genes are known to be associated with this disease, but the majority of both familial and sporadic cases remain unexplained. Predicting candidate genes related to rare genetic diseases is challenging in that little prior knowledge is available for these diseases. These phenotypes are related to two different tissue-categories (reproductive and neurological systems), enabling us to highlight the broad applications of our approach across organ systems. We used these two examples and experimental confirmations to demonstrate the power of tissue-specific networks to discover disease genes.
First, we used male fertility related phenotypes to test the performance of tissue-specific networks to predict phenotype-related genes. To do so, we utilized a recent, nearly comprehensive literature review of genes involved in mammalian spermatogenesis and male fertility phenotypes [40], which we organized into a hierarchy of male fertility-related phenotypes (Dataset S4). This curation effort is independent of, and more comprehensive than, the current GO or MP annotations related to male fertility, which makes these lists excellent, non-circular test sets. We tested whether the testis-specific network could predict male fertility genes more robustly than the globally integrated network, and found that the testis-specific network significantly improved our ability to predict spermatogenesis-related phenotypes. For example, for predicting genes related to ‘spermatid head and nuclear modifications,’ we achieved 4.5-fold improvement in precision at 1 percent recall; for ‘acrosome-related genes,’ we achieved 3.6-fold improvement; and for ‘germ/Sertoli cell interaction genes,’ we achieved 3.3-fold improvement. On the other hand, for terms that are not specifically related to male-reproductive systems, such as ‘association with methylation and acetylation,’ and ‘association with Golgi Apparatus,’ we observe no performance improvements using the testis-specific network. This illustrates that tissue-specific functional relationship networks are tuned to predict phenotypes closely related to these tissues.
We selected Mybl1 to demonstrate the specific utility of the male-reproductive network to predict fertility related genes. Mybl1 (MGI:99925) is among our top candidates in multiple phenotypes related to male fertility, including ‘association with chromatoid body and manchette’, ‘transcription factor involved in spermatogenesis’ and, ‘spermatogenesis’. However, in the global network, Mybl1 was not a strong candidate for these phenotypes, as it was predicted with negative values. Therefore Mybl1 is an ideal candidate to test the accuracy of our tissue-specific network-based phenotype predictions. In our male-reproductive network, the majority of the top interactors of Mybl1 are indeed well-known male fertility genes (Figure 5A), including Dmc1 (required for meiosis and male fertility [41]), Ddx4 (a DEAD-box helicase required for male, but not female, germ cell development [42]), Cyct (encoding testis-specific cytochrome c [43]) and Lhx9 (a LIM homeobox required for sex differentiation and normal fertility [44]). Moreover, Mybl1 was independently identified recently in an unbiased mutagenesis screen for infertility phenotypes involving meiotic arrest [45]. We found that the Mybl1 mutants are characterized by low testis weight and depletion of male germ cells, as shown in Figure 5B. Additionally, analysis of the mutant testis transcriptome suggested that MYBL1 is a “master regulator” of the meiotic cell cycle and transcriptional program [46], and at least one gene regulated by MYBL1, Cyct, is among the top interactors of MYBL1 predicted by our network. Together, these findings on an infertility phenotype and suggestions of a corresponding potential mechanism confirm the accuracy of predictions from our tissue-specific network and show that when taken with expression analyses and other data, they can be used as a basis for functional testing.
In addition to the well-studied phenotype of male infertility, we also examined a less well-understood disease, ataxia, to investigate whether our tissue-specific networks can identify genes related to phenotypes or diseases with limited prior knowledge. Gene identification through genetic approaches, such as pedigree analyses, has had a major impact on our understanding of ataxia (over 40 candidate genes identified so far). Genetic testing is now an integral part of assessment. Routinely, a blood sample of any new ataxia case is mailed in for laboratory evaluation. However, the majority of the sporadic cases as well as the familial cases are so far unexplained. We curated the known gene list (43 in total) related to human ataxia, mapped these genes to their mouse orthologs, and used this list as seeds to predict additional candidate genes using our cerebellum-specific network, which is the major tissue affected by ataxia.
Our cerebellum-specific network reveals connections of ataxia-related genes not shown in the global network. A key, known ataxia gene is Atcay (ataxia, cerebellar, Cayman type homolog (human)), and in the cerebellum-specific network, two of its top interactors are Cacna1e (with connection confidence 0.943, ranked 18) and Grm1 (0.902, ranked 46) (Figure 6). These are plausible candidate genes since Grm1 is a known mouse ataxia gene [47], and Cacna1e encodes a subunit of an R-type calcium channel, while mutations to the related protein family member Cacna1a, encoding a subunit of an L-type calcium channel, causing spinocerebellar ataxia. However, in the global network these interactions are much weaker (0.647 for Cacna1e and 0.763 for Grm1 respectively), and would not be identified in the top 100 connections of Atcay, which supports the utility of tissue-specific networks relevant to ataxia to identify candidate genes (Figure 6B).
In addition to identifying these novel, likely correct edges, we also identified novel candidates using our SVM-based approach described above. Out of our top 10 novel candidates, we found strong evidence in the literature for 4 of these genes to be associated to ataxia (Table 2); suggesting at least a 40% success rate at low levels of recall. Among these, SORBS1 physically interacts with ATXN7, an autosomal dominant gene causing cerebellar ataxia [48]. RBFOX1 physically interacts with the c-terminus of ATXN2, another autosomal dominant gene causing cerebellar ataxia [48], [49]. It is thought that RBFOX1 might contribute to the restricted pathology of spinocerebellar ataxia type 2 (SCA2) [50]. The homozygous mouse knockout of a third gene, Plcb4 induces ataxia, although no human patients have been identified with mutations in this gene. A fourth gene, Plp1, is implicated in Spastic paraplegia-2 and Pelizaeus-Merzbacher diseases [51], which are disorders closely related to ataxia. It is also a homologue of Pmp22, which is involved in Charcot Marie Tooth disease type 1A, a sensory neuropathy common in some forms of ataxia [52]. Thus, even in the case of less well-studied phenotypes or diseases, our tissue-specific approach is able to identify likely candidates as evidenced by our success rate of at least 40% for ataxia-related predictions based on the cerebellar network, compared to a background detection rate of less than 1/500.
Genetic diseases often manifest tissue-specific pathologies [1], [2], [3], [4]. Therefore, acquiring tissue-specific functional information is essential for biomarker identification, diagnosis, and drug discovery. Current integrative functional genomics approaches to study diseases or phenotypes generally do not analyze them in the context of specific tissues. Our work represents a conceptual advance to address tissue-specificity in genome-scale functional studies of phenotypes. We describe a strategy to systematically generate tissue-specific functional networks that are robust and accurate for mining phenotype-related genes, demonstrating the importance of tissue-specific approaches for understanding human diseases.
Our approach addresses the twin challenges of incomplete systematic knowledge of tissue-specific protein functions and of limited availability and coverage of tissue-specific high-throughput functional data. Due to this lack of systematically defined tissue-specific genomic data, our approach uses highly reliable, low-throughput measures of gene expression to constrain our gold standard examples into tissue-specific sets. As more tissue-specific protein functions are defined systematically, perhaps with the help of hypotheses generated by approaches such as this, tissue-specific functional interactions will be directly used for experimental testing. Many genomic datasets, especially physical interaction studies, such as yeast 2-hybrid screens, and large-scale genetic screens, utilize artificial or in vitro contexts that may or may not reflect tissue-specific functional roles. Other data, however, such as high-throughput gene expression datasets (e.g. microarrays or RNA-seq), is often collected in a specific tissue or cellular context and may thus reflect a more restricted, tissue-specific set of genes or proteins. In our approach, we use the power of Bayesian machine learning to learn the predictive power of each dataset, whether in vivo or in vitro, by utilizing training sets restricted to gene pairs that are both expressed in the same tissue or context. In this way, data from empirically relevant contexts are trusted, while irrelevant data are disregarded.
While our current study focuses on predicting genotype-phenotype associations using tissue-specific functional relationship networks, the potential application of tissue-specific networks extends far beyond predicting phenotype-associated genes. For example, just as perturbations of the same gene may lead to different phenotypic outcomes across different tissues; treatments with bioactive chemicals or drugs may manifest differential effects across different tissues. Our broad conceptual framework of utilizing tissue-specific expression to refine a global network could be brought into these application domains such as drug target identification.
For each pair of networks, we quantify how much each gene has changed in the network relative to its neighbors. Suppose in network X, the connection weight between gene i and gene j is Xij, and in network Y, the connection weight between i and j is Yij. The prior for network X is Px, and the prior for network Y is Py. The score representing how much the gene i changed from network X to network Y is defined as:(8)n represents all other genes. We then calculate the gene ontology enrichment of top 100 changed genes for each network against the global network using established techniques [62].
To examine the role of Myb1 in spermatogenesis, as described in detail elsewhere [46], novel ENU-induced fertility mutations were identified in a 3-generation breeding scheme. Standard histological methods were used for preliminary characterization of the Mybl1 mutant phenotype.
We curated 43 genes causing ataxia that have been confirmed in human pedigree studies. These 43 genes were mapped to mouse one-to-one orthologs using the orthology defined by MGI [19], and were used as seeding genes for predicting additional candidates.
To allow dynamic visualization and cross-network comparison of our integration results, we developed the mouseMAP software (http://mouseMAP.princeton.edu), based on the open-source viewing framework Graphle that we developed in [63]. MouseMAP is based on the Prefuse Java visualization library, the Args4j command line parsing tool, and the SQLiteJDBC SQLite database driver. The basic functionality of mouseMAP allows querying one or multiple genes and retrieving the local network surrounding the query, with user-variable node number and confidence level cutoffs.
Our public, web-based system features cross-comparison of different networks that highlights connections in the newly queried network vs. the previously queried network, which allows us to compare the connections between different tissues of the same query gene(s). Gene information, including annotation, phenotype and disease association is retrievable through the interface. To facilitate general public use, mouseMAP also dynamically generates figure descriptions based on the current query and network structure.
|
10.1371/journal.pntd.0000776 | Diagnostic Accuracy of the Leishmania OligoC-TesT and NASBA-Oligochromatography for Diagnosis of Leishmaniasis in Sudan | The Leishmania OligoC-TesT and NASBA-Oligochromatography (OC) were recently developed for simplified and standardised molecular detection of Leishmania parasites in clinical specimens. We here present the phase II evaluation of both tests for diagnosis of visceral leishmaniasis (VL), cutaneous leishmaniasis (CL) and post kala-azar dermal leishmaniasis (PKDL) in Sudan.
The diagnostic accuracy of the tests was evaluated on 90 confirmed and 90 suspected VL cases, 7 confirmed and 8 suspected CL cases, 2 confirmed PKDL cases and 50 healthy endemic controls from Gedarif state and Khartoum state in Sudan.
The OligoC-TesT as well as the NASBA-OC showed a sensitivity of 96.8% (95% CI: 83.8%–99.4%) on lymph node aspirates and of 96.2% (95% CI: 89.4%–98.7%) on blood from the confirmed VL cases. The sensitivity on bone marrow was 96.9% (95% CI: 89.3%–99.1%) and 95.3% (95% CI: 87.1%–98.4%) for the OligoC-TesT and NASBA-OC, respectively. All confirmed CL and PKDL cases were positive with both tests. On the suspected VL cases, we observed a positive OligoC-TesT and NASBA-OC result in 37.1% (95% CI: 23.2%–53.7%) and 34.3% (95% CI: 20.8%–50.9%) on lymph, in 72.7% (95% CI: 55.8%–84.9%) and 63.6% (95% CI: 46.6%–77.8%) on bone marrow and in 76.9% (95% CI: 49.7%–91.8%) and 69.2% (95% CI: 42.4%–87.3%) on blood. Seven out of 8 CL suspected cases were positive with both tests. The specificity on the healthy endemic controls was 90% (95% CI: 78.6%–95.7%) for the OligoC-TesT and 100% (95% CI: 92.9%–100.0%) for the NASBA-OC test.
Both tests showed high sensitivity on lymph, blood and tissue scrapings for diagnosis of VL, CL and PKDL in Sudan, but the specificity for clinical VL was significantly higher with NASBA-OC.
| The leishmaniases are a group of vector-borne diseases caused by protozoan parasites of the genus Leishmania. The parasites are transmitted by phlebotomine sand flies and can cause, depending on the infecting species, three clinical manifestations of leishmaniasis: visceral leishmaniasis (VL), post kala-azar dermal leishmaniasis (PKDL) and cutaneous leishmaniasis (CL) including the mucocutaneous form. VL, PKDL as well as CL are endemic in several parts of Sudan, and VL especially represents a major health problem in this country. Molecular tests such as the polymerase chain reaction (PCR) or nucleic acid sequence based assay (NASBA) are powerful techniques for accurate detection of the parasite in clinical specimens, but broad use is hampered by their complexity and lack of standardisation. Recently, the Leishmania OligoC-TesT and NASBA-Oligochromatography were developed as simplified and standardised PCR and NASBA formats. In this study, both tests were phase II evaluated for diagnosis of VL, PKDL and CL in Sudan.
| The leishmaniases are a group of vector-borne diseases caused by parasites of the genus Leishmania. The parasites are transmitted by phlebotomine sand flies and can cause, depending on the infecting species, three main clinical manifestations of leishmaniasis: visceral leishmaniasis (VL), post kala-azar dermal leishmaniasis (PKDL) and cutaneous leishmaniasis (CL) including the mucocutaneous form [1]. VL is the most severe form in which the internal organs are invaded and tends to be 100% fatal if not appropriately treated. While CL and MCL are clinical manifestations as a result from replication of the parasite in the dermis and naso-oropharyngeal mucosa respectively, PKDL is a skin disorder seen in a number of treated VL patients. VL, PKDL and CL are endemic in several parts of Sudan and especially VL represents a major health problem in this country [2].
Although serological tests such as the direct agglutination test (DAT) [3], [4] and the rK39 dipstick test [5]–[7] have become the mainstay in VL diagnosis [8], parasite detection by microscopic analysis of aspirates from the lymph nodes, bone marrow or spleen is still used in some endemic regions. The diagnostic standard for CL and PKDL is microscopic analysis of tissue biopsies or scrapings. However, microscopy is hampered by its low and variable sensitivity and the need for rather invasive sampling in the case of VL. Sensitivity may be increased by prior in vitro cultivation of the parasite, but this technique is cumbersome and time consuming. Serological tests are of high value to support clinical diagnosis of VL but they are less useful in patients co-infected with HIV [9] and antibodies remain detectable for years after successful treatment [10], [11]. Antibody detection with the rapid rK39 dipstick test is not yet implemented in Sudan due to the reported low diagnostic performance of the test in this region [12]–[14].
Molecular diagnostics such as the polymerase chain reaction (PCR) and nucleic acid sequence based assay (NASBA) offer attractive alternatives to conventional parasite detection as they are generally highly sensitive and specific. PCR detects the parasite's DNA through in vitro thermocyclic conditions [15], while NASBA amplifies the RNA by an isothermal reaction [16]. However, broad application of these powerful techniques in diagnosis and control of leishmaniasis is hindered by their complexity and lack of standardised test formats. Recently, the Leishmania OligoC-TesT and NASBA-Oligochromatography (OC) were introduced as promising PCR and NASBA formats for simplified and standardised molecular detection of Leishmania parasites [17], [18]. The tests are based on amplification of a short sequence within the Leishmania 18S ribosomal DNA (PCR) or RNA (NASBA), followed by simple and rapid detection of amplified products in dipstick format. Both tests are available as self-containing kits including all components needed except for the Taq polymerase enzyme in the OligoC-TesT and the Nuclisense Basic Kit in the NASBA-OC due to licensing issues. The tests showed high sensitivity and specificity on experimentally prepared specimens and on clinical specimens from a limited number of confirmed cases and healthy controls (phase I) [17], [18] and satisfactory repeatability and reproducibility in a multicenter evaluation study [19]. In this phase II study, we evaluated the two tests in Sudan on confirmed and suspected VL, CL and PKDL patients and on healthy endemic controls.
VL and PKDL suspects and endemic healthy controls were recruited in Gedarif State and CL suspects in Khartoum State between October 2008 and January 2009. VL and PKDL suspects were recruited in the health centers of villages within the endemic areas while the endemic healthy controls were volunteers from the same villages but not visiting the health centers. CL suspects were recruited at the Dermatology Hospital in Khartoum. Confirmed leishmaniasis cases were given appropriate treatment in the same health center or hospital as they were diagnosed. A participant was included in the study if 2 years or older and written consent was obtained from the individual or his/her guardian. Specimen collection was performed by the Faculty of Medicine of Khartoum University. Ethical clearance for the study was obtained from the ethical committees of the Federal Ministry of Health Committee in Sudan and the University of Antwerp in Belgium.
A participant was classified as (i) confirmed VL case if there was clinical suspicion for VL, DAT on serum was positive (titer ≥1∶3200) and parasites were observed in lymph or bone marrow aspirates by microscopic analysis; (ii) suspected VL case with positive DAT if there was clinical suspicion for VL, DAT on serum was positive, no parasites were observed in lymph or bone marrow aspirates and no previous history of VL was reported; (iii) suspected VL case with negative DAT if there was clinical suspicion for VL, DAT on serum was negative, no parasites were observed in lymph or bone marrow aspirates and no previous history of VL was reported; (iv) endemic healthy control if there was no clinical suspicion for VL, no previous history of VL and DAT on serum was negative; (v) confirmed CL or PDKL case if there was clinical suspicion for CL or PKDL and parasites were observed in lesion or skin scrapings by microscopic analysis; and (vi) suspected CL case if there was clinical suspicion for CL and no parasites were observed in lesion or skin scrapings. Clinical suspicion for VL was defined as a history of fever for two weeks or more and splenomegaly or lymphadenopathy and for CL and PKDL the presence of skin lesions or nodules. The reference tests were performed by experienced laboratory technicians immediately after specimen taking at the collection sites as described in the WHO manual on visceral leishmaniasis [20].
Two hundred µl anti-coagulated blood (confirmed and suspected VL cases and healthy endemic controls), and/or inguinal lymph node aspirate, and/or bone marrow aspirate (confirmed and suspected VL cases) or lesion or skin scrapings (confirmed and suspected CL and PKDL cases) was mixed with 200 µl of AngeroNA buffer (Mallinckrodt Baker, USA). This buffer allows specimen storage without loss of DNA and RNA quality. The same specimens were used for the reference tests and for the index tests. Specimens were shipped at 4°C from the collection site to Soba University hospital laboratory of Khartoum University and stored at 4°C for a maximum of two weeks. Nucleic acids of the specimens were extracted according to the method described by Boom et al. [21]. Elution was done in 50 µl of Tris-EDTA (TE) buffer and stored at −20°C until analysis. The extracts were tested with the Leishmania OligoC-TesT and NASBA-OC between October 2008 and February 2009 as described by Deborggraeve et al. [17] and Mugasa et al. [18]. In brief, with the OligoC-TesT DNA of the parasite is amplified by PCR and subsequently 40 µl of denatured PCR product is mixed with 40 µl of migration buffer preheated at 55°C for at least 20 minutes followed by dipping the Oligo-strip into the solution. The NASBA-OC amplifies an RNA sequence of the parasite by NASBA reaction after which 4 µl of the amplified product is mixed with 76 µl of migration buffer preheated at 55°C and the Oligo-strip is dipped into the solution. Test results are read after 10 minutes for both tests. Executors of the index tests were trained during a one-week workshop held in June 2007 at Makerere University, Kampala, Uganda. No external quality control confirming the reference test or index test results could be performed during the study. Executors of the index tests were not blinded to the results of the reference tests.
The sensitivity and specificity of the Leishmania OligoC-TesT and NASBA-OC were calculated from data entered into contingency tables. Differences in sensitivity and specificity between the two tests and differences in test results of specimen types were estimated by the Mc Nemar test. Concordances between the two tests were determined using the kappa index. All calculations were estimated at a 95% confidence interval (95% CI).
In the study the following participants were recruited: 90 confirmed and 67 suspected VL cases with positive DAT, 23 suspected VL cases with negative DAT, 7 confirmed and 8 suspected CL cases, 2 confirmed PKDL cases and 50 healthy endemic controls.
Molecular diagnostics are attractive alternatives to conventional parasite detection as they combine sensitivity with specificity. Recently, the Leishmania OligoC-TesT and NASBA-OC were developed as standardised formats for molecular detection of Leishmania parasites [17], [18]. Here we present the phase II evaluation of the two tests on confirmed and suspected VL, CL and PKDL cases and on health endemic controls from Sudan. Although the study was set out as a phase II diagnostic evaluation, it has some limitations. A weak point is the lack of external quality control on a subset of specimens carried out by another reference laboratory. Furthermore, the executors of the index tests were not blinded to the participant classification and thus the results of the reference tests. Although we are convinced that these limitations did not influence the study results, these shortcomings should be avoided in future phase II and III evaluation trials. In addition, response to treatment was not included in the standard references as we were not able to follow up the patients after treatment. Lastly, the number of participants in some subgroups is too low to make major conclusions.
The Leishmania OligoC-TesT as well as the NASBA-OC showed a high sensitivity (>95%) on lymph, blood and bone marrow from the confirmed VL patients. The observation that analysing lymph or blood yielded similar sensitivity as bone marrow is very promising towards less invasive diagnosis. Indeed this means that the OligoC-TesT and NASBA-OC on lymph or blood could indicate the infection status of VL suspected cases. Similar findings were reported in a study on the phase III evaluation of conventional PCR for VL diagnosis in Nepal [23]. Although the number of confirmed CL and PKDL cases in our study was limited, all were positive with both tests indicating their potential for accurate detection of the parasites in skin tissues. In addition, the high sensitivity is confirmed by the observation that the tests were able to detect Leishmania RNA or DNA in the blood or lesion scrapings of to the majority of the suspected VL and CL cases. Several cases in the suspected patient with positive DAT group might be true leishmaniasis cases given the suboptimal sensitivity of the conventional parasite detection tests. In 2008, Deborggraeve et al. reported a sensitivity of conventional PCR on probable VL cases (defined as clinical suspicion of VL and positive DAT test but negative in conventional parasite detection) of 67.6% on blood and 71.8% on bone marrow [23]. The sensitivities of the index tests on the lymph node aspirates of the suspected VL cases is surprisingly low (34–37%). This might indicate that the parasite load in blood is higher than in lymph. The fact that this is not observed in the confirmed VL cases is probably due to the general higher parasite load in this patient group because of the low sensitivity of the reference test. PCR positivity in the suspected VL group with negative DAT was not significantly lower than with positive DAT. This confirms the low correlation in DAT and PCR status in these groups as observed earlier [23]. While antibody levels in the blood are a marker for the host response, the PCR/NASBA status of the blood is a marker for the presence of parasites and might therefore be complementary. PCR and NASBA can offer an added value compared to immunodiagnosis in HIV co-infected VL patients.
On the 50 healthy endemic controls, the OligoC-TesT showed a specificity of 90% while this was 100% for the NASBA-OC. The positive OligoC-TesT results might be due to asymptomatic infections, which are known to be common in VL endemic regions. The inclusion of these asymptomatic carriers in the control group could be explained by the low concordance between negative DAT status and PCR outcome on blood from endemic control persons as described by Deborggraeve et al. [23] and Bhattarai et al. [24]. However, while NASBA-OC showed equal sensitivity on the confirmed and suspected VL cases, the test did not detect these assumed asymptomatic infections in the endemic control group. This discrepancy can be explained in two ways. Firstly, although not indicated by the negative controls taken along in specimen analysis, contamination of the PCR can never be fully excluded. Secondly, the OligoC-TesT might be slightly more sensitive than the NASBA-OC and thus pick up asymptomatic infections which might have very low parasite loads in the blood. The observed equal sensitivity for both tests on the confirmed and suspected VL cases might be biased by the fact that these cases are individuals presenting syndromes and thus probably have higher parasite loads than healthy parasite carriers. Hence, NASBA-OC might provide a better marker for active disease than the OligoC-TesT, as NASBA-OC does not detect asymptomatic infections while more than 95% of the active VL cases are still positive with the test. Furthermore, both tests might be useful as a test of cure after treatment of VL. As cure does not always equals parasite clearance, NASBA-OC might be more suitable than the OligoC-TesT. However, specific evaluation studies are needed to confirm this hypothesis.
One should keep in mind that the PCR-OC and NASBA-OC are not yet an option for routine diagnosis at the primary care level as they require basic molecular biology lab facilities. VL typically affects populations in rural areas where health centers most often suffer from infrastructural limitations and thus only apply less sophisticated diagnostic methods. Yet, they can be valuable tools in leishmaniasis diagnosis at reference hospitals with basic molecular biology lab facilities. In addition, the evaluated tools can offer an added value in disease surveillance and epidemiological studies in which specimens are analysed at central reference laboratories.
In conclusion, the Leishmania OligoC-TesT and NASBA-OC showed high sensitivity on lymph and blood of VL patients and on scrapings from CL and PKDL patients from Sudan. A significant higher specificity for active VL with the NASBA-OC than with the OligoC-TesT was observed. Both tests are not yet an option for routine diagnosis of leishmaniasis at the primary care level due to their infrastructural requirements but they might be powerful diagnostic tests in disease surveillance programmes and in monitoring intervention studies.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.